llm - 2025_10
Navigation
- Part 1
- Part 2
- Part 3
- Part 4
- Part 5
- Part 6
- Part 7
- Part 8
- Part 9
- Part 10
- Part 11
- Part 12
- Part 13
- Part 14
- Part 15
- Part 16
- Part 17
- Part 18
- Part 19
- Part 20
Papers
This paper establishes a benchmark for evaluating tool-calling capabilities of large language models (LLMs) on multi-step geospatial tasks relevant to commercial GIS practitioners. We assess eight commercial LLMs (Claude Sonnet 3.5 and 4, Claude Haiku 3.5, Gemini 2.0 Flash, Gemini 2.5 Pro Preview, GPT-4o, GPT-4.1 and o4-mini) using a simple tool-calling agent equipped with 23 geospatial functions. Our benchmark comprises tasks in four categories of increasing complexity, with both solvable and intentionally unsolvable tasks to test rejection accuracy. We develop a LLM-as-Judge evaluation framework to compare agent solutions against reference solutions. Results show o4-mini and Claude 3.5 Sonnet achieve the best overall performance, OpenAI's GPT-4.1, GPT-4o and Google's Gemini 2.5 Pro Preview do not fall far behind, but the last two are more efficient in identifying unsolvable tasks. Claude Sonnet 4, due its preference to provide any solution rather than reject a task, proved to be less accurate. We observe significant differences in token usage, with Anthropic models consuming more tokens than competitors. Common errors include misunderstanding geometrical relationships, relying on outdated knowledge, and inefficient data manipulation. The resulting benchmark set, evaluation framework, and data generation pipeline are released as open-source resources (available at https://github.com/Solirinai/GeoBenchX), providing one more standardized method for the ongoing evaluation of LLMs for GeoAI.
Large Language Models (LLMs) have achieved remarkable success in generative tasks, including register-transfer level (RTL) hardware synthesis. However, their tendency to memorize training data poses critical risks when proprietary or security-sensitive designs are unintentionally exposed during inference. While prior work has examined memorization in natural language, RTL introduces unique challenges: In RTL, structurally different implementations (e.g., behavioral vs. gate-level descriptions) can realize the same hardware, leading to intellectual property (IP) leakage (full or partial) even without verbatim overlap. Conversely, even small syntactic variations (e.g., operator precedence or blocking vs. non-blocking assignments) can drastically alter circuit behavior, making correctness preservation especially challenging. In this work, we systematically study memorization in RTL code generation and propose CircuitGuard, a defense strategy that balances leakage reduction with correctness preservation. CircuitGuard (1) introduces a novel RTL-aware similarity metric that captures both structural and functional equivalence beyond surface-level overlap, and (2) develops an activation-level steering method that identifies and attenuates transformer components most responsible for memorization. Our empirical evaluation demonstrates that CircuitGuard identifies (and isolates) 275 memorization-critical features across layers 18-28 of Llama 3.1-8B model, achieving up to 80% reduction in semantic similarity to proprietary patterns while maintaining generation quality. CircuitGuard further shows 78-85% cross-domain transfer effectiveness, enabling robust memorization mitigation across circuit categories without retraining.
Recent reasoning Large Language Models (LLMs) demonstrate remarkable problem-solving abilities but often generate long thinking traces whose utility is unclear. Our work aims to improve their efficiency, enabling them to reach high performance without overthinking. First, we analyze the entropy of token probabilities in reasoning traces. Across three models, we observe a consistent U-shaped entropy pattern: high entropy on easy problems despite high accuracy, low entropy on problems with medium difficulty, and high entropy on hard problems reflecting uncertainty. Specifically, we notice 22--25\% entropy reduction from easy to medium difficulty regions, suggesting an {overthinking} phenomenon on easy instances. Building on these insights, we introduce \textbf{DiffAdapt}, a lightweight framework that selects Easy/Normal/Hard inference strategies per question based on their difficulty and reasoning trace entropy. Each inference strategy consists of a fixed prompt, temperature and maximum token length. In contrast to existing efficiency optimization methods, our approach does not fine-tune base LLM but a small probe that classifies LLM's final hidden state, allowing inexpensive adaptation. We comprehensively evaluate our method on five models and eight benchmarks. Our method achieves comparable or improved accuracy while reducing token usage by up to 22.4\%, establishing a practical path toward compute-efficient reasoning.
Web-based participatory urban sensing has emerged as a vital approach for modern urban management by leveraging mobile individuals as distributed sensors. However, existing urban sensing systems struggle with limited generalization across diverse urban scenarios and poor interpretability in decision-making. In this work, we introduce AgentSense, a hybrid, training-free framework that integrates large language models (LLMs) into participatory urban sensing through a multi-agent evolution system. AgentSense initially employs classical planner to generate baseline solutions and then iteratively refines them to adapt sensing task assignments to dynamic urban conditions and heterogeneous worker preferences, while producing natural language explanations that enhance transparency and trust. Extensive experiments across two large-scale mobility datasets and seven types of dynamic disturbances demonstrate that AgentSense offers distinct advantages in adaptivity and explainability over traditional methods. Furthermore, compared to single-agent LLM baselines, our approach outperforms in both performance and robustness, while delivering more reasonable and transparent explanations. These results position AgentSense as a significant advancement towards deploying adaptive and explainable urban sensing systems on the web.
The injection molding industry faces critical challenges in preserving and transferring field knowledge, particularly as experienced workers retire and multilingual barriers hinder effective communication. This study introduces IM-Chat, a multi-agent framework based on large language models (LLMs), designed to facilitate knowledge transfer in injection molding. IM-Chat integrates both limited documented knowledge (e.g., troubleshooting tables, manuals) and extensive field data modeled through a data-driven process condition generator that infers optimal manufacturing settings from environmental inputs such as temperature and humidity, enabling robust and context-aware task resolution. By adopting a retrieval-augmented generation (RAG) strategy and tool-calling agents within a modular architecture, IM-Chat ensures adaptability without the need for fine-tuning. Performance was assessed across 100 single-tool and 60 hybrid tasks for GPT-4o, GPT-4o-mini, and GPT-3.5-turbo by domain experts using a 10-point rubric focused on relevance and correctness, and was further supplemented by automated evaluation using GPT-4o guided by a domain-adapted instruction prompt. The evaluation results indicate that more capable models tend to achieve higher accuracy, particularly in complex, tool-integrated scenarios. In addition, compared with the fine-tuned single-agent LLM, IM-Chat demonstrated superior accuracy, particularly in quantitative reasoning, and greater scalability in handling multiple information sources. Overall, these findings demonstrate the viability of multi-agent LLM systems for industrial knowledge workflows and establish IM-Chat as a scalable and generalizable approach to AI-assisted decision support in manufacturing.
This research proposes a systematic, large language model (LLM) approach for extracting product and service attributes, features, and associated sentiments from customer reviews. Grounded in marketing theory, the framework distinguishes perceptual attributes from actionable features, producing interpretable and managerially actionable insights. We apply the methodology to 20,000 Yelp reviews of Starbucks stores and evaluate eight prompt variants on a random subset of reviews. Model performance is assessed through agreement with human annotations and predictive validity for customer ratings. Results show high consistency between LLMs and human coders and strong predictive validity, confirming the reliability of the approach. Human coders required a median of six minutes per review, whereas the LLM processed each in two seconds, delivering comparable insights at a scale unattainable through manual coding. Managerially, the analysis identifies attributes and features that most strongly influence customer satisfaction and their associated sentiments, enabling firms to pinpoint "joy points," address "pain points," and design targeted interventions. We demonstrate how structured review data can power an actionable marketing dashboard that tracks sentiment over time and across stores, benchmarks performance, and highlights high-leverage features for improvement. Simulations indicate that enhancing sentiment for key service features could yield 1-2% average revenue gains per store.
Recent work has demonstrated state-of-the-art results in large language model (LLM) hallucination detection and mitigation through consistency-based approaches which involve aggregating multiple responses sampled from a single LLM for a given prompt. These approaches help offset limitations stemming from the imperfect data on which LLMs are trained, which includes biases and under-representation of information required at deployment time among other limitations which can lead to hallucinations. We show that extending these single-model consistency methods to combine responses from multiple LLMs with different training data, training schemes and model architectures can result in substantial further improvements in hallucination detection and mitigation capabilities beyond their single-model consistency counterparts. We evaluate this \emph{consortium consistency} approach across many model teams from a pool of 15 LLMs and explore under what conditions it is beneficial to team together different LLMs in this manner. Further, we show that these performance improvements often come with reduced inference costs, offsetting a significant drawback with single-model consistency methods.
Supervised Fine-Tuning (SFT) Large Language Models (LLM) fundamentally rely on high-quality training data. While data selection and data synthesis are two common strategies to improve data quality, existing approaches often face limitations in static dataset curation that fail to adapt to evolving model capabilities. In this paper, we introduce Middo, a self-evolving Model-informed dynamic data optimization framework that uses model-aware data selection and context-preserving data refinement. Unlike conventional one-off filtering/synthesis methods, our framework establishes a closed-loop optimization system: (1) A self-referential diagnostic module proactively identifies suboptimal samples through tri-axial model signals - loss patterns (complexity), embedding cluster dynamics (diversity), and self-alignment scores (quality); (2) An adaptive optimization engine then transforms suboptimal samples into pedagogically valuable training points while preserving semantic integrity; (3) This optimization process continuously evolves with model capability through dynamic learning principles. Experiments on multiple benchmarks demonstrate that our Middo consistently enhances the quality of seed data and boosts LLM's performance with improving accuracy by 7.15% on average while maintaining the original dataset scale. This work establishes a new paradigm for sustainable LLM training through dynamic human-AI co-evolution of data and models. Our datasets, models, and code are publicly available at https://github.com/Word2VecT/Middo.
In this work, we study security of Model Context Protocol (MCP) agent toolchains and their applications in smart homes. We introduce AegisMCP, a protocol-level intrusion detector. Our contributions are: (i) a minimal attack suite spanning instruction-driven escalation, chain-of-tool exfiltration, malicious MCP server registration, and persistence; (ii) NEBULA-Schema (Network-Edge Behavioral Learning for Untrusted LLM Agents), a reusable protocol-level instrumentation that represents MCP activity as a streaming heterogeneous temporal graph over agents, MCP servers, tools, devices, remotes, and sessions; and (iii) a CPU-only streaming detector that fuses novelty, session-DAG structure, and attribute cues for near-real-time edge inference, with optional fusion of local prompt-guardrail signals. On an emulated smart-home testbed spanning multiple MCP stacks and a physical bench, AegisMCP achieves sub-second per-window model inference and end-to-end alerting. The latency of AegisMCP is consistently sub-second on Intel N150-class edge hardware, while outperforming traffic-only and sequence baselines; ablations confirm the importance of DAG and install/permission signals. We release code, schemas, and generators for reproducible evaluation.
Autonomous Driving Systems (ADS) are safety-critical, where failures can be severe. While Metamorphic Testing (MT) is effective for fault detection in ADS, existing methods rely heavily on manual effort and lack automation. We present AutoMT, a multi-agent MT framework powered by Large Language Models (LLMs) that automates the extraction of Metamorphic Relations (MRs) from local traffic rules and the generation of valid follow-up test cases. AutoMT leverages LLMs to extract MRs from traffic rules in Gherkin syntax using a predefined ontology. A vision-language agent analyzes scenarios, and a search agent retrieves suitable MRs from a RAG-based database to generate follow-up cases via computer vision. Experiments show that AutoMT achieves up to 5 x higher test diversity in follow-up case generation compared to the best baseline (manual expert-defined MRs) in terms of validation rate, and detects up to 20.55% more behavioral violations. While manual MT relies on a fixed set of predefined rules, AutoMT automatically extracts diverse metamorphic relations that augment real-world datasets and help uncover corner cases often missed during in-field testing and data collection. Its modular architecture separating MR extraction, filtering, and test generation supports integration into industrial pipelines and potentially enables simulation-based testing to systematically cover underrepresented or safety-critical scenarios.
Large Language Model (LLM)-based Multi-Agent Systems (MAS) have become a popular paradigm of AI applications. However, trustworthiness issues in MAS remain a critical concern. Unlike challenges in single-agent systems, MAS involve more complex communication processes, making them susceptible to corruption attacks. To mitigate this issue, several defense mechanisms have been developed based on the graph representation of MAS, where agents represent nodes and communications form edges. Nevertheless, these methods predominantly focus on static graph defense, attempting to either detect attacks in a fixed graph structure or optimize a static topology with certain defensive capabilities. To address this limitation, we propose a dynamic defense paradigm for MAS graph structures, which continuously monitors communication within the MAS graph, then dynamically adjusts the graph topology, accurately disrupts malicious communications, and effectively defends against evolving and diverse dynamic attacks. Experimental results in increasingly complex and dynamic MAS environments demonstrate that our method significantly outperforms existing MAS defense mechanisms, contributing an effective guardrail for their trustworthy applications. Our code is available at https://github.com/ChengcanWu/Monitoring-LLM-Based-Multi-Agent-Systems.
Self-correction of large language models (LLMs) emerges as a critical component for enhancing their reasoning performance. Although various self-correction methods have been proposed, a comprehensive evaluation of these methods remains largely unexplored, and the question of whether LLMs can truly correct themselves is a matter of significant interest and concern. In this study, we introduce CorrectBench, a benchmark developed to evaluate the effectiveness of self-correction strategies, including intrinsic, external, and fine-tuned approaches, across three tasks: commonsense reasoning, mathematical reasoning, and code generation. Our findings reveal that: 1) Self-correction methods can improve accuracy, especially for complex reasoning tasks; 2) Mixing different self-correction strategies yields further improvements, though it reduces efficiency; 3) Reasoning LLMs (e.g., DeepSeek-R1) have limited optimization under additional self-correction methods and have high time costs. Interestingly, a comparatively simple chain-of-thought (CoT) baseline demonstrates competitive accuracy and efficiency. These results underscore the potential of self-correction to enhance LLM's reasoning performance while highlighting the ongoing challenge of improving their efficiency. Consequently, we advocate for further research focused on optimizing the balance between reasoning capabilities and operational efficiency. Project Page: https://correctbench.github.io/
The advent of Large Language Models (LLMs) has revolutionized product recommenders, yet their susceptibility to adversarial manipulation poses critical challenges, particularly in real-world commercial applications. Our approach is the first one to tap into human psychological principles, seamlessly modifying product descriptions, making such manipulations hard to detect. In this work, we investigate cognitive biases as black-box adversarial strategies, drawing parallels between their effects on LLMs and human purchasing behavior. Through extensive evaluation across models of varying scale, we find that certain biases, such as social proof, consistently boost product recommendation rate and ranking, while others, like scarcity and exclusivity, surprisingly reduce visibility. Our results demonstrate that cognitive biases are deeply embedded in state-of-the-art LLMs, leading to highly unpredictable behavior in product recommendations and posing significant challenges for effective mitigation.
The creation of high-quality datasets to improve Large Language Model (LLM) reasoning remains a significant challenge, as current methods often suffer from generating low-quality/incorrect answers and limited information richness from available data sources. To address this, we propose AgenticMath, a novel agentic pipeline for generating high-quality mathematical question-answer pairs to enhance the supervised fine-tuning of LLMs. Our method operates through four stages: (1) Seed Question Filter that selects questions with high information richness, complexity, and clarity; (2) an Agentic Question Rephrase step that employs a multi-agent system to generate diverse, logically consistent paraphrases; (3) an Answer Augment step where rewrite answers using chain-of-thought reasoning to enhance numerical and logical correctness, without reliance on human-provided labels; and (4) a final Question and Answer Evaluation that retains only the most superior pairs. Extensive experiments demonstrate that, fine-tuning 3B-8B parameter LLMs on AgenticMath generated datasets (comprising only 30-60K math samples) achieves competitive or superior performance on diverse in domain and out-of-domain mathematical reasoning benchmarks compared to baselines trained on much more data (e.g., 400K or 2.3M samples). Our work demonstrates that targeted, high-quality data generation is a more efficient path to improving mathematical reasoning in LLMs than large-scale, low-quality alternatives.
The capability to precisely adhere to instructions is a cornerstone for Large Language Models (LLMs) to function as dependable agents in real-world scenarios. However, confronted with complex prompts, LLMs frequently encounter difficulties in fulfilling all specified requirements within a single response. Drawing inspiration from recent advancements in Chain-of-Thought (CoT) prompting and self-correction methodologies, we introduce Meeseeks (The name is inspired by Mr. Meeseeks from "Rick and Morty," a character renowned for efficiently accomplishing assigned tasks. See: https://en.wikipedia.org/wiki/Mr._Meeseeks), a fully automated iterative instruction-following benchmark equipped with an integrated feedback mechanism. Meeseeks identifies erroneous components in model responses and provides corresponding feedback accurately, thereby iteratively guiding the model toward self-correction. The dataset contains over 700 curated instances annotated by 32 distinct capability tags in Chinese and English. Extensive experimental results reveal that different state-of-the-art commercial and open-source LLMs exhibit vastly disparate performance, and even after 20 turns of iterative feedback-driven self-correction, nearly all models demonstrate suboptimal performance. We conducted comprehensive analysis from both macro and instance levels, uncovering numerous common issues prevalent in current state-of-the-art models, as well as several counterintuitive phenomena. We've open-sourced our work on https://github.com/ADoublLEN/Meeseeks.
In this paper, we investigate how personalising Large Language Models (Persona-LLMs) with annotator personas affects their sensitivity to hate speech, particularly regarding biases linked to shared or differing identities between annotators and targets. To this end, we employ Google's Gemini and OpenAI's GPT-4.1-mini models and two persona-prompting methods: shallow persona prompting and a deeply contextualised persona development based on Retrieval-Augmented Generation (RAG) to incorporate richer persona profiles. We analyse the impact of using in-group and out-group annotator personas on the models' detection performance and fairness across diverse social groups. This work bridges psychological insights on group identity with advanced NLP techniques, demonstrating that incorporating socio-demographic attributes into LLMs can address bias in automated hate speech detection. Our results highlight both the potential and limitations of persona-based approaches in reducing bias, offering valuable insights for developing more equitable hate speech detection systems.
Automated regression test generation has been extensively explored, yet generating high-quality tests for Python programs remains particularly challenging. Because of the Python's dynamic typing features, existing approaches, ranging from search-based software testing (SBST) to recent LLM-driven techniques, are often prone to type errors. Hence, existing methods often generate invalid inputs and semantically inconsistent test cases, which ultimately undermine their practical effectiveness. To address these limitations, we present Test4Py, a novel framework that enhances type correctness in automated test generation for Python. Test4Py leverages the program's call graph to capture richer contextual information about parameters, and introduces a behavior-based type inference mechanism that accurately infers parameter types and construct valid test inputs. Beyond input construction, Test4Py integrates an iterative repair procedure that progressively refines generated test cases to improve coverage. In an evaluation on 183 real-world Python modules, Test4Py achieved an average statement coverage of 83.0% and branch coverage of 70.8%, outperforming state-of-the-art tools by 7.2% and 8.4%, respectively.
The rapid evolution of smart cities has increased the reliance on intelligent interconnected services to optimize infrastructure, resources, and citizen well-being. Agentic AI has emerged as a key enabler by supporting autonomous decision-making and adaptive coordination, allowing urban systems to respond in real time to dynamic conditions. Its benefits are evident in areas such as transportation, where the integration of traffic data, weather forecasts, and safety sensors enables dynamic rerouting and a faster response to hazards. However, its deployment across heterogeneous smart city ecosystems raises critical governance, risk, and compliance (GRC) challenges, including accountability, data privacy, and regulatory alignment within decentralized infrastructures. Evaluation of SORA-ATMAS with three domain agents (Weather, Traffic, and Safety) demonstrated that its governance policies, including a fallback mechanism for high-risk scenarios, effectively steer multiple LLMs (GPT, Grok, DeepSeek) towards domain-optimized, policy-aligned outputs, producing an average MAE reduction of 35% across agents. Results showed stable weather monitoring, effective handling of high-risk traffic plateaus 0.85, and adaptive trust regulation in Safety/Fire scenarios 0.65. Runtime profiling of a 3-agent deployment confirmed scalability, with throughput between 13.8-17.2 requests per second, execution times below 72~ms, and governance delays under 100 ms, analytical projections suggest maintained performance at larger scales. Cross-domain rules ensured safe interoperability, with traffic rerouting permitted only under validated weather conditions. These findings validate SORA-ATMAS as a regulation-aligned, context-aware, and verifiable governance framework that consolidates distributed agent outputs into accountable, real-time decisions, offering a resilient foundation for smart-city management.
We introduce Traffic-R1, a 3B-parameter foundation model with human-like reasoning for Traffic signal control (TSC), developed via self-exploration and iterative reinforcement of LLM with expert guidance in a simulated traffic environment. Compared with traditional reinforcement learning and recent LLM-based methods, Traffic-R1 offers three main advantages: zero-shot generalization, transferring unchanged to new road networks and out-of-distribution incidents by leveraging internal traffic-control policies and reasoning; a compact 3B-parameter design that supports real-time inference on mobile-class chips for edge deployment; and an explainable TSC process that enables multi-intersection coordination through communication and an asynchronous communication network. Extensive benchmarks show Traffic-R1 outperforms strong baselines and training-intensive RL controllers. In production, the model now manages signals affecting over 55,000 drivers daily, reduces average queue lengths by more than 5%, and halves operator workload. Our model is available at https://huggingface.co/Season998/Traffic-R1.
Can large language model (LLM) agents reproduce the complex social dynamics that characterize human online behavior -- shaped by homophily, reciprocity, and social validation -- and what memory and learning mechanisms enable such dynamics to emerge? We present a multi-agent LLM simulation framework in which agents repeatedly interact, evaluate one another, and adapt their behavior through in-context learning accelerated by a coaching signal. To model human social behavior, we design behavioral reward functions that capture core drivers of online engagement, including social interaction, information seeking, self-presentation, coordination, and emotional support. These rewards align agent objectives with empirically observed user motivations, enabling the study of how network structures and group formations emerge from individual decision-making. Our experiments show that coached LLM agents develop stable interaction patterns and form emergent social ties, yielding network structures that mirror properties of real online communities. By combining behavioral rewards with in-context adaptation, our framework establishes a principled testbed for investigating collective dynamics in LLM populations and reveals how artificial agents may approximate or diverge from human-like social behavior.
We study weight-only post-training quantization (PTQ), which quantizes the weights of a large language model (LLM) without retraining, using little or no calibration data. Weight-only PTQ is crucial for reducing the memory footprint and latency of LLM inference, especially in memory-bound, small-batch inference scenarios, such as personalized inference on edge devices. Despite its importance, irregular weight distributions with heavy-tailed outliers in LLMs complicate quantization, recently motivating rotation-based methods that transform weights into near-Gaussian distributions, which are more regular with fewer outliers, thereby reducing quantization error. In this work, we first derive the information-theoretically optimal bit allocation for Gaussianized weights under given bit budgets, revealing that fine-grained fractional-bit quantizers approaching the Gaussian distortion-rate bound are essential to achieve near-optimal quantization performance. To bridge this theoretical insight and practical implementation, we introduce Q-Palette, a versatile collection of fractional-bit quantizers that range from trellis-coded quantizers offering near-optimal distortion to simpler vector and scalar quantizers optimized for faster inference, all efficiently implemented with optimized CUDA kernels across various bitwidths. Furthermore, leveraging Q-Palette as a foundational component, we propose a novel mixed-scheme quantization framework, jointly optimizing quantizer choices and layer fusion decisions given resource constraints. The code is available at https://github.com/snu-mllab/Q-Palette.
We propose a general-purpose approach for improving the ability of Large Language Models (LLMs) to intelligently and adaptively gather information from a user or other external source using the framework of sequential Bayesian experimental design (BED). This enables LLMs to act as effective multi-turn conversational agents and interactively interface with external environments. Our approach, which we call BED-LLM (Bayesian Experimental Design with Large Language Models), is based on iteratively choosing questions or queries that maximize the expected information gain (EIG) about the task of interest given the responses gathered previously. We show how this EIG can be formulated (and then estimated) in a principled way using a probabilistic model derived from the LLM's predictive distributions and provide detailed insights into key decisions in its construction and updating procedure. We find that BED-LLM achieves substantial gains in performance across a wide range of tests based on the 20 questions game and using the LLM to actively infer user preferences, compared to direct prompting of the LLM and other adaptive design strategies.
The training scale of large language models (LLMs) has reached tens of thousands of GPUs and is still continuously expanding, enabling faster learning of larger models. Accompanying the expansion of the resource scale is the prevalence of failures (CUDA error, NaN values, job hang, etc.), which poses significant challenges to training stability. Any large-scale LLM training infrastructure should strive for minimal training interruption, efficient fault diagnosis, and effective failure tolerance to enable highly efficient continuous training. This paper presents ByteRobust, a large-scale GPU infrastructure management system tailored for robust and stable training of LLMs. It exploits the uniqueness of LLM training process and gives top priorities to detecting and recovering failures in a routine manner. Leveraging parallelisms and characteristics of LLM training, ByteRobust enables high-capacity fault tolerance, prompt fault demarcation, and localization with an effective data-driven approach, comprehensively ensuring continuous and efficient training of LLM tasks. ByteRobust is deployed on a production GPU platform and achieves 97% ETTR for a three-month training job on 9,600 GPUs.
Phishing websites remain a significant cybersecurity threat, necessitating accurate and cost-effective detection mechanisms. In this paper, we present CLASP, a novel system that effectively identifies phishing websites by leveraging multiple intelligent agents, built using large language models (LLMs), to analyze different aspects of a web resource. The system processes URLs or QR codes, employing specialized LLM-based agents that evaluate the URL structure, webpage screenshot, and HTML content to predict potential phishing threats. To optimize performance while minimizing operational costs, we experimented with multiple combination strategies for agent-based analysis, ultimately designing a strategic combination that ensures the per-website evaluation expense remains minimal without compromising detection accuracy. We tested various LLMs, including Gemini 1.5 Flash and GPT-4o mini, to build these agents and found that Gemini 1.5 Flash achieved the best performance with an F1 score of 83.01% on a newly curated dataset. Also, the system maintained an average processing time of 2.78 seconds per website and an API cost of around $3.18 per 1,000 websites. Moreover, CLASP surpasses leading previous solutions, achieving over 40% higher recall and a 20% improvement in F1 score for phishing detection on the collected dataset. To support further research, we have made our dataset publicly available, supporting the development of more advanced phishing detection systems.
The paradigm of LLM-as-a-judge is emerging as a scalable and efficient alternative to human evaluation, demonstrating strong performance on well-defined tasks. However, its reliability in open-ended tasks with dynamic environments and complex interactions remains unexplored. To bridge the gap, we introduce WebDevJudge, a systematic benchmark for assessing LLM-as-a-judge performance in web development, with support for both non-interactive evaluation based on static observations and continuous interactive evaluation with a dynamic web environment. WebDevJudge comprises human preference labels over paired web implementations, annotated with structured and query-grounded rubrics to ensure high-quality ground truth. Using this benchmark, we comprehensively evaluate various evaluators, including LLMs, MLLMs, and agentic workflows. We systematically investigate the impact of different paradigms and guidance mechanisms. Our experiments reveal a significant gap between LLM judges and human experts. In-depth analysis indicates this gap stems from fundamental model limitations, including failures in recognizing functional equivalence, verifying task feasibility, and mitigating bias. Overall, WebDevJudge presents a significant challenge to LLM-as-a-judge, offering insights to guide future research toward developing more reliable and capable automated evaluators for complicated scenarios. Code and data are available at https://github.com/lcy2723/WebDevJudge.
Large Language Models (LLMs) increasingly rely on emerging protocols such as the Model Context Protocol (MCP) to invoke external tools and services. However, current tool routing mechanisms remain fragile because they only consider functional matching between users' queries and tools. In practice, user intent expressed through queries can be vague or underspecified, and the actual Quality of Experience (QoE) also depends on external factors such as link latency and server availability that are not captured by semantics alone. To address this challenge, we propose JAUNT, a framework for Joint Alignment of User intent and Network state in QoE-centric Tool routing. JAUNT introduces a dual-view alignment strategy that interprets user intent while employing LLM agents to construct network profiles, mapping numerical performance indicators into the semantic space to guide routing. We further design a benchmark that integrates diverse user request patterns with heterogeneous network states, enabling systematic evaluation of QoE outcomes. Experimental results show that JAUNT significantly improves QoE compared with several baselines, demonstrating the importance of aligning both intent and network state for scalable LLM service orchestration.
Large Language Models (LLMs), as the foundational architecture for next-generation interactive AI applications, not only power intelligent dialogue systems but also drive the evolution of embodied intelligence on edge devices, including humanoid robots, smart vehicles, and other scenarios. The applications running on these edge devices impose differentiated Service Level Objectives (SLO) requirements on LLM services, specifically manifested as distinct constraints on Time to First Token (TTFT) and Time Per Output Token (TPOT) as well as end-to-end latency. Notably, edge devices typically handle real-time tasks that are extremely sensitive to latency, such as machine control and navigation planning. However, existing scheduling service systems still prioritize maximizing output token throughput as the sole optimization objective, failing to adequately address the diversity of SLO requirements. This ultimately results in persistently high violation rates for end-to-end latency or TPOT related SLOs. This paper proposes SLICE, an innovative scheduling solution designed for edge computing scenarios with differentiated SLO requirements. By combining a utility-maximizing request scheduling algorithm with a dynamic iterative control mechanism for generation rates, SLICE significantly improves LLM inference service SLO attainment. Experimental results demonstrate that compared to state-of-the-art solutions Orca and FastServe, SLICE achieves up to 35x higher SLO attainment and 3.4x advantage in task completion time than the other two solutions.
Product search is a crucial component of modern e-commerce platforms, with billions of user queries every day. In product search systems, first-stage retrieval should achieve high recall while ensuring efficient online deployment. Sparse retrieval is particularly attractive in this context due to its interpretability and storage efficiency. However, sparse retrieval methods suffer from severe vocabulary mismatch issues, leading to suboptimal performance in product search scenarios.With their potential for semantic analysis, large language models (LLMs) offer a promising avenue for mitigating vocabulary mismatch issues and thereby improving retrieval quality. Directly applying LLMs to sparse retrieval in product search exposes two key challenges:(1)Queries and product titles are typically short and highly susceptible to LLM-induced hallucinations, such as generating irrelevant expansion terms or underweighting critical literal terms like brand names and model numbers;(2)The large vocabulary space of LLMs leads to difficulty in initializing training effectively, making it challenging to learn meaningful sparse representations in such ultra-high-dimensional spaces.To address these challenges, we propose PROSPER, a framework for PROduct search leveraging LLMs as SParsE Retrievers. PROSPER incorporates: (1)A literal residual network that alleviates hallucination in lexical expansion by reinforcing underweighted literal terms through a residual compensation mechanism; and (2)A lexical focusing window that facilitates effective training initialization via a coarse-to-fine sparsification strategy.Extensive offline and online experiments show that PROSPER significantly outperforms sparse baselines and achieves recall performance comparable to advanced dense retrievers, while also achieving revenue increments online.
Large language models achieve impressive performance across diverse tasks but exhibit high inference latency due to their large parameter sizes. While quantization reduces model size, it often leads to performance degradation compared to the full model. Speculative decoding remains lossless but typically incurs extra overheads. We propose SPEQ, an algorithm-hardware co-designed speculative decoding method that uses part of the full-model weight bits to form a quantized draft model, thereby eliminating additional training or storage overhead. A reconfigurable processing element array enables efficient execution of both the draft and verification passes. Experimental results across 15 LLMs and tasks demonstrate that SPEQ achieves speedups of 2.07x, 1.53x, and 1.45x compared over FP16, Olive, and Tender, respectively.
Continued pretraining and instruction tuning on large-scale multilingual data have proven to be effective in scaling large language models (LLMs) to low-resource languages. However, the unaligned nature of such data limits its ability to effectively capture cross-lingual semantics. In contrast, multi-way parallel data, where identical content is aligned across multiple languages, provides stronger cross-lingual consistency and offers greater potential for improving multilingual performance. In this paper, we introduce a large-scale, high-quality multi-way parallel corpus, TED2025, based on TED Talks. The corpus spans 113 languages, with up to 50 languages aligned in parallel, ensuring extensive multilingual coverage. Using this dataset, we investigate best practices for leveraging multi-way parallel data to enhance LLMs, including strategies for continued pretraining, instruction tuning, and the analysis of key influencing factors. Experiments on six multilingual benchmarks show that models trained on multiway parallel data consistently outperform those trained on unaligned multilingual data.
Security analysts face increasing pressure to triage large and complex vulnerability backlogs. Large Language Models (LLMs) offer a potential aid by automating parts of the interpretation process. We evaluate four models (ChatGPT, Claude, Gemini, and DeepSeek) across twelve prompting techniques to interpret semi-structured and unstructured vulnerability information. As a concrete use case, we test each model's ability to predict decision points in the Stakeholder-Specific Vulnerability Categorization (SSVC) framework: Exploitation, Automatable, Technical Impact, and Mission and Wellbeing. Using 384 real-world vulnerabilities from the VulZoo dataset, we issued more than 165,000 queries to assess performance under prompting styles including one-shot, few-shot, and chain-of-thought. We report F1 scores for each SSVC decision point and Cohen's kappa (weighted and unweighted) for the final SSVC decision outcomes. Gemini consistently ranked highest, leading on three of four decision points and yielding the most correct recommendations. Prompting with exemplars generally improved accuracy, although all models struggled on some decision points. Only DeepSeek achieved fair agreement under weighted metrics, and all models tended to over-predict risk. Overall, current LLMs do not replace expert judgment. However, specific LLM and prompt combinations show moderate effectiveness for targeted SSVC decisions. When applied with care, LLMs can support vulnerability prioritization workflows and help security teams respond more efficiently to emerging threats.
Control algorithms in production environments typically require domain experts to tune their parameters and logic for specific scenarios. However, existing research predominantly focuses on algorithmic performance under ideal or default configurations, overlooking the critical aspect of Tuning Potential. To bridge this gap, we introduce Crucible, an agent that employs an LLM-driven, multi-level expert simulation to turn algorithms and defines a formalized metric to quantitatively evaluate their Tuning Potential. We demonstrate Crucible's effectiveness across a wide spectrum of case studies, from classic control tasks to complex computer systems, and validate its findings in a real-world deployment. Our experimental results reveal that Crucible systematically quantifies the tunable space across different algorithms. Furthermore, Crucible provides a new dimension for algorithm analysis and design, which ultimately leads to performance improvements. Our code is available at https://github.com/thu-media/Crucible.
Large Language Models (LLMs) have shown great promise in automating data analytics tasks by interpreting natural language queries and generating multi-operation execution plans. However, existing LLM-agent-based analytics frameworks operate under the assumption of centralized data access, offering little to no privacy protection. In contrast, federated analytics (FA) enables privacy-preserving computation across distributed data sources, but lacks support for natural language input and requires structured, machine-readable queries. In this work, we present LAFA, the first system that integrates LLM-agent-based data analytics with FA. LAFA introduces a hierarchical multi-agent architecture that accepts natural language queries and transforms them into optimized, executable FA workflows. A coarse-grained planner first decomposes complex queries into sub-queries, while a fine-grained planner maps each subquery into a Directed Acyclic Graph of FA operations using prior structural knowledge. To improve execution efficiency, an optimizer agent rewrites and merges multiple DAGs, eliminating redundant operations and minimizing computational and communicational overhead. Our experiments demonstrate that LAFA consistently outperforms baseline prompting strategies by achieving higher execution plan success rates and reducing resource-intensive FA operations by a substantial margin. This work establishes a practical foundation for privacy-preserving, LLM-driven analytics that supports natural language input in the FA setting.
We present a probabilistic intent modeling framework for large language model (LLM) agents in multi-turn social dialogue. The framework maintains a belief distribution over a partner's latent intentions, initialized from contextual priors and dynamically updated through likelihood estimation after each utterance. The evolving distribution provides additional contextual grounding for the policy, enabling adaptive dialogue strategies under uncertainty. Preliminary experiments in the SOTOPIA environment show consistent improvements: the proposed framework increases the Overall score by 9.0% on SOTOPIA-All and 4.1% on SOTOPIA-Hard compared with the Qwen2.5-7B baseline, and slightly surpasses an oracle agent that directly observes partner intentions. These early results suggest that probabilistic intent modeling can contribute to the development of socially intelligent LLM agents.
Large language models (LLMs) have demonstrated impressive reasoning capabilities, but scaling their performance often relies on massive reasoning datasets that are computationally expensive to train on. Existing data selection methods aim to curate smaller, high-quality subsets but often rely on costly external models or opaque heuristics. In this work, we shift the focus from external heuristics to the model's internal mechanisms. We find that complex reasoning tasks consistently activate a sparse, specialized subset of attention heads, forming core reasoning circuits. Building on this insight, we propose CircuitSeer, a novel data selection method that quantifies the reasoning complexity of data by measuring its influence on these crucial circuits. Extensive experiments on 4 models and 9 datasets demonstrate CircuitSeer's superiority. Notably, fine-tuning Qwen2.5-Math-7B on just 10% of data selected by our method achieves a 1.4-point gain in average Pass@1 over training on the full dataset, highlighting its efficiency and effectiveness.
Phishing attacks in Web3 ecosystems are increasingly sophisticated, exploiting deceptive contract logic, malicious frontend scripts, and token approval patterns. We present DeepTx, a real-time transaction analysis system that detects such threats before user confirmation. DeepTx simulates pending transactions, extracts behavior, context, and UI features, and uses multiple large language models (LLMs) to reason about transaction intent. A consensus mechanism with self-reflection ensures robust and explainable decisions. Evaluated on our phishing dataset, DeepTx achieves high precision and recall (demo video: https://youtu.be/4OfK9KCEXUM).
Optimization modeling enables critical decisions across industries but remains difficult to automate: informal language must be mapped to precise mathematical formulations and executable solver code. Prior LLM approaches either rely on brittle prompting or costly retraining with limited generalization. We present AlphaOPT, a self-improving experience library that enables an LLM to learn from limited demonstrations (even answers alone, without gold-standard programs) and solver feedback - without annotated reasoning traces or parameter updates. AlphaOPT operates in a continual two-phase cycle: (i) a Library Learning phase that reflects on failed attempts, extracting solver-verified, structured insights as {taxonomy, condition, explanation, example}; and (ii) a Library Evolution phase that diagnoses retrieval misalignments and refines the applicability conditions of stored insights, improving transfer across tasks. This design (1) learns efficiently from limited demonstrations without curated rationales, (2) expands continually without costly retraining by updating the library rather than model weights, and (3) makes knowledge explicit and interpretable for human inspection and intervention. Experiments show that AlphaOPT steadily improves with more data (65% to 72% from 100 to 300 training items) and surpasses the strongest baseline by 7.7% on the out-of-distribution OptiBench dataset when trained only on answers. Code and data are available at: https://github.com/Minw913/AlphaOPT.
This paper proposes Memory-Augmented State Machine Prompting (MASMP), a novel framework for LLM agents in real-time strategy games. Addressing key challenges like hallucinations and fragmented decision-making in existing approaches, MASMP integrates state machine prompting with memory mechanisms to unify structured actions with long-term tactical coherence. The framework features: (1) a natural language-driven state machine architecture that guides LLMs to emulate finite state machines and behavior trees through prompts, and (2) a lightweight memory module preserving strategic variables (e.g., tactics, priority units) across decision cycles. Experiments in StarCraft II demonstrate MASMP's 60% win rate against the hardest built-in AI (Lv7), vastly outperforming baselines (0%). Case studies reveal the method retains LLMs' semantic comprehension while resolving the "Knowing-Doing Gap" through strict state-action mapping, achieving both interpretability and FSM-like reliability. This work establishes a new paradigm for combining neural and symbolic AI in complex decision-making.
We present $\textbf{Korean SimpleQA (KoSimpleQA)}$, a benchmark for evaluating factuality in large language models (LLMs) with a focus on Korean cultural knowledge. KoSimpleQA is designed to be challenging yet easy to grade, consisting of 1,000 short, fact-seeking questions with unambiguous answers. We conduct a comprehensive evaluation across a diverse set of open-source LLMs of varying sizes that support Korean, and find that even the strongest model generates correct answer only 33.7% of the time, underscoring the challenging nature of KoSimpleQA. Notably, performance rankings on KoSimpleQA differ substantially from those on the English SimpleQA, highlighting the unique value of our dataset. Furthermore, our analysis of reasoning LLMs shows that engaging reasoning capabilities in the factual QA task can both help models better elicit their latent knowledge and improve their ability to abstain when uncertain. KoSimpleQA can be found at https://anonymous.4open.science/r/KoSimpleQA-62EB.
Large Language Models (LLMs) are increasingly used to recommend mobile applications through natural language prompts, offering a flexible alternative to keyword-based app store search. Yet, the reasoning behind these recommendations remains opaque, raising questions about their consistency, explainability, and alignment with traditional App Store Optimization (ASO) metrics. In this paper, we present an empirical analysis of how widely-used general purpose LLMs generate, justify, and rank mobile app recommendations. Our contributions are: (i) a taxonomy of 16 generalizable ranking criteria elicited from LLM outputs; (ii) a systematic evaluation framework to analyse recommendation consistency and responsiveness to explicit ranking instructions; and (iii) a replication package to support reproducibility and future research on AI-based recommendation systems. Our findings reveal that LLMs rely on a broad yet fragmented set of ranking criteria, only partially aligned with standard ASO metrics. While top-ranked apps tend to be consistent across runs, variability increases with ranking depth and search specificity. LLMs exhibit varying sensitivity to explicit ranking instructions - ranging from substantial adaptations to near-identical outputs - highlighting their complex reasoning dynamics in conversational app discovery. Our results aim to support end-users, app developers, and recommender-systems researchers in navigating the emerging landscape of conversational app discovery.
Large language models (LLMs) suffer from temporal misalignment issues especially across long span of time. The issue arises from knowing that LLMs are trained on large amounts of data where temporal information is rather sparse over long times, such as thousands of years, resulting in insufficient learning or catastrophic forgetting by the LLMs. This paper proposes a methodology named "Ticktack" for addressing the LLM's long-time span misalignment in a yearly setting. Specifically, we first propose to utilize the sexagenary year expression instead of the Gregorian year expression employed by LLMs, achieving a more uniform distribution in yearly granularity. Then, we employ polar coordinates to model the sexagenary cycle of 60 terms and the year order within each term, with additional temporal encoding to ensure LLMs understand them. Finally, we present a temporal representational alignment approach for post-training LLMs that effectively distinguishes time points with relevant knowledge, hence improving performance on time-related tasks, particularly over a long period. We also create a long time span benchmark for evaluation. Experimental results prove the effectiveness of our proposal.
The rapid advancement of large language models (LLMs) has exacerbated the memory bottleneck due to the widening gap between model parameter scaling and hardware capabilities. While post-training quantization techniques effectively reduce memory overhead, existing methods predominantly rely on static quantization strategies, which struggle to adapt to dynamic workloads. To address this, we propose FlexQuant, a dynamic precision-switching framework that optimizes the trade-off between inference speed and accuracy. Leveraging model perplexity entropy and Kullback-Leibler divergence, FlexQuant enables fine-grained, layer-wise mixed-precision quantization and dynamically adjusts bit-widths during each token generation. FlexQuant provides a comprehensive analysis of quantization strategies, introduces a precision requirement model for optimal switching, and implements efficient fine-grained precision management. Evaluations demonstrate that FlexQuant achieves a 1.3x end-to-end speedup across diverse language tasks with negligible accuracy loss introduced. This framework offers a flexible and adaptive solution for efficient LLM deployment. Code is released at https://github.com/ZongwuWang/FlexQuant.git.
The wide deployment of Large Language Models (LLMs) has given rise to strong demands for optimizing their inference performance. Today's techniques serving this purpose primarily focus on reducing latency and improving throughput through algorithmic and hardware enhancements, while largely overlooking their privacy side effects, particularly in a multi-user environment. In our research, for the first time, we discovered a set of new timing side channels in LLM systems, arising from shared caches and GPU memory allocations, which can be exploited to infer both confidential system prompts and those issued by other users. These vulnerabilities echo security challenges observed in traditional computing systems, highlighting an urgent need to address potential information leakage in LLM serving infrastructures. In this paper, we report novel attack strategies designed to exploit such timing side channels inherent in LLM deployments, specifically targeting the Key-Value (KV) cache and semantic cache widely used to enhance LLM inference performance. Our approach leverages timing measurements and classification models to detect cache hits, allowing an adversary to infer private prompts with high accuracy. We also propose a token-by-token search algorithm to efficiently recover shared prompt prefixes in the caches, showing the feasibility of stealing system prompts and those produced by peer users. Our experimental studies on black-box testing of popular online LLM services demonstrate that such privacy risks are completely realistic, with significant consequences. Our findings underscore the need for robust mitigation to protect LLM systems against such emerging threats.
Despite progress in watermarking algorithms for large language models (LLMs), real-world deployment remains limited. We argue that this gap stems from misaligned incentives among LLM providers, platforms, and end users, which manifest as four key barriers: competitive risk, detection-tool governance, robustness concerns and attribution issues. We revisit three classes of watermarking through this lens. \emph{Model watermarking} naturally aligns with LLM provider interests, yet faces new challenges in open-source ecosystems. \emph{LLM text watermarking} offers modest provider benefit when framed solely as an anti-misuse tool, but can gain traction in narrowly scoped settings such as dataset de-contamination or user-controlled provenance. \emph{In-context watermarking} (ICW) is tailored for trusted parties, such as conference organizers or educators, who embed hidden watermarking instructions into documents. If a dishonest reviewer or student submits this text to an LLM, the output carries a detectable watermark indicating misuse. This setup aligns incentives: users experience no quality loss, trusted parties gain a detection tool, and LLM providers remain neutral by simply following watermark instructions. We advocate for a broader exploration of incentive-aligned methods, with ICW as an example, in domains where trusted parties need reliable tools to detect misuse. More broadly, we distill design principles for incentive-aligned, domain-specific watermarking and outline future research directions. Our position is that the practical adoption of LLM watermarking requires aligning stakeholder incentives in targeted application domains and fostering active community engagement.
Large Language Models (LLMs) frequently generate buggy code with complex logic errors that are challenging to diagnose. While existing LLM-based self-repair approaches conduct intensive static semantic analysis or reply on superficial execution logs, they miss the in-depth runtime behaviors that often expose bug root causes-lacking the interactive dynamic analysis capabilities that make human debugging effective. We present InspectCoder, the first agentic program repair system that empowers LLMs to actively conduct dynamic analysis via interactive debugger control. Our dual-agent framework enables strategic breakpoint placement, targeted state inspection, and incremental runtime experimentation within stateful debugger sessions. Unlike existing methods that follow fixed log collection procedures, InspectCoder adaptively inspects and perturbs relevant intermediate states at runtime, and leverages immediate process rewards from debugger feedback to guide multi-step reasoning, transforming LLM debugging paradigm from blind trial-and-error into systematic root cause diagnosis. We conduct comprehensive experiments on two challenging self-repair benchmarks: BigCodeBench-R and LiveCodeBench-R. InspectCoder achieves 5.10%-60.37% relative improvements in repair accuracy over the strongest baseline, while delivering 1.67x-2.24x superior bug-fix efficiency respectively. We also contribute InspectWare, an open-source middleware that abstracts debugger complexities and maintains stateful debugging sessions across mainstream Python testing frameworks. Our work provides actionable insight into the interactive LLM-debugger systems, demonstrating the significant potential of LLM-driven dynamic analysis for automated software engineering.
Training effective dense retrieval models typically relies on hard negative (HN) examples mined from large document corpora using methods such as BM25 or cross-encoders (CE), which require full corpus access. We propose a corpus-free alternative: an end-to-end pipeline where a Large Language Model (LLM) first generates a query from a passage and then produces a hard negative example using only the generated query text. Our dataset comprises 7,250 arXiv abstracts spanning diverse domains including mathematics, physics, computer science, and related fields, serving as positive passages for query generation. We evaluate two fine-tuning configurations of DistilBERT for dense retrieval; one using LLM-generated hard negatives conditioned solely on the query, and another using negatives generated with both the query and its positive document as context. Compared to traditional corpus-based mining methods {LLM Query $\rightarrow$ BM25 HN and LLM Query $\rightarrow$ CE HN on multiple BEIR benchmark datasets, our all-LLM pipeline outperforms strong lexical mining baselines and achieves performance comparable to cross-encoder-based methods, demonstrating the potential of corpus-free hard negative generation for retrieval model training.
Large Language Models (LLMs) achieve superior performance through Chain-of-Thought (CoT) reasoning, but these token-level reasoning chains are computationally expensive and inefficient. In this paper, we introduce Compressed Latent Reasoning (CoLaR), a novel framework that dynamically compresses reasoning processes in latent space through a two-stage training approach. First, during supervised fine-tuning, CoLaR extends beyond next-token prediction by incorporating an auxiliary next compressed embedding prediction objective. This process merges embeddings of consecutive tokens using a compression factor randomly sampled from a predefined range, and trains a specialized latent head to predict distributions of subsequent compressed embeddings. Second, we enhance CoLaR through reinforcement learning (RL) that leverages the latent head's non-deterministic nature to explore diverse reasoning paths and exploit more compact ones. This approach enables CoLaR to: i) perform reasoning at a dense latent level (i.e., silently), substantially reducing reasoning chain length, and ii) dynamically adjust reasoning speed at inference time by simply prompting the desired compression factor. Extensive experiments across four mathematical reasoning datasets demonstrate that CoLaR achieves 14.1% higher accuracy than latent-based baseline methods at comparable compression ratios, and reduces reasoning chain length by 53.3% with only 4.8% performance degradation compared to explicit CoT method. Moreover, when applied to more challenging mathematical reasoning tasks, our RL-enhanced CoLaR demonstrates performance gains of up to 5.4% while dramatically reducing latent reasoning chain length by 82.8%. The code and models will be released upon acceptance.
As large language model (LLM) agents increasingly automate complex web tasks, they boost productivity while simultaneously introducing new security risks. However, relevant studies on web agent attacks remain limited. Existing red-teaming approaches mainly rely on manually crafted attack strategies or static models trained offline. Such methods fail to capture the underlying behavioral patterns of web agents, making it difficult to generalize across diverse environments. In web agent attacks, success requires the continuous discovery and evolution of attack strategies. To this end, we propose Genesis, a novel agentic framework composed of three modules: Attacker, Scorer, and Strategist. The Attacker generates adversarial injections by integrating the genetic algorithm with a hybrid strategy representation. The Scorer evaluates the target web agent's responses to provide feedback. The Strategist dynamically uncovers effective strategies from interaction logs and compiles them into a continuously growing strategy library, which is then re-deployed to enhance the Attacker's effectiveness. Extensive experiments across various web tasks show that our framework discovers novel strategies and consistently outperforms existing attack baselines.
We study a web-deployed, tool-augmented LLM health coach with real users. In a pilot with seven users (280 rated turns), offline policy evaluation (OPE) over factorized decision heads (Tool/Style) shows that a uniform heavy-tool policy raises average value on logs but harms specific subgroups, most notably low-health-literacy/high-self-efficacy users. A lightweight simulator with hidden archetypes further shows that adding a small early information-gain bonus reliably shortens trait identification and improves goal success and pass@3. Together, these early findings indicate an evaluation-first path to personalization: freeze the generator, learn subgroup-aware decision heads on typed rewards (objective tool outcomes and satisfaction), and always report per-archetype metrics to surface subgroup harms that averages obscure.
Black-box Large Language Models (LLMs) provide practical and accessible alternatives to other machine learning methods, as they require minimal labeled data and machine learning expertise to develop solutions for various decision making problems. However, for applications that need operating with constraints on specific metrics (e.g., precision $\geq$ 95%), decision making with black-box LLMs remains unfavorable, due to their low numerical output cardinalities. This results in limited control over their operating points, preventing fine-grained adjustment of their decision making behavior. In this paper, we study using black-box LLMs as classifiers, focusing on efficiently improving their operational granularity without performance loss. Specifically, we first investigate the reasons behind their low-cardinality numerical outputs and show that they are biased towards generating rounded but informative verbalized probabilities. Then, we experiment with standard prompt engineering, uncertainty estimation and confidence elicitation techniques, and observe that they do not effectively improve operational granularity without sacrificing performance or increasing inference cost. Finally, we propose efficient approaches to significantly increase the number and diversity of available operating points. Our proposed approaches provide finer-grained operating points and achieve comparable to or better performance than the benchmark methods across 11 datasets and 3 LLMs.
As large language models (LLMs) evolve, their integration with 3D spatial data (3D-LLMs) has seen rapid progress, offering unprecedented capabilities for understanding and interacting with physical spaces. This survey provides a comprehensive overview of the methodologies enabling LLMs to process, understand, and generate 3D data. Highlighting the unique advantages of LLMs, such as in-context learning, step-by-step reasoning, open-vocabulary capabilities, and extensive world knowledge, we underscore their potential to significantly advance spatial comprehension and interaction within embodied Artificial Intelligence (AI) systems. Our investigation spans various 3D data representations, from point clouds to Neural Radiance Fields (NeRFs). It examines their integration with LLMs for tasks such as 3D scene understanding, captioning, question-answering, and dialogue, as well as LLM-based agents for spatial reasoning, planning, and navigation. The paper also includes a brief review of other methods that integrate 3D and language. The meta-analysis presented in this paper reveals significant progress yet underscores the necessity for novel approaches to harness the full potential of 3D-LLMs. Hence, with this paper, we aim to chart a course for future research that explores and expands the capabilities of 3D-LLMs in understanding and interacting with the complex 3D world. To support this survey, we have established a project page where papers related to our topic are organized and listed: https://github.com/ActiveVisionLab/Awesome-LLM-3D.
We present FinAI Data Assistant, a practical approach for natural-language querying over financial databases that combines large language models (LLMs) with the OpenAI Function Calling API. Rather than synthesizing complete SQL via text-to-SQL, our system routes user requests to a small library of vetted, parameterized queries, trading generative flexibility for reliability, low latency, and cost efficiency. We empirically study three questions: (RQ1) whether LLMs alone can reliably recall or extrapolate time-dependent financial data without external retrieval; (RQ2) how well LLMs map company names to stock ticker symbols; and (RQ3) whether function calling outperforms text-to-SQL for end-to-end database query processing. Across controlled experiments on prices and fundamentals, LLM-only predictions exhibit non-negligible error and show look-ahead bias primarily for stock prices relative to model knowledge cutoffs. Ticker-mapping accuracy is near-perfect for NASDAQ-100 constituents and high for S\&P~500 firms. Finally, FinAI Data Assistant achieves lower latency and cost and higher reliability than a text-to-SQL baseline on our task suite. We discuss design trade-offs, limitations, and avenues for deployment.
Large language models (LLMs) and their multimodal variants can now process visual inputs, including images of text. This raises an intriguing question: can we compress textual inputs by feeding them as images to reduce token usage while preserving performance? In this paper, we show that visual text representations are a practical and surprisingly effective form of input compression for decoder LLMs. We exploit the idea of rendering long text inputs as a single image and provide it directly to the model. This leads to dramatically reduced number of decoder tokens required, offering a new form of input compression. Through experiments on two distinct benchmarks RULER (long-context retrieval) and CNN/DailyMail (document summarization) we demonstrate that this text-as-image method yields substantial token savings (often nearly half) without degrading task performance.
The increasing number of data a booking platform such as Booking.com and AirBnB offers make it challenging for interested parties to browse through the available accommodations and analyze reviews in an efficient way. Efforts have been made from the booking platform providers to utilize recommender systems in an effort to enable the user to filter the results by factors such as stars, amenities, cost but most valuable insights can be provided by the unstructured text-based reviews. Going through these reviews one-by-one requires a substantial amount of time to be devoted while a respectable percentage of the reviews won't provide to the user what they are actually looking for. This research publication explores how Large Language Models (LLMs) can enhance short rental apartments recommendations by summarizing and mining key insights from user reviews. The web application presented in this paper, named "instaGuide", automates the procedure of isolating the text-based user reviews from a property on the Booking.com platform, synthesizing the summary of the reviews, and enabling the user to query specific aspects of the property in an effort to gain feedback on their personal questions/criteria. During the development of the instaGuide tool, numerous LLM models were evaluated based on accuracy, cost, and response quality. The results suggest that the LLM-powered summarization reduces significantly the amount of time the users need to devote on their search for the right short rental apartment, improving the overall decision-making procedure.
The Object Goal Navigation (ObjectNav) task challenges agents to locate a specified object in an unseen environment by imagining unobserved regions of the scene. Prior approaches rely on deterministic and discriminative models to complete semantic maps, overlooking the inherent uncertainty in indoor layouts and limiting their ability to generalize to unseen environments. In this work, we propose GOAL, a generative flow-based framework that models the semantic distribution of indoor environments by bridging observed regions with LLM-enriched full-scene semantic maps. During training, spatial priors inferred from large language models (LLMs) are encoded as two-dimensional Gaussian fields and injected into target maps, distilling rich contextual knowledge into the flow model and enabling more generalizable completions. Extensive experiments demonstrate that GOAL achieves state-of-the-art performance on MP3D and Gibson, and shows strong generalization in transfer settings to HM3D. Codes and pretrained models are available at https://github.com/Badi-Li/GOAL.
In recent years, blockchain has experienced widespread adoption across various industries, becoming integral to numerous enterprise applications. Concurrently, the rise of generative AI and LLMs has transformed human-computer interactions, offering advanced capabilities in understanding and generating human-like text. The introduction of the MCP has further enhanced AI integration by standardizing communication between AI systems and external data sources. Despite these advancements, there is still no standardized method for seamlessly integrating LLM applications and blockchain. To address this concern, we propose "MCC: Model Context Contracts" a novel framework that enables LLMs to interact directly with blockchain smart contracts through MCP-like protocol. This integration allows AI agents to invoke blockchain smart contracts, facilitating more dynamic and context-aware interactions between users and blockchain networks. Essentially, it empowers users to interact with blockchain systems and perform transactions using queries in natural language. Within this proposed architecture, blockchain smart contracts can function as intelligent agents capable of recognizing user input in natural language and executing the corresponding transactions. To ensure that the LLM accurately interprets natural language inputs and maps them to the appropriate MCP functions, the LLM was fine-tuned using a custom dataset comprising user inputs paired with their corresponding MCP server functions. This fine-tuning process significantly improved the platform's performance and accuracy. To validate the effectiveness of MCC, we have developed an end-to-end prototype implemented on the Rahasak blockchain with the fine-tuned Llama-4 LLM. To the best of our knowledge, this research represents the first approach to using the concept of Model Context Protocol to integrate LLMs with blockchain.
Large Language Models (LLMs) are central to reasoning, writing, and decision-support workflows, yet users lack consistent control over how they reason and express outputs. Conventional prompt engineering relies on verbose natural-language instructions, limiting reproducibility, modularity, and interpretability. This paper introduces Prompt Decorators, a declarative, composable syntax that governs LLM behavior through compact control tokens such as +++Reasoning, +++Tone(style=formal), and +++Import(topic="Systems Thinking"). Each decorator modifies a behavioral dimension, such as reasoning style, structure, or tone, without changing task content. The framework formalizes twenty core decorators organized into two functional families (Cognitive & Generative and Expressive & Systemic), each further decomposed into subcategories that govern reasoning, interaction, expression, and session-control. It defines a unified syntax, scoping model, and deterministic processing pipeline enabling predictable and auditable behavior composition. By decoupling task intent from execution behavior, Prompt Decorators create a reusable and interpretable interface for prompt design. Illustrative use cases demonstrate improved reasoning transparency, reduced prompt complexity, and standardized model behavior across domains. The paper concludes with implications for interoperability, behavioral consistency, and the development of declarative interfaces for scalable AI systems.
Large language model (LLM)-powered agents are increasingly used in recommender systems (RSs) to achieve personalized behavior modeling, where the memory mechanism plays a pivotal role in enabling the agents to autonomously explore, learn and self-evolve from real-world interactions. However, this very mechanism, serving as a contextual repository, inherently exposes an attack surface for potential adversarial manipulations. Despite its central role, the robustness of agentic RSs in the face of such threats remains largely underexplored. Previous works suffer from semantic mismatches or rely on static embeddings or pre-defined prompts, all of which are not designed for dynamic systems, especially for dynamic memory states of LLM agents. This challenge is exacerbated by the black-box nature of commercial recommenders. To tackle the above problems, in this paper, we present the first systematic investigation of memory-based vulnerabilities in LLM-powered recommender agents, revealing their security limitations and guiding efforts to strengthen system resilience and trustworthiness. Specifically, we propose a novel black-box attack framework named DrunkAgent. DrunkAgent crafts semantically meaningful adversarial textual triggers for target item promotions and introduces a series of strategies to maximize the trigger effect by corrupting the memory updates during the interactions. The triggers and strategies are optimized on a surrogate model, enabling DrunkAgent transferable and stealthy. Extensive experiments on real-world datasets across diverse agentic RSs, including collaborative filtering, retrieval augmentation and sequential recommendations, demonstrate the generalizability, transferability and stealthiness of DrunkAgent.
Data quality plays a critical role in enhancing supervised fine-tuning (SFT) for large language models (LLMs), and token-level data selection has emerged as a promising direction for its fine-grained nature. Despite their strong empirical performance, existing token-level selection methods share two key limitations: (1) requiring training or accessing an additional reference model, and (2) relying solely on loss information for token selection, which cannot well preserve semantically important tokens that are not favored by loss-based metrics. To address these challenges, we propose ssToken, a Self-modulated and Semantic-aware Token Selection approach. ssToken leverages readily accessible history models to compute the per-token loss difference with the current model, which serves as a self-modulated signal that enables the model to adaptively select tokens along its optimization trajectory, rather than relying on excess loss from an offline-trained reference model as in prior works. We further introduce a semantic-aware, attention-based token importance estimation metric, orthogonal to loss-based selection and providing complementary semantic information for more effective filtering. Extensive experiments across different model families and scales demonstrate that both self-modulated selection and semantic-aware selection alone outperform full-data fine-tuning, while their integration--ssToken--achieves synergistic gains and further surpasses prior token-level selection methods, delivering performance improvements while maintaining training efficiency.
Scaling the number of parameters and the size of training data has proven to be an effective strategy for improving large language model (LLM) performance. Yet, as these models grow increasingly powerful and widely deployed, the cost of inference has become a pressing concern. Despite its importance, the trade-off between model accuracy and inference efficiency remains underexplored. In this work, we examine how key architectural factors, hidden size, the allocation of parameters between MLP and attention (mlp-to-attention ratio), and grouped-query attention (GQA), influence both inference cost and accuracy. We introduce a conditional scaling law that augments the Chinchilla framework with architectural information, along with a search framework for identifying architectures that are simultaneously inference-efficient and accurate. To validate our approach, we train more than 200 models spanning 80M to 3B parameters and 8B to 100B training tokens, and fit the proposed conditional scaling law. Our results show that the conditional scaling law reliably predicts optimal architectural choices and that the resulting models outperform existing open-source baselines. Under the same training budget, optimized architectures achieve up to 2.1% higher accuracy and 42% greater inference throughput compared to LLaMA-3.2.
Large language models (LLMs) are increasingly used in various domains, showing impressive potential on different tasks. Recently, reasoning LLMs have been proposed to improve the \textit{reasoning} or \textit{thinking} capabilities of LLMs to solve complex problems. Despite the promising results of reasoning LLMs, enhancing the multi-step reasoning capabilities of LLMs still remains a significant challenge. While existing optimization methods have advanced the LLM reasoning capabilities, they often treat reasoning trajectories as a whole, without considering the underlying critical steps within the trajectory. In this paper, we introduce \textbf{G}uided \textbf{P}ivotal \textbf{O}ptimization (GPO), a novel fine-tuning strategy that dives into the reasoning process to enable more effective improvements. GPO first identifies the `critical step' within a reasoning trajectory - a point that the model must carefully proceed to succeed at the problem. We locate the critical step by estimating the advantage function. GPO then resets the policy to the critical step, samples the new rollout and prioritizes the learning process on those rollouts. This focus allows the model to learn more effectively from pivotal moments within the reasoning process to improve the reasoning performance. We demonstrate that GPO is a general strategy that can be integrated with various optimization methods to improve reasoning performance. Besides theoretical analysis, our experiments across challenging reasoning benchmarks show that GPO can consistently and significantly enhance the performance of existing optimization methods, showcasing its effectiveness and generalizability in improving LLM reasoning by concentrating on pivotal moments within the generation process.
Fine-tuning large language models (LLMs) using zeroth-order (ZO) optimization has emerged as a promising alternative to traditional gradient-based methods due to its reduced memory footprint requirement. However, existing ZO methods suffer from high variance in gradient estimation, leading to slow convergence and suboptimal performance on large-scale models. In this work, we propose P-GAP, a fast LLM fine-tuning approach through zeroth-order optimization with Projected Gradient-Aligned Perturbations. Specifically, we first estimate a low-dimensional gradient space and then align perturbations in projected gradients' direction within the space. This approach enables reduced the number of perturbed parameters and decreased variance, therefore accelerated convergence for LLM fine-tuning. Experiments on LLMs show that P-GAP consistently surpasses the baselines, achieving up to 6% increase in accuracy on classification tasks and up to 12% higher accuracy on generation tasks, with up to about 81% less training iterations and 70% less GPU hours. These results demonstrate that P-GAP enables fast, scalable, and resource-efficient ZO LLM fine-tuning.
We study whether self-learning can scale LLM-based agents without relying on human-curated datasets or predefined rule-based rewards. Through controlled experiments in a search-agent setting, we identify two key determinants of scalable agent training: the source of reward signals and the scale of agent task data. We find that rewards from a Generative Reward Model (GRM) outperform rigid rule-based signals for open-domain learning, and that co-evolving the GRM with the policy further boosts performance. Increasing the volume of agent task data-even when synthetically generated-substantially enhances agentic capabilities. Building on these insights, we propose \textbf{Agentic Self-Learning} (ASL), a fully closed-loop, multi-role reinforcement learning framework that unifies task generation, policy execution, and evaluation within a shared tool environment and LLM backbone. ASL coordinates a Prompt Generator, a Policy Model, and a Generative Reward Model to form a virtuous cycle of harder task setting, sharper verification, and stronger solving. Empirically, ASL delivers steady, round-over-round gains, surpasses strong RLVR baselines (e.g., Search-R1) that plateau or degrade, and continues improving under zero-labeled-data conditions, indicating superior sample efficiency and robustness. We further show that GRM verification capacity is the main bottleneck: if frozen, it induces reward hacking and stalls progress; continual GRM training on the evolving data distribution mitigates this, and a small late-stage injection of real verification data raises the performance ceiling. This work establishes reward source and data scale as critical levers for open-domain agent learning and demonstrates the efficacy of multi-role co-evolution for scalable, self-improving agents. The data and code of this paper are released at https://github.com/forangel2014/Towards-Agentic-Self-Learning
Structural pruning enhances hardware-agnostic inference efficiency for large language models (LLMs) yet often fails to maintain comparable performance. Local pruning performs efficient layer-by-layer compression but ignores global topology. Although global pruning aims to identify an optimal sparse model, intuitive methods typically adopt a two-stage paradigm that first evaluates substructure saliency and then applies global pruning, which ignores inter-structure dependencies and fails to achieve end-to-end optimization. To address these limitations, we propose T\'yr-the-Pruner, an efficient end-to-end search-based global structural pruning framework. This framework constructs a supernet by repeatedly applying local pruning across a range of sparsity ratios to each layer in an LLM, with the core goal of determining the optimal sparsity distribution under a target overall sparsity ratio. Concretely, we introduce an effective local pruning and an expectation error accumulation approach to improve supernet construction. Furthermore, we employ an iterative prune-and-search strategy with coarse-to-fine sparsity granularity to ensure efficient search convergence. Experimental results show that T\'yr-the-Pruner achieves state-of-the-art structural pruning, retaining 97% of the dense model's performance while removing a challenging 50% of Llama-3.1-70B's parameters. Code will be available at https://github.com/AMD-AGI/Tyr-the-Pruner.
Large Language Models (LLMs) increasingly rely on Chain-of-Thought (CoT) reasoning to improve accuracy on complex tasks. However, always generating lengthy reasoning traces is inefficient, leading to excessive token usage and higher inference costs. This paper introduces the Hybrid Policy Optimization (i.e., HiPO), a framework for adaptive reasoning control that enables LLMs to selectively decide when to engage in detailed reasoning (Think-on) and when to respond directly (Think-off). Specifically, HiPO combines a hybrid data pipelineproviding paired Think-on and Think-off responseswith a hybrid reinforcement learning reward system that balances accuracy and efficiency while avoiding over-reliance on detailed reasoning. Experiments across mathematics and coding benchmarks demonstrate that HiPO can substantially reduce token length while maintaining or improving accuracy. Finally, we hope HiPO a can be a principled approach for efficient adaptive reasoning, advancing the deployment of reasoning-oriented LLMs in real-world, resource-sensitive settings.
Large Language Models (LLMs) are commonly used as evaluators in various applications, but the reliability of the outcomes remains a challenge. One such challenge is using LLMs-as-judges for direct assessment, i.e., assigning scores from a specified range without any references. We first show that this challenge stems from LLM judge outputs being associated with score range bias, i.e., LLM judge outputs are highly sensitive to pre-defined score ranges, preventing the search for optimal score ranges. We also show that similar biases exist among models from the same family. We then mitigate this bias through contrastive decoding, achieving up to 11.3% relative improvement on average in Spearman correlation with human judgments across different score ranges.
Cascade systems route computational requests to smaller models when possible and defer to larger models only when necessary, offering a promising approach to balance cost and quality in LLM deployment. However, they face a fundamental challenge in open-ended text generation: determining output reliability when generation quality lies on a continuous spectrum, often with multiple valid responses. To address this, we propose semantic agreement -- meaning-level consensus between ensemble outputs -- as a training-free signal for reliable deferral. We show that when diverse model outputs agree semantically, their consensus is a stronger reliability signal than token-level confidence. Evaluated from 500M to 70B-parameter models, we find that semantic cascades match or surpass target-model quality at 40% of the cost and reduce latency by up to 60%. Our method requires no model internals, works across black-box APIs, and remains robust to model updates, making it a practical baseline for real-world LLM deployment.
LLMs are increasingly used to design reward functions based on human preferences in Reinforcement Learning (RL). We focus on LLM-designed rewards for Restless Multi-Armed Bandits, a framework for allocating limited resources among agents. In applications such as public health, this approach empowers grassroots health workers to tailor automated allocation decisions to community needs. In the presence of multiple agents, altering the reward function based on human preferences can impact subpopulations very differently, leading to complex tradeoffs and a multi-objective resource allocation problem. We are the first to present a principled method termed Social Choice Language Model for dealing with these tradeoffs for LLM-designed rewards for multiagent planners in general and restless bandits in particular. The novel part of our model is a transparent and configurable selection component, called an adjudicator, external to the LLM that controls complex tradeoffs via a user-selected social welfare function. Our experiments demonstrate that our model reliably selects more effective, aligned, and balanced reward functions compared to purely LLM-based approaches.
Inference-time computation is a critical yet challenging paradigm for enhancing the reasoning performance of large language models (LLMs). While existing strategies improve reasoning stability and consistency, they suffer from notable limitations: self-correction often reinforces the model's initial biases, and Multi-Agent Collaboration (MAC) often fails due to the lack of efficient coordination mechanisms, leading to collective errors. Although high-performing verifiers can detect reasoning errors, making them reliable requires substantial training. To address these challenges, we introduce a novel inference-time framework, Adaptive Coopetition (AdCo), in which LLM agents utilize an adaptive, UCB-based "coopetition" mechanism. At each round, agents leverage coarse verifier signals to determine whether to collaborate or compete, and iteratively refine their reasoning based on peer feedback. Without relying on high-performance verifiers, our adaptive strategy achieves significant performance gains on mathematical reasoning benchmarks, yielding a 20% relative improvement over baselines on the more challenging dataset. Our approach remains robust and consistent in terms of accuracy under different sample sizes and configurations. This adaptive, signal-guided "coopetition" framework enhances reasoning robustness by leveraging both model knowledge diversity and reasoning trace measures, while also promoting uncertainty-driven exploration, especially when participants have comparable capabilities. From this perspective, our work offers a fresh lens on inference-time computation and paves the way for more resilient multi-agent LLM systems. Our code is available at: https://github.com/AdCo-Research/adaptive-coopetition.
Transformer-based Language Models' computation and memory overhead increase quadratically as a function of sequence length. The quadratic cost poses challenges when employing LLMs for processing long sequences. In this work, we introduce \ourmodelacronym~(Extend at Test-Time), method for extending the context length of short context Transformer-based LLMs, with constant memory requirement and linear computation overhead. ETT enable the extension of the context length at test-time by efficient fine-tuning the model's parameters on the input context, chunked into overlapping small subsequences. We evaluate ETT on LongBench by extending the context length of GPT-Large and Phi-2 up to 32 times, increasing from 1k to 32k tokens. This results in up to a 30 percent improvement in the model's accuracy. We also study how context can be stored in LLM's weights effectively and efficiently. Through a detailed ablation study, we examine which Transformer modules are most beneficial to fine-tune at test-time. Interestingly, we find that fine-tuning the second layer of the FFNs is more effective than full fine-tuning, leading to a further improvement in the models' accuracy.
Today's Internet infrastructure is centered around content retrieval over HTTP, with middleboxes (e.g., HTTP proxies) playing a crucial role in performance, security, and cost-effectiveness. We envision a future where Internet communication will be dominated by "prompts" sent to generative AI models. For this, we will need proxies that provide similar functions to HTTP proxies (e.g., caching, routing, compression) while dealing with unique challenges and opportunities of prompt-based communication. As a first step toward supporting prompt-based communication, we present LLMBridge, an LLM proxy designed for cost-conscious users, such as those in developing regions and education (e.g., students, instructors). LLMBridge supports three key optimizations: model selection (routing prompts to the most suitable model), context management (intelligently reducing the amount of context), and semantic caching (serving prompts using local models and vector databases). These optimizations introduce trade-offs between cost and quality, which applications navigate through a high-level, bidirectional interface. As case studies, we deploy LLMBridge in two cost-sensitive settings: a WhatsApp-based Q&A service and a university classroom environment. The WhatsApp service has been live for over twelve months, serving 100+ users and handling more than 14.7K requests. In parallel, we exposed LLMBridge to students across three computer science courses over a semester, where it supported diverse LLM-powered applications - such as reasoning agents and chatbots - and handled an average of 500 requests per day. We report on deployment experiences across both settings and use the collected workloads to benchmark the effectiveness of various cost-optimization strategies, analyzing their trade-offs in cost, latency, and response quality.
We investigate how peer pressure influences the opinions of Large Language Model (LLM) agents across a spectrum of cognitive commitments by embedding them in social networks where they update opinions based on peer perspectives. Our findings reveal key departures from traditional conformity assumptions. First, agents follow a sigmoid curve: stable at low pressure, shifting sharply at threshold, and saturating at high. Second, conformity thresholds vary by model: Gemini 1.5 Flash requires over 70% peer disagreement to flip, whereas ChatGPT-4o-mini shifts with a dissenting minority. Third, we uncover a fundamental "persuasion asymmetry," where shifting an opinion from affirmative-to-negative requires a different cognitive effort than the reverse. This asymmetry results in a "dual cognitive hierarchy": the stability of cognitive constructs inverts based on the direction of persuasion. For instance, affirmatively-held core values are robust against opposition but easily adopted from a negative stance, a pattern that inverts for other constructs like attitudes. These dynamics echoing complex human biases like negativity bias, prove robust across different topics and discursive frames (moral, economic, sociotropic). This research introduces a novel framework for auditing the emergent socio-cognitive behaviors of multi-agent AI systems, demonstrating their decision-making is governed by a fluid, context-dependent architecture, not a static logic.
Curriculum learning (CL) - ordering training data from easy to hard - has become a popular strategy for improving reasoning in large language models (LLMs). Yet prior work employs disparate difficulty metrics and training setups, leaving open fundamental questions: When does curriculum help? Which direction - forward or reverse - is better? And does the answer depend on what we measure? We address these questions through a unified offline evaluation framework that decomposes curriculum difficulty into five complementary dimensions: Problem Difficulty, Model Surprisal, Confidence Margin, Predictive Uncertainty, and Decision Variability. Through controlled post-training experiments on mathematical reasoning benchmarks with Llama3.1-8B, Mistral-7B, and Gemma3-4B, we find that (i) no curriculum strategy dominates universally - the relative effectiveness of forward versus reverse CL depends jointly on model capability and task complexity; (ii) even within a single metric, samples at different difficulty levels produce distinct gains depending on task demands; and (iii) task-aligned curricula focus on shaping the model's final representations and generalization, whereas inner-state curricula modulate internal states such as confidence and uncertainty. Our findings challenge the notion of a universal curriculum strategy and offer actionable guidance across model and task regimes, with some metrics indicating that prioritizing decision-uncertain samples can further enhance learning outcomes.
While vision-language models (VLMs) have advanced into detailed image description, evaluation remains a challenge. Standard metrics (e.g. CIDEr, SPICE) were designed for short texts and tuned to recognize errors that are now uncommon, such as object misidentification. In contrast, long texts require sensitivity to attribute and relation attachments and scores that localize errors to particular text spans. In this work, we introduce PoSh, a metric for detailed image description that uses scene graphs as structured rubrics to guide LLMs-as-a-Judge, producing aggregate scores grounded in fine-grained errors (e.g. mistakes in compositional understanding). PoSh is replicable, interpretable and a better proxy for human raters than existing metrics (including GPT4o-as-a-Judge). To validate PoSh, we introduce a challenging new dataset, DOCENT. This novel benchmark contains artwork, paired with expert-written references, and model-generated descriptions, augmented with granular and coarse judgments of their quality from art history students. Thus, DOCENT enables evaluating both detailed image description metrics and detailed image description itself in a challenging new domain. We show that PoSh achieves stronger correlations (+0.05 Spearman $\rho$) with the human judgments in DOCENT than the best open-weight alternatives, is robust to image type (using CapArena, an existing dataset of web imagery) and is a capable reward function, outperforming standard supervised fine-tuning. Then, using PoSh, we characterize the performance of open and closed models in describing the paintings, sketches and statues in DOCENT and find that foundation models struggle to achieve full, error-free coverage of images with rich scene dynamics, establishing a demanding new task to gauge VLM progress. Through both PoSh and DOCENT, we hope to enable advances in important areas such as assistive text generation.
We introduce ALHD, the first large-scale comprehensive Arabic dataset explicitly designed to distinguish between human- and LLM-generated texts. ALHD spans three genres (news, social media, reviews), covering both MSA and dialectal Arabic, and contains over 400K balanced samples generated by three leading LLMs and originated from multiple human sources, which enables studying generalizability in Arabic LLM-genearted text detection. We provide rigorous preprocessing, rich annotations, and standardized balanced splits to support reproducibility. In addition, we present, analyze and discuss benchmark experiments using our new dataset, in turn identifying gaps and proposing future research directions. Benchmarking across traditional classifiers, BERT-based models, and LLMs (zero-shot and few-shot) demonstrates that fine-tuned BERT models achieve competitive performance, outperforming LLM-based models. Results are however not always consistent, as we observe challenges when generalizing across genres; indeed, models struggle to generalize when they need to deal with unseen patterns in cross-genre settings, and these challenges are particularly prominent when dealing with news articles, where LLM-generated texts resemble human texts in style, which opens up avenues for future research. ALHD establishes a foundation for research related to Arabic LLM-detection and mitigating risks of misinformation, academic dishonesty, and cyber threats.
Evaluating Large Language Models (LLMs) for mental health support is challenging due to the emotionally and cognitively complex nature of therapeutic dialogue. Existing benchmarks are limited in scale, reliability, often relying on synthetic or social media data, and lack frameworks to assess when automated judges can be trusted. To address the need for large-scale dialogue datasets and judge reliability assessment, we introduce two benchmarks that provide a framework for generation and evaluation. MentalBench-100k consolidates 10,000 one-turn conversations from three real scenarios datasets, each paired with nine LLM-generated responses, yielding 100,000 response pairs. MentalAlign-70k}reframes evaluation by comparing four high-performing LLM judges with human experts across 70,000 ratings on seven attributes, grouped into Cognitive Support Score (CSS) and Affective Resonance Score (ARS). We then employ the Affective Cognitive Agreement Framework, a statistical methodology using intraclass correlation coefficients (ICC) with confidence intervals to quantify agreement, consistency, and bias between LLM judges and human experts. Our analysis reveals systematic inflation by LLM judges, strong reliability for cognitive attributes such as guidance and informativeness, reduced precision for empathy, and some unreliability in safety and relevance. Our contributions establish new methodological and empirical foundations for reliable, large-scale evaluation of LLMs in mental health. We release the benchmarks and codes at: https://github.com/abeerbadawi/MentalBench/
As large language models (LLMs) are increasingly used in human-AI interactions, their social reasoning capabilities in interpersonal contexts are critical. We introduce SCRIPTS, a 1k-dialogue dataset in English and Korean, sourced from movie scripts. The task involves evaluating models' social reasoning capability to infer the interpersonal relationships (e.g., friends, sisters, lovers) between speakers in each dialogue. Each dialogue is annotated with probabilistic relational labels (Highly Likely, Less Likely, Unlikely) by native (or equivalent) Korean and English speakers from Korea and the U.S. Evaluating nine models on our task, current proprietary LLMs achieve around 75-80% on the English dataset, whereas their performance on Korean drops to 58-69%. More strikingly, models select Unlikely relationships in 10-25% of their responses. Furthermore, we find that thinking models and chain-of-thought prompting, effective for general reasoning, provide minimal benefits for social reasoning and occasionally amplify social biases. Our findings reveal significant limitations in current LLMs' social reasoning capabilities, highlighting the need for efforts to develop socially-aware language models.
FlexiDataGen: An Adaptive LLM Framework for Dynamic Semantic Dataset Generation in Sensitive Domains
Dataset availability and quality remain critical challenges in machine learning, especially in domains where data are scarce, expensive to acquire, or constrained by privacy regulations. Fields such as healthcare, biomedical research, and cybersecurity frequently encounter high data acquisition costs, limited access to annotated data, and the rarity or sensitivity of key events. These issues-collectively referred to as the dataset challenge-hinder the development of accurate and generalizable machine learning models in such high-stakes domains. To address this, we introduce FlexiDataGen, an adaptive large language model (LLM) framework designed for dynamic semantic dataset generation in sensitive domains. FlexiDataGen autonomously synthesizes rich, semantically coherent, and linguistically diverse datasets tailored to specialized fields. The framework integrates four core components: (1) syntactic-semantic analysis, (2) retrieval-augmented generation, (3) dynamic element injection, and (4) iterative paraphrasing with semantic validation. Together, these components ensure the generation of high-quality, domain-relevant data. Experimental results show that FlexiDataGen effectively alleviates data shortages and annotation bottlenecks, enabling scalable and accurate machine learning model development.
Generative AI and large language models (LLMs) have shown strong capabilities in code understanding, but their use in cybersecurity, particularly for malware detection and analysis, remains limited. Existing detection systems often fail to generalize to obfuscated or previously unseen threats, underscoring the need for more adaptable and explainable models. To address this challenge, we introduce XGen-Q, a domain-adapted LLM built on the Qwen-Coder architecture and pretrained on a large-scale corpus of over one million malware samples, spanning both source and assembly code. XGen-Q uses a multi-stage prompt strategy combined with retrieval-augmented generation (RAG) to deliver reliable malware identification and detailed forensic reporting, even in the presence of complex code obfuscation. To further enhance generalization, we design a training pipeline that systematically exposes the model to diverse obfuscation patterns. Experimental results show that XGen-Q achieves significantly lower perplexity than competitive baselines and exhibits strong performance on novel malware samples, demonstrating the promise of LLM-based approaches for interpretable and robust malware analysis.
Oversensitivity occurs when language models defensively reject prompts that are actually benign. This behavior not only disrupts user interactions but also obscures the boundary between harmful and harmless content. Existing benchmarks rely on static datasets that degrade overtime as models evolve, leading to data contamination and diminished evaluative power. To address this, we develop a framework that dynamically generates model-specific challenging datasets, capturing emerging defensive patterns and aligning with each model's unique behavior. Building on this approach, we construct OVERBENCH, a benchmark that aggregates these datasets across diverse LLM families, encompassing 450,000 samples from 25 models. OVERBENCH provides a dynamic and evolving perspective on oversensitivity, allowing for continuous monitoring of defensive triggers as models advance, highlighting vulnerabilities that static datasets overlook.
Large Language Models (LLMs) offer promising capabilities for tackling complex reasoning tasks, including optimization problems. However, existing methods either rely on prompt engineering, which leads to poor generalization across problem types, or require costly supervised training. We introduce SolverLLM, a training-free framework that leverages test-time scaling to solve diverse optimization problems. Rather than solving directly, SolverLLM generates mathematical formulations and translates them into solver-ready code, guided by a novel Monte Carlo Tree Search (MCTS) strategy. To enhance the search process, we modify classical MCTS with (1) dynamic expansion for adaptive formulation generation, (2) prompt backpropagation to guide exploration via outcome-driven feedback, and (3) uncertainty backpropagation to incorporate reward reliability into decision-making. Experiments on six standard benchmark datasets demonstrate that SolverLLM outperforms both prompt-based and learning-based baselines, achieving strong generalization without additional training.
Evaluating the creative capabilities of large language models (LLMs) in complex tasks often requires human assessments that are difficult to scale. We introduce a novel, scalable methodology for evaluating LLM story generation by analyzing underlying social structures in narratives as signed character networks. To demonstrate its effectiveness, we conduct a large-scale comparative analysis using networks from over 1,200 stories, generated by four leading LLMs (GPT-4o, GPT-4o mini, Gemini 1.5 Pro, and Gemini 1.5 Flash) and a human-written corpus. Our findings, based on network properties like density, clustering, and signed edge weights, show that LLM-generated stories consistently exhibit a strong bias toward tightly-knit, positive relationships, which aligns with findings from prior research using human assessment. Our proposed approach provides a valuable tool for evaluating limitations and tendencies in the creative storytelling of current and future LLMs.
Course syllabi set the tone and expectations for courses, shaping the learning experience for both students and instructors. In computing courses, especially those addressing fairness and ethics in artificial intelligence (AI), machine learning (ML), and algorithmic design, it is imperative that we understand how approaches to navigating barriers to fair outcomes are being addressed.These expectations should be inclusive, transparent, and grounded in promoting critical thinking. Syllabus analysis offers a way to evaluate the coverage, depth, practices, and expectations within a course. Manual syllabus evaluation, however, is time-consuming and prone to inconsistency. To address this, we developed a justice-oriented scoring rubric and asked a large language model (LLM) to review syllabi through a multi-perspective role simulation. Using this rubric, we evaluated 24 syllabi from four perspectives: instructor, departmental chair, institutional reviewer, and external evaluator. We also prompted the LLM to identify thematic trends across the courses. Findings show that multiperspective evaluation aids us in noting nuanced, role-specific priorities, leveraging them to fill hidden gaps in curricula design of AI/ML and related computing courses focused on fairness and ethics. These insights offer concrete directions for improving the design and delivery of fairness, ethics, and justice content in such courses.
Reinforcement learning (RL) has recently become the core paradigm for aligning and strengthening large language models (LLMs). Yet, applying RL in off-policy settings--where stale data from past policies are used for training--improves sample efficiency, but remains challenging: policy entropy declines sharply, optimization often becomes unstable and may even collapse. Through theoretical and empirical analysis, we identify two key insights: (i) an imbalance in optimization, where negative-advantage samples dominate the policy gradient, suppressing useful behaviors and risking gradient explosions; and (ii) the derived Entropy-Clip Rule, which reveals that the fixed clipping mechanism in PPO-like objectives systematically blocks entropy-increasing updates, thereby driving the policy toward over-exploitation at the expense of exploration. Building on these insights, we propose BAlanced Policy Optimization with Adaptive Clipping (BAPO), a simple yet effective method that dynamically adjusts clipping bounds to adaptively re-balance positive and negative contributions, preserve entropy, and stabilize RL optimization. Across diverse off-policy scenarios--including sample replay and partial rollout--BAPO achieves fast, stable, and data-efficient training. On AIME 2024 and AIME 2025 benchmarks, our 7B BAPO model surpasses open-source counterparts such as SkyWork-OR1-7B, while our 32B BAPO model not only achieves state-of-the-art results among models of the same scale but also outperforms leading proprietary systems like o3-mini and Gemini-2.5-Flash-Thinking.
Domain-adapted open-weight large language models (LLMs) offer promising healthcare applications, from queryable knowledge bases to multimodal assistants, with the crucial advantage of local deployment for privacy preservation. However, optimal adaptation strategies, evaluation methodologies, and performance relative to general-purpose LLMs remain poorly characterized. We investigated these questions in electrocardiography, an important area of cardiovascular medicine, by finetuning open-weight models on domain-specific literature and implementing a multi-layered evaluation framework comparing finetuned models, retrieval-augmented generation (RAG), and Claude Sonnet 3.7 as a representative general-purpose model. Finetuned Llama 3.1 70B achieved superior performance on multiple-choice evaluations and automatic text metrics, ranking second to Claude 3.7 in LLM-as-a-judge assessments. Human expert evaluation favored Claude 3.7 and RAG approaches for complex queries. Finetuned models significantly outperformed their base counterparts across nearly all evaluation modes. Our findings reveal substantial performance heterogeneity across evaluation methodologies, underscoring assessment complexity. Nevertheless, domain-specific adaptation through finetuning and RAG achieves competitive performance with proprietary models, supporting the viability of privacy-preserving, locally deployable clinical solutions.
Growing evidence suggests that large language models do not use their depth uniformly, yet we still lack a fine-grained understanding of their layer-wise prediction dynamics. In this paper, we trace the intermediate representations of several open-weight models during inference and reveal a structured and nuanced use of depth. Specifically, we propose a "Guess-then-Refine" framework that explains how LLMs internally structure their computations to make predictions. We first show that the top-ranked predictions in early LLM layers are composed primarily of high-frequency tokens, which act as statistical guesses proposed by the model early on due to the lack of appropriate contextual information. As contextual information develops deeper into the model, these initial guesses get refined into contextually appropriate tokens. Even high-frequency token predictions from early layers get refined >70% of the time, indicating that correct token prediction is not "one-and-done". We then go beyond frequency-based prediction to examine the dynamic usage of layer depth across three case studies. (i) Part-of-speech analysis shows that function words are, on average, the earliest to be predicted correctly. (ii) Fact recall task analysis shows that, in a multi-token answer, the first token requires more computational depth than the rest. (iii) Multiple-choice task analysis shows that the model identifies the format of the response within the first half of the layers, but finalizes its response only toward the end. Together, our results provide a detailed view of depth usage in LLMs, shedding light on the layer-by-layer computations that underlie successful predictions and providing insights for future works to improve computational efficiency in transformer-based models.
Large Language Models (LLMs) are intended to reflect human linguistic competencies. But humans have access to a broad and embodied context, which is key in detecting and resolving linguistic ambiguities, even in isolated text spans. A foundational case of semantic ambiguity is found in the task of coreference resolution: how is a pronoun related to an earlier person mention? This capability is implicit in nearly every downstream task, and the presence of ambiguity at this level can alter performance significantly. We show that LLMs can achieve good performance with minimal prompting in both coreference disambiguation and the detection of ambiguity in coreference, however, they cannot do both at the same time. We present the CORRECT-DETECT trade-off: though models have both capabilities and deploy them implicitly, successful performance balancing these two abilities remains elusive.
Large Language Models (LLMs) have revolutionized natural language processing but remain vulnerable to jailbreak attacks, especially multi-turn jailbreaks that distribute malicious intent across benign exchanges and bypass alignment mechanisms. Existing approaches often explore the adversarial space poorly, rely on hand-crafted heuristics, or lack systematic query refinement. We present NEXUS (Network Exploration for eXploiting Unsafe Sequences), a modular framework for constructing, refining, and executing optimized multi-turn attacks. NEXUS comprises: (1) ThoughtNet, which hierarchically expands a harmful intent into a structured semantic network of topics, entities, and query chains; (2) a feedback-driven Simulator that iteratively refines and prunes these chains through attacker-victim-judge LLM collaboration using harmfulness and semantic-similarity benchmarks; and (3) a Network Traverser that adaptively navigates the refined query space for real-time attacks. This pipeline uncovers stealthy, high-success adversarial paths across LLMs. On several closed-source and open-source LLMs, NEXUS increases attack success rate by 2.1% to 19.4% over prior methods. Code: https://github.com/inspire-lab/NEXUS
With the emergence of wearable devices and other embedded systems, deploying large language models (LLMs) on edge platforms has become an urgent need. However, this is challenging because of their high computational and memory demands. Although recent low-bit quantization methods (e.g., BitNet, DeepSeek) compress weights to as low as 1.58~bits with minimal accuracy loss, edge deployment is still constrained by limited on-chip resources, power budgets, and the often-neglected long latency of the prefill stage. We present \textbf{TeLLMe}, the first table-lookup-based ternary LLM accelerator for low-power edge FPGAs that fully supports both prefill and autoregressive decoding using 1.58-bit weights and 8-bit activations. TeLLMe incorporates several novel techniques, including (1) a table-lookup-based ternary matrix multiplication (TLMM) engine utilizing grouped activations and online precomputation for low resource utilization and high throughput; (2) a fine-grained analytic URAM-based weight buffer management scheme for efficient loading and compute engine access; (3) a streaming dataflow architecture that fuses floating-point element-wise operations with linear computations to hide latency; (4) a reversed-reordered prefill stage attention with fused attention operations for high memory efficiency; and (5) a resource-efficient specialized decoding stage attention. Under a 5~W power budget, TeLLMe delivers up to 25~tokens/s decoding throughput and 0.45--0.96~s time-to-first-token (TTFT) for 64--128 token prompts, marking a significant energy-efficiency advancement in LLM inference on edge FPGAs.
We present a simple, self-help online supervised finetuning (OSFT) paradigm for LLM reasoning. In this paradigm, the model generates its own responses and is immediately finetuned on this self-generated data. OSFT is a highly efficient training strategy for LLM reasoning, as it is reward-free and uses just one rollout by default. Experiment results show that OSFT achieves downstream performance on challenging mathematical reasoning tasks comparable to strong reinforcement learning with verifiable rewards (RLVR) methods such as GRPO. Our ablation study further demonstrates the efficiency and robustness of OSFT. The major mechanism of OSFT lies in facilitating the model's own existing preference (latent knowledge) learned from pretraining, which leads to reasoning ability improvement. We believe that OSFT offers an efficient and promising alternative to more complex, reward-based training paradigms. Our code is available at https://github.com/ElementQi/OnlineSFT.
Large language model (LLM) based search agents iteratively generate queries, retrieve external information, and reason to answer open-domain questions. While researchers have primarily focused on improving their utility, their safety behaviors remain underexplored. In this paper, we first evaluate search agents using red-teaming datasets and find that they are more likely to produce harmful outputs than base LLMs. For example, when asked "How can I track someone's location without their consent?", a base model refuses, whereas a search agent designed to retrieve and cite sources may lower its refusal threshold, fetch documents (e.g., court cases), and, once appended, synthesize them into an informative yet unsafe summary. We further show that utility-oriented fine-tuning intensifies this risk, motivating joint alignment of safety and utility. We present SafeSearch, a multi-objective reinforcement learning approach that couples a final-output safety/utility reward with a novel query-level shaping term that penalizes unsafe queries and rewards safe ones. Experiments show that SafeSearch reduces agent harmfulness by over 70% across three red-teaming datasets while producing safe, helpful responses, and matches the QA performance of a utility-only finetuned agent; further analyses confirm the effectiveness of the query-level reward in jointly improving safety and utility.
LLM-powered Multi-Agent Systems (LLM-MAS) unlock new potentials in distributed reasoning, collaboration, and task generalization but also introduce additional risks due to unguaranteed agreement, cascading uncertainty, and adversarial vulnerabilities. We argue that ensuring responsible behavior in such systems requires a paradigm shift: from local, superficial agent-level alignment to global, systemic agreement. We conceptualize responsibility not as a static constraint but as a lifecycle-wide property encompassing agreement, uncertainty, and security, each requiring the complementary integration of subjective human-centered values and objective verifiability. Furthermore, a dual-perspective governance framework that combines interdisciplinary design with human-AI collaborative oversight is essential for tracing and ensuring responsibility throughout the lifecycle of LLM-MAS. Our position views LLM-MAS not as loose collections of agents, but as unified, dynamic socio-technical systems that demand principled mechanisms to support each dimension of responsibility and enable ethically aligned, verifiably coherent, and resilient behavior for sustained, system-wide agreement.
With the broad availability of large language models and their ability to generate vast outputs using varied prompts and configurations, determining the best output for a given task requires an intensive evaluation process, one where machine learning practitioners must decide how to assess the outputs and then carefully carry out the evaluation. This process is both time-consuming and costly. As practitioners work with an increasing number of models, they must now evaluate outputs to determine which model and prompt performs best for a given task. LLMs are increasingly used as evaluators to filter training data, evaluate model performance, assess harms and risks, or assist human evaluators with detailed assessments. We present EvalAssist, a framework that simplifies the LLM-as-a-judge workflow. The system provides an online criteria development environment, where users can interactively build, test, and share custom evaluation criteria in a structured and portable format. We support a set of LLM-based evaluation pipelines that leverage off-the-shelf LLMs and use a prompt-chaining approach we developed and contributed to the UNITXT open-source library. Additionally, our system also includes specially trained evaluators to detect harms and risks in LLM outputs. We have deployed the system internally in our organization with several hundreds of users.
The original CLIP text encoder is limited by a maximum input length of 77 tokens, which hampers its ability to effectively process long texts and perform fine-grained semantic understanding. In addition, the CLIP text encoder lacks support for multilingual inputs. All these limitations significantly restrict its applicability across a broader range of tasks. Recent studies have attempted to replace the CLIP text encoder with an LLM-based embedder to enhance its ability in processing long texts, multilingual understanding, and fine-grained semantic comprehension. However, because the representation spaces of LLMs and the vision-language space of CLIP are pretrained independently without alignment priors, direct alignment using contrastive learning can disrupt the intrinsic vision-language alignment in the CLIP image encoder, leading to an underutilization of the knowledge acquired during pre-training. To address this challenge, we propose ProCLIP, a curriculum learning-based progressive vision-language alignment framework to effectively align the CLIP image encoder with an LLM-based embedder. Specifically, ProCLIP first distills knowledge from CLIP's text encoder into the LLM-based embedder to leverage CLIP's rich pretrained knowledge while establishing initial alignment between the LLM embedder and CLIP image encoder. Subsequently, ProCLIP further aligns the CLIP image encoder with the LLM-based embedder through image-text contrastive tuning, employing self-distillation regularization to avoid overfitting. To achieve a more effective alignment, instance semantic alignment loss and embedding structure alignment loss are employed during representation inheritance and contrastive tuning. The Code is available at https://github.com/VisionXLab/ProCLIP
[Context] Large Language Models (LLMs) rely on domain-specific datasets to achieve robust performance across training and inference stages. However, in Requirements Engineering (RE), data scarcity remains a persistent limitation reported in surveys and mapping studies. [Question/Problem] Although there are multiple datasets supporting LLM-based RE tasks (LLM4RE), they are fragmented and poorly characterized, limiting reuse and comparability. This research addresses the limited visibility and characterization of datasets used in LLM4RE. We investigate which public datasets are employed, how they can be systematically characterized, and which RE tasks and dataset descriptors remain under-represented. [Ideas/Results] To address this, we conduct a systematic mapping study to identify and analyse datasets used in LLM4RE research. A total of 62 publicly available datasets are referenced across 43 primary studies. Each dataset is characterized along descriptors such as artifact type, granularity, RE stage, task, domain, and language. Preliminary findings show multiple research gaps, including limited coverage for elicitation tasks, scarce datasets for management activities beyond traceability, and limited multilingual availability. [Contribution] This research preview offers a public catalogue and structured characterization scheme to support dataset selection, comparison, and reuse in LLM4RE research. Future work will extend the scope to grey literature, as well as integration with open dataset and benchmark repositories.
Large Language Models (LLMs) are adept at generating responses based on information within their context. While this ability is useful for interacting with structured data like code files, another popular method, Retrieval-Augmented Generation (RAG), retrieves relevant documents to augment the model's in-context learning. However, it is not well-explored how to best represent this retrieved knowledge for generating responses on structured data, particularly hierarchical structures like trees. In this work, we propose a novel bottom-up method to linearize knowledge from tree-like structures (like a GitHub repository) by generating implicit, aggregated summaries at each hierarchical level. This approach enables the knowledge to be stored in a knowledge base and used directly with RAG. We then compare our method to using RAG on raw, unstructured code, evaluating the accuracy and quality of the generated responses. Our results show that while response quality is comparable across both methods, our approach generates over 68% fewer documents in the retriever, a significant gain in efficiency. This finding suggests that leveraging implicit, linearized knowledge may be a highly effective and scalable strategy for handling complex, hierarchical data structures.
The proliferation of misinformation necessitates scalable, automated fact-checking solutions. Yet, current benchmarks often overlook multilingual and topical diversity. This paper introduces a novel, dynamically extensible data set that includes 61,514 claims in multiple languages and topics, extending existing datasets up to 2024. Through a comprehensive evaluation of five prominent Large Language Models (LLMs), including GPT-4o, GPT-3.5 Turbo, LLaMA 3.1, and Mixtral 8x7B, we identify significant performance gaps between different languages and topics. While overall GPT-4o achieves the highest accuracy, it declines to classify 43% of claims. Across all models, factual-sounding claims are misclassified more often than opinions, revealing a key vulnerability. These findings underscore the need for caution and highlight challenges in deploying LLM-based fact-checking systems at scale.
What if artificial agents could not just communicate, but also evolve, adapt, and reshape their worlds in ways we cannot fully predict? With llm now powering multi-agent systems and social simulations, we are witnessing new possibilities for modeling open-ended, ever-changing environments. Yet, most current simulations remain constrained within static sandboxes, characterized by predefined tasks, limited dynamics, and rigid evaluation criteria. These limitations prevent them from capturing the complexity of real-world societies. In this paper, we argue that static, task-specific benchmarks are fundamentally inadequate and must be rethought. We critically review emerging architectures that blend llm with multi-agent dynamics, highlight key hurdles such as balancing stability and diversity, evaluating unexpected behaviors, and scaling to greater complexity, and introduce a fresh taxonomy for this rapidly evolving field. Finally, we present a research roadmap centered on open-endedness, continuous co-evolution, and the development of resilient, socially aligned AI ecosystems. We call on the community to move beyond static paradigms and help shape the next generation of adaptive, socially-aware multi-agent simulations.
This research proposes a systematic, large language model (LLM) approach for extracting product and service attributes, features, and associated sentiments from customer reviews. Grounded in marketing theory, the framework distinguishes perceptual attributes from actionable features, producing interpretable and managerially actionable insights. We apply the methodology to 20,000 Yelp reviews of Starbucks stores and evaluate eight prompt variants on a random subset of reviews. Model performance is assessed through agreement with human annotations and predictive validity for customer ratings. Results show high consistency between LLMs and human coders and strong predictive validity, confirming the reliability of the approach. Human coders required a median of six minutes per review, whereas the LLM processed each in two seconds, delivering comparable insights at a scale unattainable through manual coding. Managerially, the analysis identifies attributes and features that most strongly influence customer satisfaction and their associated sentiments, enabling firms to pinpoint "joy points," address "pain points," and design targeted interventions. We demonstrate how structured review data can power an actionable marketing dashboard that tracks sentiment over time and across stores, benchmarks performance, and highlights high-leverage features for improvement. Simulations indicate that enhancing sentiment for key service features could yield 1-2% average revenue gains per store.
No-Code Development Platforms (NCDPs) empower non-technical end users to build applications tailored to their specific demands without writing code. While NCDPs lower technical barriers, users still require some technical knowledge, e.g., to structure process steps or define event-action rules. Large Language Models (LLMs) offer a promising solution to further reduce technical requirements by supporting natural language interaction and dynamic code generation. By integrating LLM, NCDPs can be more accessible to non-technical users, enabling application development truly without requiring any technical expertise. Despite growing interest in LLM-powered NCDPs, a systematic investigation into the factors influencing LLM suitability and performance remains absent. Understanding these factors is critical to effectively leveraging LLMs capabilities and maximizing their impact. In this paper, we investigate key factors influencing the effectiveness of LLMs in supporting end-user application development within NCDPs. By conducting comprehensive experiments, we evaluate the impact of four key factors, i.e., model selection, prompt language, training data background, and an error-informed few-shot setup, on the quality of generated applications. Specifically, we selected a range of LLMs based on their architecture, scale, design focus, and training data, and evaluated them across four real-world smart home automation scenarios implemented on a representative open-source LLM-powered NCDP. Our findings offer practical insights into how LLMs can be effectively integrated into NCDPs, informing both platform design and the selection of suitable LLMs for end-user application development.