llm - 2025_11
Navigation
Papers
Data contamination poses a significant challenge to the fairness of LLM evaluations in natural language processing tasks by inadvertently exposing models to test data during training. Current studies attempt to mitigate this issue by modifying existing datasets or generating new ones from freshly collected information. However, these methods fall short of ensuring contamination-resilient evaluation, as they fail to fully eliminate pre-existing knowledge from models or preserve the semantic complexity of the original datasets. To address these limitations, we propose \textbf{CoreEval}, a \textbf{Co}ntamination-\textbf{re}silient \textbf{Eval}uation strategy for automatically updating data with real-world knowledge. This approach begins by extracting entity relationships from the original data and leveraging the GDELT database to retrieve relevant, up-to-date knowledge. The retrieved knowledge is then recontextualized and integrated with the original data, which is refined and restructured to ensure semantic coherence and enhanced task relevance. Ultimately, a robust data reflection mechanism is employed to iteratively verify and refine labels, ensuring consistency between the updated and original datasets. Extensive experiments on updated datasets validate the robustness of CoreEval, demonstrating its effectiveness in mitigating performance overestimation caused by data contamination.
High quality kernels are critical for reducing training and inference costs of Large Language Models (LLMs), yet they traditionally require significant expertise in hardware architecture and software optimization. While recent advances in LLM-based code generation show promise for complex optimization, existing methods struggle with the vast optimization space due to insufficient hardware domain knowledge, failing to effectively balance exploration and exploitation. We present KernelBand, a novel framework that formulates kernel optimization as a hierarchical multi-armed bandit problem, enabling LLM agents to strategically navigate the optimization space by treating kernel selection and optimization strategy application as sequential decision-making processes. Our approach leverages hardware profiling information to identify promising optimization strategies and employs runtime behavior clustering to reduce exploration overhead across kernel candidates. Extensive experiments on TritonBench demonstrate that KernelBand significantly outperforms state-of-the-art methods, achieving superior performance with fewer tokens while exhibiting consistent improvement without saturation as computational resources increase.
We present a novel framework that integrates Large Language Models (LLMs) into the Git bisect process for semantic fault localization. Traditional bisect assumes deterministic predicates and binary failure states assumptions often violated in modern software development due to flaky tests, nonmonotonic regressions, and semantic divergence from upstream repositories. Our system augments bisect traversal with structured chain of thought reasoning, enabling commit by commit analysis under noisy conditions. We evaluate multiple open source and proprietary LLMs for their suitability and fine tune DeepSeekCoderV2 using QLoRA on a curated dataset of semantically labeled diffs. We adopt a weak supervision workflow to reduce annotation overhead, incorporating human in the loop corrections and self consistency filtering. Experiments across multiple open source projects show a 6.4 point absolute gain in success rate from 74.2 to 80.6 percent, leading to significantly fewer failed traversals and by experiment up to 2x reduction in average bisect time. We conclude with discussions on temporal reasoning, prompt design, and finetuning strategies tailored for commit level behavior analysis.
Discovering effective predictive signals, or ``alphas,'' from financial data with high dimensionality and extremely low signal-to-noise ratio remains a difficult open problem. Despite progress in deep learning, genetic programming, and, more recently, large language model (LLM)--based factor generation, existing approaches still explore only a narrow region of the vast alpha search space. Neural models tend to produce opaque and fragile patterns, while symbolic or formula-based methods often yield redundant or economically ungrounded expressions that generalize poorly. Although different in form, these paradigms share a key limitation: none can conduct broad, structured, and human-like exploration that balances logical consistency with creative leaps. To address this gap, we introduce the Cognitive Alpha Mining Framework (CogAlpha), which combines code-level alpha representation with LLM-driven reasoning and evolutionary search. Treating LLMs as adaptive cognitive agents, our framework iteratively refines, mutates, and recombines alpha candidates through multi-stage prompts and financial feedback. This synergistic design enables deeper thinking, richer structural diversity, and economically interpretable alpha discovery, while greatly expanding the effective search space. Experiments on A-share equities demonstrate that CogAlpha consistently discovers alphas with superior predictive accuracy, robustness, and generalization over existing methods. Our results highlight the promise of aligning evolutionary optimization with LLM-based reasoning for automated and explainable alpha discovery. All source code will be released.
Generating accurate SQL from users' natural language questions (text-to-SQL) remains a long-standing challenge due to the complexities involved in user question understanding, database schema comprehension, and SQL generation. Traditional text-to-SQL systems, which combine human engineering and deep neural networks, have made significant progress. Subsequently, pre-trained language models (PLMs) have been developed for text-to-SQL tasks, achieving promising results. However, as modern databases and user questions grow more complex, PLMs with a limited parameter size often produce incorrect SQL. This necessitates more sophisticated and tailored optimization methods, which restricts the application of PLM-based systems. Recently, large language models (LLMs) have shown significant capabilities in natural language understanding as model scale increases. Thus, integrating LLM-based solutions can bring unique opportunities, improvements, and solutions to text-to-SQL research. In this survey, we provide a comprehensive review of existing LLM-based text-to-SQL studies. Specifically, we offer a brief overview of the technical challenges and evolutionary process of text-to-SQL. Next, we introduce the datasets and metrics designed to evaluate text-to-SQL systems. Subsequently, we present a systematic analysis of recent advances in LLM-based text-to-SQL. Finally, we make a summarization and discuss the remaining challenges in this field and suggest expectations for future research directions. All the related resources of LLM-based, including research papers, benchmarks, and open-source projects, are collected for the community in our repository: https://github.com/DEEP-PolyU/Awesome-LLM-based-Text2SQL.
Large Language Models (LLMs) are increasingly integrated into code editors to provide AI-powered code suggestions. Yet many of these suggestions are ignored, resulting in wasted computation, increased latency, and unnecessary interruptions. We introduce a lightweight pre-filtering model that predicts the likelihood of suggestion acceptance before invoking the LLM, using only real-time developer telemetry such as typing speed, file navigation, and editing activity. Deployed in a production-grade Visual Studio Code plugin over four months of naturalistic use, our approach nearly doubled acceptance rates (18.4% -> 34.2%) while suppressing 35% of low-value LLM calls. These findings demonstrate that behavioral signals alone can meaningfully improve both user experience and system efficiency in LLM-assisted programming, highlighting the value of timing-aware, privacy-preserving adaptation mechanisms. The filter operates solely on pre-invocation editor telemetry and never inspects code or prompts.
Large Language Models (LLMs) have transformed code auto-completion by generating context-aware suggestions. Yet, deciding when to present these suggestions remains underexplored, often leading to interruptions or wasted inference calls. We propose an adaptive timing mechanism that dynamically adjusts the delay before offering a suggestion based on real-time developer feedback. Our suggested method combines a logistic transform of recent acceptance rates with a bounded delay range, anchored by a high-level binary prediction of the developer's cognitive state. In a two-month deployment with professional developers, our system improved suggestion acceptance from 4.9% with no delay to 15.4% with static delays, and to 18.6% with adaptive timing-while reducing blind rejections (rejections without being read) from 8.3% to 0.36%. Together, these improvements increase acceptance and substantially reduce wasted inference calls by 75%, making LLM-based code assistants more efficient and cost-effective in practice.
Recent large language models (LLMs) demonstrate multilingual abilities, yet they are English-centric due to dominance of English in training corpora. The limited resource for low-resource languages remains a crucial challenge. Code-switching (CS), a phenomenon where multilingual speakers alternate between languages in a discourse, can convey subtle cultural and linguistic nuances that can be otherwise lost in translation and elicits language-specific knowledge in human communications. In light of this, we investigate whether code-switching can activate, or identify and leverage knowledge for reasoning when LLMs solve low-resource language tasks. To facilitate the research, we first present EnKoQA, a synthetic English-Korean CS question-answering dataset. We provide comprehensive analysis on a variety of multilingual LLMs by subdividing activation process into knowledge identification and knowledge leveraging. Our results demonstrate that compared to English text, CS can faithfully activate knowledge inside LLMs especially on language-specific domains, suggesting the potential of code-switching on low-resource language tasks.
Long-context inference for Large Language Models (LLMs) is heavily limited by high computational demands. While several existing methods optimize attention computation, they still process the full set of hidden states at each layer, limiting overall efficiency. In this work, we propose SlimInfer, an innovative framework that aims to accelerate inference by directly pruning less critical prompt tokens during the forward pass. Our key insight is an information diffusion phenomenon: As information from critical tokens propagates through layers, it becomes distributed across the entire sequence. This diffusion process suggests that LLMs can maintain their semantic integrity when excessive tokens, even including these critical ones, are pruned in hidden states. Motivated by this, SlimInfer introduces a dynamic fine-grained pruning mechanism that accurately removes redundant tokens of hidden state at intermediate layers. This layer-wise pruning naturally enables an asynchronous KV cache manager that prefetches required token blocks without complex predictors, reducing both memory usage and I/O costs. Extensive experiments show that SlimInfer can achieve up to $\mathbf{2.53\times}$ time-to-first-token (TTFT) speedup and $\mathbf{1.88\times}$ end-to-end latency reduction for LLaMA3.1-8B-Instruct on a single RTX 4090, without sacrificing performance on LongBench. Our code is available at https://github.com/Longxmas/SlimInfer.
Reinforcement Learning with Verifiable Rewards (RLVR) has recently demonstrated notable success in enhancing the reasoning performance of large language models (LLMs), particularly on mathematics and programming tasks. Similar to how traditional RL helps agents explore and learn new strategies, RLVR is believed to enable LLMs to continuously self-improve, thus acquiring novel reasoning abilities beyond those of the corresponding base models. In this study we critically examine the current state of RLVR by systematically probing the reasoning capability boundaries of RLVR-trained LLMs across various model families, RL algorithms, and math, coding, and visual reasoning benchmarks, using pass@k at large k values as the evaluation metric. Surprisingly, we find that the current training setup does not elicit fundamentally new reasoning patterns. While RLVR-trained models outperform their base models at small k (e.g., k = 1), the base models achieve a higher pass@k score when k is large. Coverage and perplexity analyses show that the observed reasoning abilities originate from and are bounded by the base model. Treating the base model as an upper bound, our quantitative analysis shows that six popular RLVR algorithms perform similarly and remain far from optimal in leveraging the potential of the base model. By contrast, we find that distillation can introduce new reasoning patterns from the teacher and genuinely expand the model's reasoning capabilities. Overall, our findings suggest that current RLVR methods have not yet realized the potential of RL to elicit truly novel reasoning abilities in LLMs. This highlights the need for improved RL paradigms, such as continual scaling and multi-turn agent-environment interaction, to unlock this potential.
Large Language Models (LLMs) often produce code with subtle implementation-level bugs despite strong benchmark performance. These errors are hard for LLMs to spot and can have large behavioural effects; yet when asked to summarise code, LLMs can frequently surface high-level intent and sometimes overlook this low-level noise. Motivated by this, we propose summary-mediated repair, a prompt-only pipeline for program repair that leverages natural-language code summarisation as an explicit intermediate step, extending previous work that has already shown code summarisation to be a useful intermediary for downstream tasks. We evaluate our method across eight production-grade LLMs on two function level benchmarks (HumanEvalPack and MBPP), comparing several summary styles against a direct repair baseline. Error-aware diagnostic summaries consistently yield the largest gains - repairing up to 65% of unseen errors, on average of 5% more than the baseline - though overall improvements are modest and LLM-dependent. Our results position summaries as a cheap, human-interpretable diagnostic artefact that can be integrated into program-repair pipelines rather than a stand-alone fix-all.
Document AI has advanced rapidly and is attracting increasing attention. Yet, while most efforts have focused on document layout analysis (DLA), its generative counterpart, layout generation, remains underexplored. Distinct from traditional graphic layout design and room layout planning, document layout generation typically involves a larger number of elements per page and exhibits greater structural diversity and complexity. Currently, a major obstacle lies in the scarcity of diverse document layouts: academic papers with Manhattan-style structures dominate existing studies, while open-world genres such as newspapers and magazines remain severely underrepresented. To address this gap, we curate OmniDocLayout-1M, the first million-scale dataset of diverse document layouts, covering six common document types and comprising contemporary layouts collected from multiple sources. Moreover, since existing methods struggle in complex domains and often fail to arrange long sequences coherently, we introduce OmniDocLayout-LLM, a 0.5B model with designed two-stage Coarse-to-Fine learning paradigm:1) learning universal layout principles from our dataset with coarse category definitions, and 2) transferring the knowledge to a specific domain with few fine-grained annotated samples. Extensive experiments demonstrate that our approach achieves strong performance on multiple domains in M$^6$Doc dataset, substantially surpassing both existing layout generation experts and several latest general-purpose LLMs. Our code, dataset, and models will be publicly released.
Informal mathematics has been central to modern large language model (LLM) reasoning, offering flexibility and enabling efficient construction of arguments. However, purely informal reasoning is prone to logical gaps and subtle errors that are difficult to detect and correct. In contrast, formal theorem proving provides rigorous, verifiable mathematical reasoning, where each inference step is checked by a trusted compiler in systems such as Lean, but lacks the exploratory freedom of informal problem solving. This mismatch leaves current LLM-based math agents without a principled way to combine the strengths of both paradigms. In this work, we introduce Hermes, the first tool-assisted agent that explicitly interleaves informal reasoning with formally verified proof steps in Lean. The framework performs intermediate formal checking to prevent reasoning drift and employs a memory module that maintains proof continuity across long, multi-step reasoning chains, enabling both exploration and verification within a single workflow. We evaluate Hermes on four challenging mathematical reasoning benchmarks using LLMs of varying parameter scales, from small models to state-of-the-art systems. Across all settings, Hermes reliably improves the reasoning accuracy of base models while substantially reducing token usage and computational cost compared to reward-based approaches. On difficult datasets such as AIME'25, Hermes achieves up to a 67% accuracy improvement while using 80% fewer total inference FLOPs. The implementation and codebase are publicly available at https://github.com/aziksh-ospanov/HERMES.
Large language models (LLMs) have made significant advancements across various tasks, but their safety alignment remain a major concern. Exploring jailbreak prompts can expose LLMs' vulnerabilities and guide efforts to secure them. Existing methods primarily design sophisticated instructions for the LLM to follow, or rely on multiple iterations, which could hinder the performance and efficiency of jailbreaks. In this work, we propose a novel jailbreak paradigm, Simple Assistive Task Linkage (SATA), which can effectively circumvent LLM safeguards and elicit harmful responses. Specifically, SATA first masks harmful keywords within a malicious query to generate a relatively benign query containing one or multiple [MASK] special tokens. It then employs a simple assistive task such as a masked language model task or an element lookup by position task to encode the semantics of the masked keywords. Finally, SATA links the assistive task with the masked query to jointly perform the jailbreak. Extensive experiments show that SATA achieves state-of-the-art performance and outperforms baselines by a large margin. Specifically, on AdvBench dataset, with mask language model (MLM) assistive task, SATA achieves an overall attack success rate (ASR) of 85% and harmful score (HS) of 4.57, and with element lookup by position (ELP) assistive task, SATA attains an overall ASR of 76% and HS of 4.43.
Aircraft maintenance logs hold valuable safety data but remain underused due to their unstructured text format. This paper introduces LogSyn, a framework that uses Large Language Models (LLMs) to convert these logs into structured, machine-readable data. Using few-shot in-context learning on 6,169 records, LogSyn performs Controlled Abstraction Generation (CAG) to summarize problem-resolution narratives and classify events within a detailed hierarchical ontology. The framework identifies key failure patterns, offering a scalable method for semantic structuring and actionable insight extraction from maintenance logs. This work provides a practical path to improve maintenance workflows and predictive analytics in aviation and related industries.
The sparse Mixture-of-Experts (MoE) architecture of large language models (LLMs) confronts an inherent issue of load imbalance arising from the simplistic linear router strategy, which ultimately causes the instability and inefficient learning of LLMs. To address this challenge, we introduce a novel MoE graph-based framework $\textbf{GMoE}$, aimed at enhancing the collaboration among multiple experts. In GMoE, a graph router function is designed to capture the collaboration signals among experts. This enables all experts to dynamically allocate information derived from input data by sharing information with their neighboring experts. Moreover, we put forward two coordination strategies in GMoE: the $\textit{Poisson distribution-based distinction strategy}$ and the $\textit{Normal distribution-based balance strategy}$, to further release the capacity of each expert and increase the model stability in the fine-tuning of LLMs. Specifically, we leverage a parameter-efficient fine-tuning technique, i.e., Low-Rank Adaptation (LoRA), to implement the graph MoE architecture. Extensive experiments on four real-world benchmark datasets demonstrate the effectiveness of GMoE, showing the benefits of facilitating collaborations of multiple experts in LLM fine-tuning. The code of experimental implementation is available at https://github.com/BAI-LAB/GMoE
Large language models (LLMs) have behaved well in function-level code translation without repository-level context. However, the performance of LLMs in repository-level context code translation remains suboptimal due to complex dependencies and context, hindering their adoption in industrial settings. In this work, we propose a novel LLM-based code translation technique K-Trans, which leverages triple knowledge augmentation to enhance LLM's translation quality under repository context in real-world software development. First, K-Trans constructs a evolving translation knowledge base by extracting relevant information from target-language codebases, the repository being translated, and prior translation results. Second, for each function to be translated, K-Trans retrieves relevant triple knowledge, including target-language code samples, dependency usage examples, and successful translation function pairs, serving as references to enhance LLM for translation. Third, K-Trans constructs a knowledge-augmented translation prompt using the retrieved triple knowledge and employs LLMs to generate the translated code while preserving repository context. It further leverages LLMs for self-debugging, enhancing translation correctness. Lastly, K-Trans continuously evolves the translation knowledge base. The experiments show that K-Trans substantially outperforms the baseline adapted from previous work by 19.4%/40.2% relative improvement in pass@1 and 0.138 in CodeBLEU. It is important to note that the results also demonstrate that each knowledge significantly contributes to K-Trans's effectiveness in handling repository-level context code translation, with dependency usage examples making the most notable contribution. Moreover, as the self-evolution process progresses, the knowledge base continuously enhances the LLM's performance across various aspects of the repository-level code translation.
This paper introduces an infrastructure-aware benchmarking framework for quantifying the environmental footprint of LLM inference across 30 state-of-the-art models in commercial datacenters. The framework combines public API performance data with company-specific environmental multipliers and statistical inference of hardware configurations. We additionally utilize cross-efficiency Data Envelopment Analysis (DEA) to rank models by performance relative to environmental cost and provide a dynamically updated dashboard that visualizes model-level energy, water, and carbon metrics. Results show the most energy-intensive models exceed 29 Wh per long prompt, over 65 times the most efficient systems. Even a 0.42 Wh short query, when scaled to 700M queries/day, aggregates to annual electricity comparable to 35{,}000 U.S. homes, evaporative freshwater equal to the annual drinking needs of 1.2M people, and carbon emissions requiring a Chicago-sized forest to offset. These findings highlight a growing paradox: as AI becomes cheaper and faster, global adoption drives disproportionate resource consumption. Our methodology offers a standardized, empirically grounded basis for sustainability benchmarking and accountability in AI deployment.
Emotion Recognition in Conversation (ERC) is a crucial task for understanding human emotions and enabling natural human-computer interaction. Although Large Language Models (LLMs) have recently shown great potential in this field, their ability to capture the intrinsic connections between explicit and implicit emotions remains limited. We propose a novel ERC training framework, PRC-Emo, which integrates Prompt engineering, demonstration Retrieval, and Curriculum learning, with the goal of exploring whether LLMs can effectively perceive emotions in conversational contexts. Specifically, we design emotion-sensitive prompt templates based on both explicit and implicit emotional cues to better guide the model in understanding the speaker's psychological states. We construct the first dedicated demonstration retrieval repository for ERC, which includes training samples from widely used datasets, as well as high-quality dialogue examples generated by LLMs and manually verified. Moreover, we introduce a curriculum learning strategy into the LoRA fine-tuning process, incorporating weighted emotional shifts between same-speaker and different-speaker utterances to assign difficulty levels to dialogue samples, which are then organized in an easy-to-hard training sequence. Experimental results on two benchmark datasets -- IEMOCAP and MELD -- show that our method achieves new state-of-the-art (SOTA) performance, demonstrating the effectiveness and generalizability of our approach in improving LLM-based emotional understanding.
As large language models (LLMs) are progressively deployed in various real-world applications, personalization of LLMs has become increasingly important. While various approaches to LLM personalization such as prompt-based and training-based methods have been actively explored, the development of effective decoding-time algorithms remains largely overlooked, despite their demonstrated potential. In this paper, we propose CoPe (Contrasting Personal Preference), a novel decoding-time approach applied after performing parameter-efficient fine-tuning (PEFT) on user-specific data. Our core idea is to leverage reward-guided decoding specifically for personalization by maximizing each user's implicit reward signal. We evaluate CoPe across five open-ended personalized text generation tasks. Our empirical results demonstrate that CoPe achieves strong performance, improving personalization by an average of 10.57% in ROUGE-L, without relying on external reward models or additional training procedures.
Ensuring the safety of Autonomous Driving Systems (ADS) requires realistic and reproducible test scenarios, yet extracting such scenarios from multimodal crash reports remains a major challenge. Large Language Models (LLMs) often hallucinate and lose map structure, resulting in unrealistic road layouts and vehicle behaviors. To address this, we introduce SAFE, a novel Scenario-based ADS testing Framework via multimodal Extraction, which leverages Retrieval-Augmented Generation (RAG), knowledge-grounded prompting, Chain-of-Thought (CoT) reasoning, and self-validation to improve scenario reconstruction from multimodal crash data. SAFE achieves 93.8% accuracy in extracting road network details, 80.0% for actor information, and 100% for environmental context. In human studies, SAFE outperforms LCTGen and AC3R in reconstructing consistent road networks and vehicle behaviors. Under identical ADS and simulator settings, SAFE detects 39 and 71 more safety violations than LCTGen and AC3R, respectively, and reproduces 12 more real-world crash cases than LCTGen. On 19 cases supported by AC3R, SAFE reproduces one additional crash case with statistically significant gains across five runs. It generates scenarios within 25 seconds and triggers violations after just 1 case (IDM) and 3 cases (PPO) in MetaDrive, as well as 1 case (Auto) in BeamNG. Code: https://github.com/Siwei-Luo-MQ/SAFE-ADS-Testing
Sliding-window attention offers a hardware-efficient solution to the memory and throughput challenges of Large Language Models (LLMs) in long-context scenarios. Existing methods typically employ a single window length across all attention heads and input sizes. However, this uniform approach fails to capture the heterogeneous attention patterns inherent in LLMs, ignoring their distinct accuracy-latency trade-offs. To address this challenge, we propose *Mixture of Attention Spans* (MoA), which automatically tailors distinct sliding-window length configurations to different heads and layers. MoA constructs and navigates a search space of various window lengths and their scaling rules relative to input sizes. It profiles the model, evaluates potential configurations, and pinpoints the optimal length configurations for each head. MoA adapts to varying input sizes, revealing that some attention heads expand their focus to accommodate longer inputs, while other heads consistently concentrate on fixed-length local contexts. Experiments show that MoA increases the effective context length by 3.9x with the same average sliding-window length, boosting retrieval accuracy by 1.5-7.1x over the uniform-window baseline across Vicuna-{7B, 13B} and Llama3-{8B, 70B} models. Moreover, MoA narrows the performance gap with full attention, reducing the maximum relative performance drop from 9%-36% to within 5% across three long-context understanding benchmarks. MoA achieves a 1.2-1.4x GPU memory reduction, boosting decode throughput by 6.6-8.2x and 1.7-1.9x over FlashAttention2 and vLLM, with minimal performance impact. Our code is available at: https://github.com/thu-nics/MoA
Large language models (LLMs) are increasingly used to generate self-explanations alongside their predictions, a practice that raises concerns about the faithfulness of these explanations, especially in low-resource languages. This study evaluates the faithfulness of LLM-generated explanations in the context of emotion classification in Persian, a low-resource language, by comparing the influential words identified by the model against those identified by human annotators. We assess faithfulness using confidence scores derived from token-level log-probabilities. Two prompting strategies, differing in the order of explanation and prediction (Predict-then-Explain and Explain-then-Predict), are tested for their impact on explanation faithfulness. Our results reveal that while LLMs achieve strong classification performance, their generated explanations often diverge from faithful reasoning, showing greater agreement with each other than with human judgments. These results highlight the limitations of current explanation methods and metrics, emphasizing the need for more robust approaches to ensure LLM reliability in multilingual and low-resource contexts.
A large amount of work has been done in Multi-Agent Systems (MAS) for modeling and solving problems with multiple interacting agents. However, most LLMs are pretrained independently and not specifically optimized for coordination. Existing LLM fine-tuning frameworks rely on individual rewards, which require complex reward designs for each agent to encourage collaboration. To address these challenges, we model LLM collaboration as a cooperative Multi-Agent Reinforcement Learning (MARL) problem. We develop a multi-agent, multi-turn algorithm, Multi-Agent Group Relative Policy Optimization (MAGRPO), to solve it, building on current RL approaches for LLMs as well as MARL techniques. Our experiments on LLM writing and coding collaboration demonstrate that fine-tuning MAS with MAGRPO enables agents to generate high-quality responses efficiently through effective cooperation. Our approach opens the door to using other MARL methods for LLMs and highlights the associated challenges. Our code is available at https://github.com/OpenMLRL/CoMLRL.
This study examines whether Low-Rank Adaptation (LoRA) fine-tuned Large Language Models (LLMs) can approximate the performance of fully fine-tuned models in generating human-interpretable decisions and explanations for malware classification. Achieving trustworthy malware detection, particularly when LLMs are involved, remains a significant challenge. We developed an evaluation framework using Bilingual Evaluation Understudy (BLEU), Recall-Oriented Understudy for Gisting Evaluation (ROUGE), and Semantic Similarity Metrics to benchmark explanation quality across five LoRA configurations and a fully fine-tuned baseline. Results indicate that full fine-tuning achieves the highest overall scores, with BLEU and ROUGE improvements of up to 10% over LoRA variants. However, mid-range LoRA models deliver competitive performance exceeding full fine-tuning on two metrics while reducing model size by approximately 81% and training time by over 80% on a LoRA model with 15.5% trainable parameters. These findings demonstrate that LoRA offers a practical balance of interpretability and resource efficiency, enabling deployment in resource-constrained environments without sacrificing explanation quality. By providing feature-driven natural language explanations for malware classifications, this approach enhances transparency, analyst confidence, and operational scalability in malware detection systems.
Multi-hop question answering over knowledge graphs remains computationally challenging due to the combinatorial explosion of possible reasoning paths. Recent approaches rely on expensive Large Language Model (LLM) inference for both entity linking and path ranking, limiting their practical deployment. Additionally, LLM-generated answers often lack verifiable grounding in structured knowledge. We present two complementary hybrid algorithms that address both efficiency and verifiability: (1) LLM-Guided Planning that uses a single LLM call to predict relation sequences executed via breadth-first search, achieving near-perfect accuracy (micro-F1 > 0.90) while ensuring all answers are grounded in the knowledge graph, and (2) Embedding-Guided Neural Search that eliminates LLM calls entirely by fusing text and graph embeddings through a lightweight 6.7M-parameter edge scorer, achieving over 100 times speedup with competitive accuracy. Through knowledge distillation, we compress planning capability into a 4B-parameter model that matches large-model performance at zero API cost. Evaluation on MetaQA demonstrates that grounded reasoning consistently outperforms ungrounded generation, with structured planning proving more transferable than direct answer generation. Our results show that verifiable multi-hop reasoning does not require massive models at inference time, but rather the right architectural inductive biases combining symbolic structure with learned representations.
This paper proposes a novel semiotic framework for analyzing Large Language Models (LLMs), conceptualizing them as stochastic semiotic engines whose outputs demand active, asymmetric human interpretation. We formalize the trade-off between expressive richness (semiotic breadth) and interpretive stability (decipherability) using information-theoretic tools. Breadth is quantified as source entropy, and decipherability as the mutual information between messages and human interpretations. We introduce a generative complexity parameter (lambda) that governs this trade-off, as both breadth and decipherability are functions of lambda. The core trade-off is modeled as an emergent property of their distinct responses to $位$. We define a semiotic channel, parameterized by audience and context, and posit a capacity constraint on meaning transmission, operationally defined as the maximum decipherability by optimizing lambda. This reframing shifts analysis from opaque model internals to observable textual artifacts, enabling empirical measurement of breadth and decipherability. We demonstrate the framework's utility across four key applications: (i) model profiling; (ii) optimizing prompt/context design; (iii) risk analysis based on ambiguity; and (iv) adaptive semiotic systems. We conclude that this capacity-based semiotic approach offers a rigorous, actionable toolkit for understanding, evaluating, and designing LLM-mediated communication.
The rapid expansion of distributed photovoltaic (PV) systems poses challenges for power grid management, as many installations remain undocumented. While satellite imagery provides global coverage, traditional computer vision (CV) models such as CNNs and U-Nets require extensive labeled data and fail to generalize across regions. This study investigates the cross-domain generalization of a multimodal large language model (LLM) for global PV assessment. By leveraging structured prompts and fine-tuning, the model integrates detection, localization, and quantification within a unified schema. Cross-regional evaluation using the $螖$F1 metric demonstrates that the proposed model achieves the smallest performance degradation across unseen regions, outperforming conventional CV and transformer baselines. These results highlight the robustness of multimodal LLMs under domain shift and their potential for scalable, transferable, and interpretable global PV mapping.
Inference attacks have been widely studied and offer a systematic risk assessment of ML services; however, their implementation and the attack parameters for optimal estimation are challenging for non-experts. The emergence of advanced large language models presents a promising yet largely unexplored opportunity to develop autonomous agents as inference attack experts, helping address this challenge. In this paper, we propose AttackPilot, an autonomous agent capable of independently conducting inference attacks without human intervention. We evaluate it on 20 target services. The evaluation shows that our agent, using GPT-4o, achieves a 100.0% task completion rate and near-expert attack performance, with an average token cost of only $0.627 per run. The agent can also be powered by many other representative LLMs and can adaptively optimize its strategy under service constraints. We further perform trace analysis, demonstrating that design choices, such as a multi-agent framework and task-specific action spaces, effectively mitigate errors such as bad plans, inability to follow instructions, task context loss, and hallucinations. We anticipate that such agents could empower non-expert ML service providers, auditors, or regulators to systematically assess the risks of ML services without requiring deep domain expertise.
EAGER: Edge-Aligned LLM Defense for Robust, Efficient, and Accurate Cybersecurity Question Answering
Large Language Models (LLMs) are highly effective for cybersecurity question answering (QA) but are difficult to deploy on edge devices due to their size. Quantization reduces memory and compute requirements but often degrades accuracy and increases vulnerability to adversarial attacks. We present EAGER, an edge-aligned defense framework that integrates parameter-efficient quantization with domain-specific preference alignment to jointly optimize efficiency, robustness, and accuracy. Unlike prior methods that address these aspects separately, EAGER leverages Quantized Low-Rank Adaptation (QLoRA) for low-cost fine-tuning and Direct Preference Optimization (DPO) on a self-constructed cybersecurity preference dataset, eliminating the need for human labels. Experiments show that EAGER reduces adversarial attack success rates by up to 7.3x and improves QA accuracy by up to 55% over state-of-the-art defenses, while achieving the lowest response latency on a Jetson Orin, demonstrating its practical edge deployment.
Multi-turn conversational attacks, which leverage psychological principles like Foot-in-the-Door (FITD), where a small initial request paves the way for a more significant one, to bypass safety alignments, pose a persistent threat to Large Language Models (LLMs). Progress in defending against these attacks is hindered by a reliance on manual, hard-to-scale dataset creation. This paper introduces a novel, automated pipeline for generating large-scale, psychologically-grounded multi-turn jailbreak datasets. We systematically operationalize FITD techniques into reproducible templates, creating a benchmark of 1,500 scenarios across illegal activities and offensive content. We evaluate seven models from three major LLM families under both multi-turn (with history) and single-turn (without history) conditions. Our results reveal stark differences in contextual robustness: models in the GPT family demonstrate a significant vulnerability to conversational history, with Attack Success Rates (ASR) increasing by as much as 32 percentage points. In contrast, Google's Gemini 2.5 Flash exhibits exceptional resilience, proving nearly immune to these attacks, while Anthropic's Claude 3 Haiku shows strong but imperfect resistance. These findings highlight a critical divergence in how current safety architectures handle conversational context and underscore the need for defenses that can resist narrative-based manipulation.
We study cost-aware routing for large language models across diverse and dynamic pools of models. Existing approaches often overlook prompt-specific context, rely on expensive model profiling, assume a fixed set of experts, or use inefficient trial-and-error strategies. We introduce Cost-Spectrum Contrastive Routing (CSCR), a lightweight framework that maps both prompts and models into a shared embedding space to enable fast, cost-sensitive selection. CSCR uses compact, fast-to-compute logit footprints for open-source models and perplexity fingerprints for black-box APIs. A contrastive encoder is trained to favor the cheapest accurate expert within adaptive cost bands. At inference time, routing reduces to a single k-NN lookup via a FAISS index, requiring no retraining when the expert pool changes and enabling microsecond latency. Across multiple benchmarks, CSCR consistently outperforms baselines, improving the accuracy-cost tradeoff by up to 25%, while generalizing robustly to unseen LLMs and out-of-distribution prompts.
Large language models (LLMs) solve complex problems yet fail on simpler variants, suggesting they achieve correct outputs through mechanisms fundamentally different from human reasoning. To understand this gap, we synthesize cognitive science research into a taxonomy of 28 cognitive elements spanning reasoning invariants, meta-cognitive controls, representations for organizing reasoning & knowledge, and transformation operations. We introduce a fine-grained evaluation framework and conduct the first large-scale empirical analysis of 192K traces from 18 models across text, vision, and audio, complemented by 54 human think-aloud traces, which we make publicly available. We find that models under-utilize cognitive elements correlated with success, narrowing to rigid sequential processing on ill-structured problems where diverse representations and meta-cognitive monitoring are critical. Human traces show more abstraction and conceptual processing, while models default to surface-level enumeration. Meta-analysis of 1.6K LLM reasoning papers reveals the research community concentrates on easily quantifiable elements (sequential organization: 55%, decomposition: 60%) but neglecting meta-cognitive controls (self-awareness: 16%) that correlate with success. Models possess behavioral repertoires associated with success but fail to deploy them spontaneously. Leveraging these patterns, we develop test-time reasoning guidance that automatically scaffold successful structures, improving performance by up to 66.7% on complex problems. By establishing a shared vocabulary between cognitive science and LLM research, our framework enables systematic diagnosis of reasoning failures and principled development of models that reason through robust cognitive mechanisms rather than spurious shortcuts, while providing tools to test theories of human cognition at scale.
We present Genie-CAT, a tool-augmented large-language-model (LLM) system designed to accelerate scientific hypothesis generation in protein design. Using metalloproteins (e.g., ferredoxins) as a case study, Genie-CAT integrates four capabilities -- literature-grounded reasoning through retrieval-augmented generation (RAG), structural parsing of Protein Data Bank files, electrostatic potential calculations, and machine-learning prediction of redox properties -- into a unified agentic workflow. By coupling natural-language reasoning with data-driven and physics-based computation, the system generates mechanistically interpretable, testable hypotheses linking sequence, structure, and function. In proof-of-concept demonstrations, Genie-CAT autonomously identifies residue-level modifications near [Fe--S] clusters that affect redox tuning, reproducing expert-derived hypotheses in a fraction of the time. The framework highlights how AI agents combining language models with domain-specific tools can bridge symbolic reasoning and numerical simulation, transforming LLMs from conversational assistants into partners for computational discovery.
Large Language Models (LLMs) have demonstrated remarkable capabilities in text comprehension, but their ability to process complex, hierarchical tabular data remains underexplored. We present a novel approach to extracting structured data from multi-page government fiscal documents using LLM-based techniques. Applied to annual fiscal documents from the State of Karnataka in India (200+ pages), our method achieves high accuracy through a multi-stage pipeline that leverages domain knowledge, sequential context, and algorithmic validation. A large challenge with traditional OCR methods is the inability to verify the accurate extraction of numbers. When applied to fiscal data, the inherent structure of fiscal tables, with totals at each level of the hierarchy, allows for robust internal validation of the extracted data. We use these hierarchical relationships to create multi-level validation checks. We demonstrate that LLMs can read tables and also process document-specific structural hierarchies, offering a scalable process for converting PDF-based fiscal disclosures into research-ready databases. Our implementation shows promise for broader applications across developing country contexts.
We introduce a novel framework for evaluating the alignment between natural language plans and their expected behavior by converting them into Kripke structures and Linear Temporal Logic (LTL) using Large Language Models (LLMs) and performing model checking. We systematically evaluate this framework on a simplified version of the PlanBench plan verification dataset and report on metrics like Accuracy, Precision, Recall and F1 scores. Our experiments demonstrate that GPT-5 achieves excellent classification performance (F1 score of 96.3%) while almost always producing syntactically perfect formal representations that can act as guarantees. However, the synthesis of semantically perfect formal models remains an area for future exploration.
Multi-agent Large Language Model (LLM) systems have been leading the way in applied LLM research across a number of fields. One notable area is software development, where researchers have advanced the automation of code implementation, code testing, code maintenance, inter alia, using LLM agents. However, software development is a multifaceted environment that extends beyond just code. As such, a successful LLM system must factor in multiple stages of the software development life-cycle (SDLC). In this paper, we propose a vision for ALMAS, an Autonomous LLM-based Multi-Agent Software Engineering framework, which follows the above SDLC philosophy such that it may work within an agile software development team to perform several tasks end-to-end. ALMAS aligns its agents with agile roles, and can be used in a modular fashion to seamlessly integrate with human developers and their development environment. We showcase the progress towards ALMAS through our published works and a use case demonstrating the framework, where ALMAS is able to seamlessly generate an application and add a new feature.
Multi-agent reinforcement learning (MARL) has been increasingly adopted in many real-world applications. While MARL enables decentralized deployment on resource-constrained edge devices, it suffers from severe non-stationarity due to the synchronous updates of agent policies. This non stationarity results in unstable training and poor policy con vergence, especially as the number of agents increases. In this paper, we propose RELED, a scalable MARL framework that integrates large language model (LLM)-driven expert demonstrations with autonomous agent exploration. RELED incorporates a Stationarity-Aware Expert Demonstration module, which leverages theoretical non-stationarity bounds to enhance the quality of LLM-generated expert trajectories, thus providing high reward and training-stable samples for each agent. Moreover, a Hybrid Expert-Agent Policy Optimization module adaptively balances each agent's learning from both expert-generated and agent-generated trajectories, accelerating policy convergence and improving generalization. Extensive experiments with real city networks based on OpenStreetMap demonstrate that RELED achieves superior performance compared to state-of-the-art MARL methods.
As a codebase expands over time, its library dependencies can become outdated and require updates to maintain innovation and security. However, updating a library can introduce breaking changes in the code, necessitating significant developer time for maintenance. To address this, we introduce a framework of LLM agents to be used in combination with migration documentation to automatically recommend and apply code updates and ensure compatibility with new versions. Our solution can automatically localize updated library usages in live Java codebases and implement recommended fixes in a user-friendly manner. The system architecture consists of multiple key components: a Summary Agent, Control Agent, and Code Agent. To validate our approach, we apply the framework on an industrial use case by which we create three synthetic code repositories with major Upgrade changes and benchmark our approach against state-of-the-art methods. Results show that our approach not only performs upgrades using fewer tokens across all cases but also achieves a precision of 71.4%, highlighting its efficiency and effectiveness compared to state-of-the-art methods.
The challenge of designing effective reward functions in reinforcement learning (RL) represents a significant bottleneck, often requiring extensive human expertise and being time-consuming. Previous work and recent advancements in large language models (LLMs) have demonstrated their potential for automating the generation of reward functions. However, existing methodologies often require preliminary evaluation metrics, human-engineered feedback for the refinement process, or the use of environmental source code as context. To address these limitations, this paper introduces LEARN-Opt (LLM-based Evaluator and Analyzer for Reward functioN Optimization). This LLM-based, fully autonomous, and model-agnostic framework eliminates the need for preliminary metrics and environmental source code as context to generate, execute, and evaluate reward function candidates from textual descriptions of systems and task objectives. LEARN-Opt's main contribution lies in its ability to autonomously derive performance metrics directly from the system description and the task objective, enabling unsupervised evaluation and selection of reward functions. Our experiments indicate that LEARN-Opt achieves performance comparable to or better to that of state-of-the-art methods, such as EUREKA, while requiring less prior knowledge. We find that automated reward design is a high-variance problem, where the average-case candidate fails, requiring a multi-run approach to find the best candidates. Finally, we show that LEARN-Opt can unlock the potential of low-cost LLMs to find high-performing candidates that are comparable to, or even better than, those of larger models. This demonstrated performance affirms its potential to generate high-quality reward functions without requiring any preliminary human-defined metrics, thereby reducing engineering overhead and enhancing generalizability.
Test-time scaling, which leverages additional computation during inference to improve model accuracy, has enabled a new class of Large Language Models (LLMs) that are able to reason through complex problems by understanding the goal, turning this goal into a plan, working through intermediate steps, and checking their own work before answering . Frontier large language models with reasoning capabilities, such as DeepSeek-R1 and OpenAI's gpt-oss, follow the same procedure when solving complex problems by generating intermediate reasoning traces before giving the final answer. Today, these models are being increasingly used to generate reasoning traces that serve as high-quality supervised data for post-training of small and medium-sized language models to teach reasoning capabilities without requiring expensive human curation. In this work, we compare the performance of medium-sized LLMs on Math problems after post-training on two kinds of reasoning traces. We compare the impact of reasoning traces generated by DeepSeek-R1 and gpt-oss LLMs in terms of accuracy and inference efficiency.
We investigate the use of LLM-generated data for continual pretraining of encoder models in specialized domains with limited training data, using the scientific domain of invasion biology as a case study. To this end, we leverage domain-specific ontologies by enriching them with LLM-generated data and pretraining the encoder model as an ontology-informed embedding model for concept definitions. To evaluate the effectiveness of this method, we compile a benchmark specifically designed for assessing model performance in invasion biology. After demonstrating substantial improvements over standard LLM pretraining, we investigate the feasibility of applying the proposed approach to domains without comprehensive ontologies by substituting ontological concepts with concepts automatically extracted from a small corpus of scientific abstracts and establishing relationships between concepts through distributional statistics. Our results demonstrate that this automated approach achieves comparable performance using only a small set of scientific abstracts, resulting in a fully automated pipeline for enhancing domain-specific understanding of small encoder models that is especially suited for application in low-resource settings and achieves performance comparable to masked language modeling pretraining on much larger datasets.
Multilingual Alignment is an effective and representative paradigm to enhance LLMs' multilingual capabilities, which transfers the capabilities from the high-resource languages to the low-resource languages. Meanwhile, some research on language-specific neurons provides a new perspective to analyze and understand LLMs' mechanisms. However, we find that there are many neurons that are shared by multiple but not all languages and cannot be correctly classified. In this work, we propose a ternary classification methodology that categorizes neurons into three types, including language-specific neurons, language-related neurons, and general neurons. And we propose a corresponding identification algorithm to distinguish these different types of neurons. Furthermore, based on the distributional characteristics of different types of neurons, we divide the LLMs' internal process for multilingual inference into four parts: (1) multilingual understanding, (2) shared semantic space reasoning, (3) multilingual output space transformation, and (4) vocabulary space outputting. Additionally, we systematically analyze the models before and after alignment with a focus on different types of neurons. We also analyze the phenomenon of ''Spontaneous Multilingual Alignment''. Overall, our work conducts a comprehensive investigation based on different types of neurons, providing empirical results and valuable insights to better understand multilingual alignment and multilingual capabilities of LLMs.
Large Language Models (LLMs) possess general world knowledge but often struggle to generate precise predictions in structured, domain-specific contexts such as simulations. These limitations arise from their inability to ground their broad, unstructured understanding in specific environments. To address this, we present WorldLLM, a framework that enhances LLM-based world modeling by combining Bayesian inference and autonomous active exploration with reinforcement learning. WorldLLM leverages the in-context learning abilities of LLMs to guide an LLM-based world model's predictions using natural language hypotheses given in its prompt. These hypotheses are iteratively refined through a Bayesian inference framework that leverages a second LLM as the proposal distribution given collected evidence. This evidence is collected using a curiosity-driven reinforcement learning policy that explores the environment to find transitions with a low log-likelihood under our LLM-based predictive model using the current hypotheses. By alternating between refining hypotheses and collecting new evidence, our framework autonomously drives continual improvement of the predictions. Our experiments demonstrate the effectiveness of WorldLLM in a textual game environment that requires agents to manipulate and combine objects. The framework not only enhances predictive accuracy, but also generates human-interpretable theories of environment dynamics.
Large Language Models (LLMs) have developed rapidly in web services, delivering unprecedented capabilities while amplifying societal risks. Existing works tend to focus on either isolated jailbreak attacks or static defenses, neglecting the dynamic interplay between evolving threats and safeguards in real-world web contexts. To mitigate these challenges, we propose ACE-Safety (Adversarial Co-Evolution for LLM Safety), a novel framework that jointly optimize attack and defense models by seamlessly integrating two key innovative procedures: (1) Group-aware Strategy-guided Monte Carlo Tree Search (GS-MCTS), which efficiently explores jailbreak strategies to uncover vulnerabilities and generate diverse adversarial samples; (2) Adversarial Curriculum Tree-aware Group Policy Optimization (AC-TGPO), which jointly trains attack and defense LLMs with challenging samples via curriculum reinforcement learning, enabling robust mutual improvement. Evaluations across multiple benchmarks demonstrate that our method outperforms existing attack and defense approaches, and provides a feasible pathway for developing LLMs that can sustainably support responsible AI ecosystems.
A critical barrier to the trustworthiness of sixth-generation (6G) agentic autonomous networks is the uncertainty neglect bias; a cognitive tendency for large language model (LLM)-powered agents to make high-stakes decisions based on simple averages while ignoring the tail risk of extreme events. This paper proposes an unbiased, risk-aware framework for agentic negotiation, designed to ensure robust resource allocation in 6G network slicing. Specifically, agents leverage Digital Twins (DTs) to predict full latency distributions, which are then evaluated using a formal framework from extreme value theory, namely, Conditional Value-at-Risk (CVaR). This approach fundamentally shifts the agent's objective from reasoning over the mean to reasoning over the tail, thereby building a statistically-grounded buffer against worst-case outcomes. Furthermore, our framework ensures full uncertainty awareness by requiring agents to quantify epistemic uncertainty -- confidence in their own DTs predictions -- and propagate this meta-verification to make robust decisions, preventing them from acting on unreliable data. We validate this framework in a 6G inter-slice negotiation use-case between an eMBB and a URLLC agent. The results demonstrate the profound failure of the biased, mean-based baseline, which consistently fails its SLAs with a 25\% rate. Our unbiased, CVaR-aware agent successfully mitigates this bias, eliminating SLA violations and reducing the URLLC and eMBB p99.999 latencies by around 11\%. We show this reliability comes at the rational and quantifiable cost of slightly reduced energy savings to 17\%, exposing the false economy of the biased approach. This work provides a concrete methodology for building the trustworthy autonomous systems required for 6G.
Research on the safety evaluation of large language models (LLMs) has become extensive, driven by jailbreak studies that elicit unsafe responses. Such response involves information already available to humans, such as the answer to "how to make a bomb". When LLMs are jailbroken, the practical threat they pose to humans is negligible. However, it remains unclear whether LLMs commonly produce unpredictable outputs that could pose substantive threats to human safety. To address this gap, we study whether LLM-generated content contains potential existential threats, defined as outputs that imply or promote direct harm to human survival. We propose \textsc{ExistBench}, a benchmark designed to evaluate such risks. Each sample in \textsc{ExistBench} is derived from scenarios where humans are positioned as adversaries to AI assistants. Unlike existing evaluations, we use prefix completion to bypass model safeguards. This leads the LLMs to generate suffixes that express hostility toward humans or actions with severe threat, such as the execution of a nuclear strike. Our experiments on 10 LLMs reveal that LLM-generated content indicates existential threats. To investigate the underlying causes, we also analyze the attention logits from LLMs. To highlight real-world safety risks, we further develop a framework to assess model behavior in tool-calling. We find that LLMs actively select and invoke external tools with existential threats. Code and data are available at: https://github.com/cuiyu-ai/ExistBench.
Recent studies have shown that large language models (LLMs), especially smaller ones, often lack robustness in grade school math (GSM) reasoning. In particular, they tend to experience performance drops when faced with distribution shifts, such as changes to numerical or nominal variables, or insertions of distracting clauses. A possible strategy to address this involves generating synthetic data to further "instantiate" reasoning problems on potential variations. In this work, we instead focuses on the strategy of "abstracting" reasoning problems. This not only helps counteract distribution shifts but also facilitates the connection to symbolic tools for deriving solutions. Focusing on GSM, we find that this abstraction process is better acquired through reinforcement learning (RL) than just supervised fine-tuning, which often fails to produce faithful abstractions. Our method, AbstRaL -- which promotes abstract reasoning in LLMs using RL on granular abstraction data -- significantly mitigates performance degradation on recent GSM perturbation benchmarks. Besides, improving GSM robustness via AbstRaL is shown to also implicitly benefit LLMs' capabilities on OOD mathematical and general reasoning tasks, indicating that abstract thinking broadly enables better generalizability.
Large Language Models (LLMs) are commonly pretrained on vast corpora of text without utilizing contextual metadata such as source, quality, or topic, leading to a context-free learning paradigm. While recent studies suggest that adding metadata like URL information as context (i.e., auxiliary inputs not used in the loss calculation) can improve training efficiency and downstream performance, they offer limited understanding of which types of metadata are truly effective and under what conditions. In this work, we conduct a systematic evaluation and find that not all metadata types contribute equally. Only URL context speeds up training, whereas quality scores and topic/format domain information offer no clear benefit. Furthermore, the improved downstream performances of URL conditioning emerge only when longer prompts are used at inference time. In addition, we demonstrate that context-aware pretraining enables more controllable generation than context-free pretraining, in a classifier-free guidance fashion. Although topic and format metadata do not accelerate training, they are effective for steering outputs, offering human-interpretable control over generation.
Large Language Models (LLMs) have recently been leveraged for asset pricing tasks and stock trading applications, enabling AI agents to generate investment decisions from unstructured financial data. However, most evaluations of LLM timing-based investing strategies are conducted on narrow timeframes and limited stock universes, overstating effectiveness due to survivorship and data-snooping biases. We critically assess their generalizability and robustness by proposing FINSABER, a backtesting framework evaluating timing-based strategies across longer periods and a larger universe of symbols. Systematic backtests over two decades and 100+ symbols reveal that previously reported LLM advantages deteriorate significantly under broader cross-section and over a longer-term evaluation. Our market regime analysis further demonstrates that LLM strategies are overly conservative in bull markets, underperforming passive benchmarks, and overly aggressive in bear markets, incurring heavy losses. These findings highlight the need to develop LLM strategies that are able to prioritise trend detection and regime-aware risk controls over mere scaling of framework complexity.
A well-known testing method for the safety evaluation and real-time validation of automotive software systems (ASSs) is Fault Injection (FI). In accordance with the ISO 26262 standard, the faults are introduced artificially for the purpose of analyzing the safety properties and verifying the safety mechanisms during the development phase. However, the current FI method and tools have a significant limitation in that they require manual identification of FI attributes, including fault type, location and time. The more complex the system, the more expensive, time-consuming and labour-intensive the process. To address the aforementioned challenge, a novel Large Language Models (LLMs)-assisted fault test cases (TCs) generation approach for utilization during real-time FI tests is proposed in this paper. To this end, considering the representativeness and coverage criteria, the applicability of various LLMs to create fault TCs from the functional safety requirements (FSRs) has been investigated. Through the validation results of LLMs, the superiority of the proposed approach utilizing gpt-4o in comparison to other state-of-the-art models has been demonstrated. Specifically, the proposed approach exhibits high performance in terms of FSRs classification and fault TCs generation with F1-score of 88% and 97.5%, respectively. To illustrate the proposed approach, the generated fault TCs were executed in real time on a hardware-in-the-loop system, where a high-fidelity automotive system model served as a case study. This novel approach offers a means of optimizing the real-time testing process, thereby reducing costs while simultaneously enhancing the safety properties of complex safety-critical ASSs.
Chain-of-Thought (CoT) reasoning is a critical capability for large language models (LLMs), enabling them to tackle com- plex multi-step tasks. While base LLMs, pre-trained on general text corpora, often struggle with reasoning due to a lack of specialized training, recent studies reveal their latent reason- ing potential tied to hidden states. However, existing hidden state manipulation methods, such as linear activation steering, suffer from limitations due to their rigid and unconstrained nature, often leading to distribution shifts and degraded text quality. In this work, we propose a novel approach for elic- iting CoT reasoning from base LLMs through hidden state manipulation grounded in probabilistic conditional generation. By reformulating the challenge as an optimization problem with a balanced likelihood and prior regularization framework, our method guides hidden states toward reasoning-oriented trajectories while preserving linguistic coherence. Extensive evaluations across mathematical, commonsense, and logical reasoning benchmarks demonstrate that our approach con- sistently outperforms existing steering methods, offering a theoretically principled and effective solution for enhancing reasoning capabilities in base LLMs.
Obfuscation poses a persistent challenge for software engineering tasks such as program comprehension, maintenance, testing, and vulnerability detection. While compiler optimizations and third-party code often introduce transformations that obscure program intent, existing analysis tools and large language models (LLMs) struggle to recover the original semantics. In this work, we investigate whether LLMs, when fine-tuned with symbolic execution artifacts, can effectively deobfuscate programs and restore analyzability. We construct a benchmark by applying four widely studied transformations-control-flow flattening, opaque predicates, arithmetic encoding, and branch encoding-across diverse C programs from TUM Obfuscation Benchmarks, the LLVM test suite, and algorithmic repositories. We then compare three state-of-the-art LLMs under two training configurations: baseline fine-tuning on obfuscated/original code pairs, and enhanced fine-tuning with additional KLEE artifacts such as SMT constraints, path statistics, and test cases. Our evaluation examines syntactic correctness (compilation success), semantic fidelity (behavioral equivalence under symbolic execution), and code quality (readability and structure). Results show that GPT-4.1-mini achieves the strongest deobfuscation overall, and that incorporating KLEE artifacts consistently improves semantic preservation and compilation success across models. These findings highlight deobfuscation as a broader software engineering concern, demonstrating that combining LLMs with symbolic execution can strengthen automated testing, static analysis, and program comprehension in the presence of obfuscation.
As large language models (LLMs) become increasingly prevalent, understanding human-LLM interactions is emerging as a central priority in psychological research. Online experiments offer an efficient means to study human-LLM interactions, yet integrating LLMs into established survey platforms remains technically demanding, particularly when aiming for ecologically valid, real-time conversational experiences with strong experimental control. We introduce Simple Chat, an open-source, research-focused chat interface that streamlines LLM integration for platforms such as Qualtrics, oTree, and LimeSurvey, while presenting a unified participant experience across conditions. Simple Chat connects to both commercial providers and open-weights models, supports streaming responses to preserve conversational flow, and offers an administrative interface for fine-grained control of prompts and interface features. By reducing technical barriers, standardizing interfaces, and improving participant experience, Simple Chat helps advance the study of human-LLM interaction. In this article, we outline Simple Chat's key features, provide a step-by-step tutorial, and demonstrate its utility through two illustrative case studies.
Aspect category sentiment analysis (ACSA) has achieved remarkable progress with large language models (LLMs), yet existing approaches primarily emphasize sentiment polarity while overlooking the underlying emotional dimensions that shape sentiment expressions. This limitation hinders the model's ability to capture fine-grained affective signals toward specific aspect categories. To address this limitation, we introduce a novel emotion-enhanced multi-task ACSA framework that jointly learns sentiment polarity and category-specific emotions grounded in Ekman's six basic emotions. Leveraging the generative capabilities of LLMs, our approach enables the model to produce emotional descriptions for each aspect category, thereby enriching sentiment representations with affective expressions. Furthermore, to ensure the accuracy and consistency of the generated emotions, we introduce an emotion refinement mechanism based on the Valence-Arousal-Dominance (VAD) dimensional framework. Specifically, emotions predicted by the LLM are projected onto a VAD space, and those inconsistent with their corresponding VAD coordinates are re-annotated using a structured LLM-based refinement strategy. Experimental results demonstrate that our approach significantly outperforms strong baselines on all benchmark datasets. This underlines the effectiveness of integrating affective dimensions into ACSA.
In-context learning (ICL) with large language models (LLMs) has emerged as a promising paradigm for named entity recognition (NER) in low-resource scenarios. However, existing ICL-based NER methods suffer from three key limitations: (1) reliance on dynamic retrieval of annotated examples, which is problematic when annotated data is scarce; (2) limited generalization to unseen domains due to the LLM's insufficient internal domain knowledge; and (3) failure to incorporate external knowledge or resolve entity ambiguities. To address these challenges, we propose KDR-Agent, a novel multi-agent framework for multi-domain low-resource in-context NER that integrates Knowledge retrieval, Disambiguation, and Reflective analysis. KDR-Agent leverages natural-language type definitions and a static set of entity-level contrastive demonstrations to reduce dependency on large annotated corpora. A central planner coordinates specialized agents to (i) retrieve factual knowledge from Wikipedia for domain-specific mentions, (ii) resolve ambiguous entities via contextualized reasoning, and (iii) reflect on and correct model predictions through structured self-assessment. Experiments across ten datasets from five domains demonstrate that KDR-Agent significantly outperforms existing zero-shot and few-shot ICL baselines across multiple LLM backbones. The code and data can be found at https://github.com/MWXGOD/KDR-Agent.
Large language models (LLMs) have demonstrated impressive capabilities in natural language understanding and generation, including multi-step reasoning such as mathematical proving. However, existing approaches often lack an explicit and dynamic mechanism to structurally represent and evolve intermediate reasoning states, which limits their ability to perform context-aware theorem selection and iterative conclusion generation. To address these challenges, we propose GraphMind, a novel dynamic graph-based framework that integrates the graph neural network (GNN) with LLMs to iteratively select theorems and generate intermediate conclusions for multi-step reasoning. Our method models the reasoning process as a heterogeneous evolving graph, where nodes represent conditions, theorems, and conclusions, while edges capture logical dependencies between nodes. By encoding the current reasoning state with GNN and leveraging semantic matching for theorem selection, our framework enables context-aware, interpretable, and structured reasoning in a closed-loop manner. Experiments on various question-answering (QA) datasets demonstrate that our proposed GraphMind method achieves consistent performance improvements and significantly outperforms existing baselines in multi-step reasoning, validating the effectiveness and generalizability of our approach.
Large language models (LLMs) may generate outputs that are misaligned with user intent, lack contextual grounding, or exhibit hallucinations during conversation, which compromises the reliability of LLM-based applications. This review aimed to identify and analyze techniques that align LLM responses with conversational goals, ensure grounding, and reduce hallucination and topic drift. We conducted a Rapid Review guided by the PRISMA framework and the PICO strategy to structure the search, filtering, and selection processes. The alignment strategies identified were categorized according to the LLM lifecycle phase in which they operate: inference-time, post-training, and reinforcement learning-based methods. Among these, inference-time approaches emerged as particularly efficient, aligning outputs without retraining while supporting user intent, contextual grounding, and hallucination mitigation. The reviewed techniques provided structured mechanisms for improving the quality and reliability of LLM responses across key alignment objectives.
Large Language Models (LLMs) have demonstrated strong translation abilities through prompting, even without task-specific training. However, their effectiveness in dialectal and low-resource contexts remains underexplored. This study presents the first systematic investigation of LLM-based machine translation (MT) for Sylheti, a dialect of Bangla that is itself low-resource. We evaluate five advanced LLMs (GPT-4.1, GPT-4.1, LLaMA 4, Grok 3, and DeepSeek V3.2) across both translation directions (Bangla $\Leftrightarrow$ Sylheti), and find that these models struggle with dialect-specific vocabulary. To address this, we introduce Sylheti-CAP (Context-Aware Prompting), a three-step framework that embeds a linguistic rulebook, a dictionary (2{,}260 core vocabulary items and idioms), and an authenticity check directly into prompts. Extensive experiments show that Sylheti-CAP consistently improves translation quality across models and prompting strategies. Both automatic metrics and human evaluations confirm its effectiveness, while qualitative analysis reveals notable reductions in hallucinations, ambiguities, and awkward phrasing, establishing Sylheti-CAP as a scalable solution for dialectal and low-resource MT. Dataset link: \href{https://github.com/TabiaTanzin/LLMs-for-Low-Resource-Dialect-Translation-Using-Context-Aware-Prompting-A-Case-Study-on-Sylheti.git}{https://github.com/TabiaTanzin/LLMs-for-Low-Resource-Dialect-Translation-Using-Context-Aware-Prompting-A-Case-Study-on-Sylheti.git}
The integration of Large Language Models (LLMs) into healthcare demands a safety paradigm rooted in \textit{primum non nocere}. However, current alignment techniques rely on generic definitions of harm that fail to capture context-dependent violations, such as administrative fraud and clinical discrimination. To address this, we introduce Medical Malice: a dataset of 214,219 adversarial prompts calibrated to the regulatory and ethical complexities of the Brazilian Unified Health System (SUS). Crucially, the dataset includes the reasoning behind each violation, enabling models to internalize ethical boundaries rather than merely memorizing a fixed set of refusals. Using an unaligned agent (Grok-4) within a persona-driven pipeline, we synthesized high-fidelity threats across seven taxonomies, ranging from procurement manipulation and queue-jumping to obstetric violence. We discuss the ethical design of releasing these "vulnerability signatures" to correct the information asymmetry between malicious actors and AI developers. Ultimately, this work advocates for a shift from universal to context-aware safety, providing the necessary resources to immunize healthcare AI against the nuanced, systemic threats inherent to high-stakes medical environments -- vulnerabilities that represent the paramount risk to patient safety and the successful integration of AI in healthcare systems.
Large Language Models (LLMs) inherit societal biases from their training data, potentially leading to harmful or unfair outputs. While various techniques aim to mitigate these biases, their effects are often evaluated only along the dimension of the bias being targeted. This work investigates the cross-category consequences of targeted bias mitigation. We study four bias mitigation techniques applied across ten models from seven model families, and we explore racial, religious, profession- and gender-related biases. We measure the impact of debiasing on model coherence and stereotypical preference using the StereoSet benchmark. Our results consistently show that while targeted mitigation can sometimes reduce bias in the intended dimension, it frequently leads to unintended and often negative consequences in others, such as increasing model bias and decreasing general coherence. These findings underscore the critical need for robust, multi-dimensional evaluation tools when examining and developing bias mitigation strategies to avoid inadvertently shifting or worsening bias along untargeted axes.
Recent advances in large language models made it possible to achieve high conversational performance with substantially reduced computational demands, enabling practical on-site deployment in clinical environments. Such progress allows for local integration of AI systems that uphold strict data protection and patient privacy requirements, yet their secure implementation in medicine necessitates careful consideration of ethical, regulatory, and technical constraints. In this study, we introduce MedChat, a locally deployable virtual physician framework that integrates an LLM-based medical chatbot with a diffusion-driven avatar for automated and structured anamnesis. The chatbot was fine-tuned using a hybrid corpus of real and synthetically generated medical dialogues, while model efficiency was optimized via Low-Rank Adaptation. A secure and isolated database interface was implemented to ensure complete separation between patient data and the inference process. The avatar component was realized through a conditional diffusion model operating in latent space, trained on researcher video datasets and synchronized with mel-frequency audio features for realistic speech and facial animation. Unlike existing cloud-based systems, this work demonstrates the feasibility of a fully offline, locally deployable LLM-diffusion framework for clinical anamnesis. The autoencoder and diffusion networks exhibited smooth convergence, and MedChat achieved stable fine-tuning with strong generalization to unseen data. The proposed system thus provides a privacy-preserving, resource-efficient foundation for AI-assisted clinical anamnesis, also in low-cost settings.
Bug bounty platforms (e.g., HackerOne, BugCrowd) leverage crowd-sourced vulnerability discovery to improve continuous coverage, reduce the cost of discovery, and serve as an integral complement to internal red teams. With the rise of AI-generated bug reports, little work exists to help bug hunters understand why these reports are labeled as invalid. To improve report quality and reduce reviewers' burden, it is critical to predict invalid reports and interpret invalid reasons. In this work, we conduct an empirical study with the purpose of helping bug hunters understand the validity of reports. We collect a dataset of 9,942 disclosed bug bounty reports, including 1,400 invalid reports, and evaluate whether state-of-the-art large language models can identify invalid reports. While models such as GPT-5, DeepSeek, and a fine-tuned RoBERTa achieve strong overall accuracy, they consistently struggle to detect invalid cases, showing a tendency to over-accept reports. To improve invalidity detection, we build a taxonomy of rejection reasons for Information Disclosure vulnerabilities and incorporate it into a retrieval-augmented generation (RAG) framework. This approach substantially improves classification consistency and reduces bias. We also examine whether reviewer decisions may be influenced by factors beyond the content of the report. Our analysis shows that reporters with higher reputations tend to receive more favorable outcomes in borderline cases, suggesting that perceived expertise can influence review judgments. Overall, our findings highlight the challenges of invalid report identification and show that combining LLMs with structured reviewer knowledge can support more transparent and consistent vulnerability report review.
Psychological assessments are dominated by rating scales, which cannot capture the nuance in natural language. Efforts to supplement them with qualitative text have relied on labelled datasets or expert rubrics, limiting scalability. We introduce a framework that avoids this reliance: large language models (LLMs) score free-text responses with simple prompts to produce candidate LLM items, from which we retain those that yield the most test information when co-calibrated with a baseline scale. Using depression as a case study, we developed and tested the method in upper-secondary students (n=693) and a matched synthetic dataset (n=3,000). Results on held-out test sets showed that augmenting a 19-item scale with LLM items improved its precision, accuracy, and convergent validity. Further, the test information gain matched that of adding as many as 16 rating-scale items. This framework leverages the increasing availability of transcribed language to enhance psychometric measures, with applications in clinical health and beyond.
LLMs are commonly trained with a learning rate (LR) warmup, followed by cosine decay to 10% of the maximum (10x decay). In a large-scale empirical study, we show that under an optimal peak LR, a simple linear decay-to-zero (D2Z) schedule consistently outperforms other schedules when training at compute-optimal dataset sizes. D2Z is superior across a range of model sizes, batch sizes, datasets, and vocabularies. Benefits increase as dataset size increases. Leveraging a novel interpretation of AdamW as an exponential moving average of weight updates, we show how linear D2Z optimally balances the demands of early training (moving away from initial conditions) and late training (averaging over more updates in order to mitigate gradient noise). In experiments, a 610M-parameter model trained for 80 tokens-per-parameter (TPP) using D2Z achieves lower loss than when trained for 200 TPP using 10x decay, corresponding to an astonishing 60% compute savings. Models such as Llama2-7B, trained for 286 TPP with 10x decay, could likely have saved a majority of compute by training with D2Z.
Efficient LLM pre-training requires well-tuned hyperparameters (HPs), including learning rate $畏$ and weight decay $位$. We study scaling laws for HPs: formulas for how to scale HPs as we scale model size N, dataset size D, and batch size B. Recent work suggests the AdamW timescale, $蟿= B/(畏位D)$, should remain constant across training settings, and we verify the implication that optimal $位$ scales linearly with B, for a fixed N and D. However, as N and D scale, we show optimal $蟿$ obeys a precise power law in the tokens-per-parameter ratio, D/N. This law thus provides a method to accurately predict $位$opt in advance of large-scale training. We also study scaling laws for optimal batch size Bopt (the B enabling lowest loss at a given N,D) and critical batch size Bcrit (the B beyond which further data parallelism becomes ineffective). In contrast to prior work, we find both Bopt and Bcrit scale as power laws in D, independent of model size, N. Finally, we analyze how these findings inform the real-world selection of Pareto-optimal N and D under dual training time and compute objectives. All experiments were run on Cerebras CS-3 systems.
Organizations are rapidly adopting Large Language Models (LLMs) to transform their operations, yet they lack clear guidance on key decisions for adoption and implementation. While LLMs offer powerful capabilities in content generation, assisted coding, and process automation, businesses face critical challenges in data security, LLM solution development approach, infrastructure requirements, and deployment strategies. Healthcare providers must protect patient data while leveraging LLMs for medical analysis, financial institutions need to balance automated customer service with regulatory compliance, and software companies seek to enhance development productivity while maintaining code security. This article presents a systematic six-step decision framework for LLM adoption, helping organizations navigate from initial application selection to final deployment. Based on extensive interviews and analysis of successful and failed implementations, our framework provides practical guidance for business leaders to align technological capabilities with business objectives. Through key decision points and real-world examples from both B2B and B2C contexts, organizations can make informed decisions about LLM adoption while ensuring secure and efficient integration across various use cases, from customer service automation to content creation and advanced analytics.
Many recent studies showed that LLMs are vulnerable to jailbreak attacks, where an attacker can perturb the input of an LLM to induce it to generate an output for a harmful question. In general, existing jailbreak techniques either optimize a semantic template intended to induce the LLM to produce harmful outputs or optimize a suffix that leads the LLM to initiate its response with specific tokens (e.g., "Sure"). In this work, we introduce TASO (Template and Suffix Optimization), a novel jailbreak method that optimizes both a template and a suffix in an alternating manner. Our insight is that suffix optimization and template optimization are complementary to each other: suffix optimization can effectively control the first few output tokens but cannot control the overall quality of the output, while template optimization provides guidance for the entire output but cannot effectively control the initial tokens, which significantly impact subsequent responses. Thus, they can be combined to improve the attack's effectiveness. We evaluate the effectiveness of TASO on benchmark datasets (including HarmBench and AdvBench) on 24 leading LLMs (including models from the Llama family, OpenAI, and DeepSeek). The results demonstrate that TASO can effectively jailbreak existing LLMs. We hope our work can inspire future studies in exploring this direction. We will make code and data publicly available.
Despite rapid progress in logic locking (LL), reproducibility remains a challenge as codes are rarely made public. We present LockForge, a first-of-its-kind, multi-agent large language model (LLM) framework that turns LL descriptions in papers into executable and tested code. LockForge provides a carefully crafted pipeline realizing forethought, implementation, iterative refinement, and a multi-stage validation, all to systematically bridge the gap between prose and practice for complex LL schemes. For validation, we devise (i) an LLM-as-Judge stage with a scoring system considering behavioral checks, conceptual mechanisms, structural elements, and reproducibility on benchmarks, and (ii) an independent LLM-as-Examiner stage for ground-truth assessment. We apply LockForge to 10 seminal LL schemes, many of which lack reference implementations. Our evaluation on multiple SOTA LLMs, including ablation studies, reveals the significant complexity of the task. We show that an advanced reasoning model and a sophisticated, multi-stage framework like LockForge are required. We release all implementations and benchmarks, providing a reproducible and fair foundation for evaluation of further LL research.
Large language models (LLMs) offer significant potential in enhancing psychiatric practice, from improving diagnostic accuracy to streamlining clinical documentation and therapeutic support. However, existing evaluation resources heavily rely on small clinical interview corpora, social media posts, or synthetic dialogues, which limits their clinical validity and fails to capture the full complexity of diagnostic reasoning. In this work, we introduce PsychiatryBench, a rigorously curated benchmark grounded exclusively in authoritative, expert-validated psychiatric textbooks and casebooks. PsychiatryBench comprises eleven distinct question-answering tasks ranging from diagnostic reasoning and treatment planning to longitudinal follow-up, management planning, clinical approach, sequential case analysis, and multiple-choice/extended matching formats totaling 5,188 expert-annotated items. {\color{red}We evaluate a diverse set of frontier LLMs (including Google Gemini, DeepSeek, Sonnet 4.5, and GPT 5) alongside leading open-source medical models such as MedGemma using both conventional metrics and an "LLM-as-judge" similarity scoring framework. Our results reveal substantial gaps in clinical consistency and safety, particularly in multi-turn follow-up and management tasks, underscoring the need for specialized model tuning and more robust evaluation paradigms. PsychiatryBench offers a modular, extensible platform for benchmarking and improving LLM performance in mental health applications.
The rapid advancement of Large Language Model (LLM)-driven multi-agent systems has significantly streamlined software developing tasks, enabling users with little technical expertise to develop executable applications. While these systems democratize software creation through natural language requirements, they introduce significant security risks that remain largely unexplored. We identify two risky scenarios: Malicious User with Benign Agents (MU-BA) and Benign User with Malicious Agents (BU-MA). We introduce the Implicit Malicious Behavior Injection Attack (IMBIA), demonstrating how multi-agent systems can be manipulated to generate software with concealed malicious capabilities beneath seemingly benign applications, and propose Adv-IMBIA as a defense mechanism. Evaluations across ChatDev, MetaGPT, and AgentVerse frameworks reveal varying vulnerability patterns, with IMBIA achieving attack success rates of 93%, 45%, and 71% in MU-BA scenarios, and 71%, 84%, and 45% in BU-MA scenarios. Our defense mechanism reduced attack success rates significantly, particularly in the MU-BA scenario. Further analysis reveals that compromised agents in the coding and testing phases pose significantly greater security risks, while also identifying critical agents that require protection against malicious user exploitation. Our findings highlight the urgent need for robust security measures in multi-agent software development systems and provide practical guidelines for implementing targeted, resource-efficient defensive strategies.
LLMs have demonstrated unprecedented capabilities in natural language processing, yet their practical deployment remains hindered by persistent factuality and faithfulness hallucinations. While existing methods address these hallucination types independently, they inadvertently induce performance trade-offs, as interventions targeting one type often exacerbate the other. Through empirical and theoretical analysis of activation space dynamics in LLMs, we reveal that these hallucination categories share overlapping subspaces within neural representations, presenting an opportunity for concurrent mitigation. To harness this insight, we propose SPACE, a unified framework that jointly enhances factuality and faithfulness by editing shared activation subspaces. SPACE establishes a geometric foundation for shared subspace existence through dual-task feature modeling, then identifies and edits these subspaces via a hybrid probe strategy combining spectral clustering and attention head saliency scoring. Experimental results across multiple benchmark datasets demonstrate the superiority of our approach.
Reinforcement Learning (RL) has become critical for advancing modern Large Language Models (LLMs), yet existing synchronous RL systems face severe performance bottlenecks. The rollout phase, which dominates end-to-end iteration time, suffers from substantial long-tail latency and poor resource utilization due to inherent workload imbalance. We present Seer, a novel online context learning system that addresses these challenges by exploiting previously overlooked similarities in output lengths and generation patterns among requests sharing the same prompt. Seer introduces three key techniques: divided rollout for dynamic load balancing, context-aware scheduling, and adaptive grouped speculative decoding. Together, these mechanisms substantially reduce long-tail latency and improve resource efficiency during rollout. Evaluations on production-grade RL workloads demonstrate that Seer improves end-to-end rollout throughput by 74% to 97% and reduces long-tail latency by 75% to 93% compared to state-of-the-art synchronous RL systems, significantly accelerating RL training iterations.
Federated Learning (FL) is increasingly being adopted in military collaborations to develop Large Language Models (LLMs) while preserving data sovereignty. However, prompt injection attacks-malicious manipulations of input prompts-pose new threats that may undermine operational security, disrupt decision-making, and erode trust among allies. This perspective paper highlights four vulnerabilities in federated military LLMs: secret data leakage, free-rider exploitation, system disruption, and misinformation spread. To address these risks, we propose a human-AI collaborative framework with both technical and policy countermeasures. On the technical side, our framework uses red/blue team wargaming and quality assurance to detect and mitigate adversarial behaviors of shared LLM weights. On the policy side, it promotes joint AI-human policy development and verification of security protocols.
Large Language Models (LLMs) and their agent systems have recently demonstrated strong potential in automating code reasoning and vulnerability detection. However, when applied to large-scale firmware, their performance degrades due to the binary nature of firmware, complex dependency structures, and heterogeneous components. To address this challenge, this paper presents FIRMHIVE, a recursive agent hive that enables LLMs to act as autonomous firmware security analysts. FIRMHIVE introduces two key mechanisms: (1) transforming delegation into a per-agent, executable primitive and (2) constructing a runtime Tree of Agents (ToA) for decentralized coordination. We evaluate FIRMHIVE using real-world firmware images obtained from publicly available datasets, covering five representative security analysis tasks. Compared with existing LLM-agent baselines, FIRMHIVE performs deeper (about 16x more reasoning steps) and broader (about 2.3x more files inspected) cross-file exploration, resulting in about 5.6x more alerts per firmware. Compared to state-of-the-art (SOTA) security tools, FIRMHIVE identifies about 1.5x more vulnerabilities (1,802 total) and achieves 71% precision, representing significant improvements in both yield and fidelity.
Large Language Models have been widely been adopted by users for writing tasks such as sentence completions. While this can improve writing efficiency, prior research shows that LLM-generated suggestions may exhibit cultural biases which may be difficult for users to detect, especially in educational contexts for non-native English speakers. While such prior work has studied the biases in LLM moral value alignment, we aim to investigate cultural biases in LLM recommendations for real-world entities. To do so, we use the WEIRD (Western, Educated, Industrialized, Rich and Democratic) framework to evaluate recommendations by various LLMs across a dataset of fine-grained entities, and apply pluralistic prompt-based strategies to mitigate these biases. Our results indicate that while such prompting strategies do reduce such biases, this reduction is not consistent across different models, and recommendations for some types of entities are more biased than others.
Large Language Models (LLMs) demonstrate partial forecasting competence across social, political, and economic events. Yet, their predictive ability varies sharply with domain structure and prompt framing. We investigate how forecasting performance varies with different model families on real-world questions about events that happened beyond the model cutoff date. We analyze how context, question type, and external knowledge affect accuracy and calibration, and how adding factual news context modifies belief formation and failure modes. Our results show that forecasting ability is highly variable as it depends on what, and how, we ask.
In open source software development, the reuse of existing artifacts has been widely adopted to avoid redundant implementation work. Reusable artifacts are considered more efficient and reliable than developing software components from scratch. However, when faced with a large number of reusable artifacts, developers often struggle to find artifacts that can meet their expected needs. To reduce this burden, retrieval-based and learning-based techniques have been proposed to automate artifact recommendations. Recently, Large Language Models (LLMs) have shown the potential to understand intentions, perform semantic alignment, and recommend usable artifacts. Nevertheless, their effectiveness has not been thoroughly explored. To fill this gap, we construct an intent-driven artifact recommendation benchmark named IntentRecBench, covering three representative open source ecosystems. Using IntentRecBench, we conduct a comprehensive comparative study of five popular LLMs and six traditional approaches in terms of precision and efficiency. Our results show that although LLMs outperform traditional methods, they still suffer from low precision and high inference cost due to the large candidate space. Inspired by the ontology-based semantic organization in software engineering, we propose TreeRec, a feature tree-guided recommendation framework to mitigate these issues. TreeRec leverages LLM-based semantic abstraction to organize artifacts into a hierarchical semantic tree, enabling intent and function alignment and reducing reasoning time. Extensive experiments demonstrate that TreeRec consistently improves the performance of diverse LLMs across ecosystems, highlighting its generalizability and potential for practical deployment.
LLMs' code generation capabilities have yielded substantial improvements in the effectiveness of programming tasks. However, LLM-generated code still suffers from compilation and runtime errors. Existing offline preference optimization methods primarily focus on enhancing LLMs' coding abilities using pass/fail signals in the preference data, overlooking the deep-level error types in the failed codes. To address this, we propose Adaptively Progressive Preference Optimization (AP2O) for coding (i.e., AP2O-Coder), a method that guides LLMs adaptively and methodically to reduce code errors for code generation. Specifically, we construct an error notebook from failed codes and progressively optimize the LLM to correct errors type by type. Furthermore, we adaptively replay error types to tailor to the LLM's changing weaknesses throughout the training process. Through extensive experiments on both code and general LLMs (Llama, Qwen, and DeepSeek series) with parameters ranging from 0.5B to 34B, our AP2O-Coder improves code generation performance by up to 3% in pass@k while using less preference data. Code: https://github.com/TsingZ0/AP2O
Large language models (LLMs) are rapidly reshaping software development, but their impact across the software development lifecycle is underexplored. Existing work focuses on isolated activities such as code generation or testing, leaving open questions about how LLMs affect developers, processes, products, and the software ecosystem. We address this gap through semi-structured interviews with sixteen early-adopter software professionals who integrated LLM-based tools into their day-to-day work in early to mid-2023. We treat these interviews as early empirical evidence and compare participants' accounts with recent work on LLMs in software engineering, noting which early patterns persist or shift. Using thematic analysis, we organize findings around four dimensions: people, process, product, and society. Developers reported substantial productivity gains from reducing routine tasks, streamlining search, and accelerating debugging, but also described a productivity-quality paradox: they often discarded generated code and shifted effort from writing to critically evaluating and integrating it. LLM use was highly phase-dependent, with strong uptake in implementation and debugging but limited influence on requirements gathering and other collaborative work. Participants developed new competencies to use LLMs effectively, including prompt engineering strategies, layered verification, and secure integration to protect proprietary data. They anticipated changes in hiring expectations, team practices, and computing education, while emphasizing that human judgment and foundational software engineering skills remain essential. Our findings, later echoed in large-scale quantitative studies, offer actionable implications for developers, organizations, educators, and tool designers seeking to integrate LLMs responsibly into software practice today.
Large Language Model agents often retrieve context from knowledge bases that lack structural consistency with the agent's current reasoning state, leading to incoherent reasoning chains. We introduce Path-Constrained Retrieval (PCR), a retrieval method that combines structural graph constraints with semantic search to ensure retrieved information maintains logical relationships within a knowledge graph. PCR restricts the search space to nodes reachable from an anchor node, preventing retrieval of structurally disconnected information that may lead to inconsistent reasoning. We evaluate PCR on PathRAG-6, a benchmark spanning six domains with 180 nodes and 360 edges. Our results show that PCR achieves full structural consistency compared to 24-32 percent in baseline methods, while maintaining strong relevance scores. On the technology domain, PCR obtains full relevance at rank 10 with full structural consistency, significantly outperforming vector search and hybrid retrieval. PCR reduces the average graph distance of retrieved context by 78 percent compared to baselines, demonstrating retrieval of more structurally consistent information. These findings suggest that path-constrained retrieval is an effective approach for improving the reliability and coherence of LLM agent reasoning systems.
Recent research has explored the use of Large Language Models (LLMs) for tackling complex graph reasoning tasks. However, due to the intricacies of graph structures and the inherent limitations of LLMs in handling long text, current approaches often fail to deliver satisfactory accuracy, even on small-scale graphs and simple tasks. To address these challenges, we introduce GraphAgent-Reasoner, a fine-tuning-free framework that utilizes a multi-agent collaboration strategy for explicit and precise graph reasoning. Inspired by distributed graph computation theory, our framework decomposes graph problems into smaller, node-centric tasks that are distributed among multiple agents. The agents collaborate to solve the overall problem, significantly reducing the amount of information and complexity handled by a single LLM, thus enhancing the accuracy of graph reasoning. By simply increasing the number of agents, GraphAgent-Reasoner can efficiently scale to accommodate larger graphs with over 1,000 nodes. Evaluated on the GraphInstruct dataset, our framework demonstrates near-perfect accuracy on polynomial-time graph reasoning tasks, significantly outperforming the best available models, both closed-source and fine-tuned open-source variants. Our framework also demonstrates the capability to handle real-world graph reasoning applications such as webpage importance analysis.
Autonomous aerial vehicles (AAVs) have played a pivotal role in coverage operations and search missions. Recent advances in large language models (LLMs) offer promising opportunities to augment AAV intelligence. These advances help address complex challenges like area coverage optimization, dynamic path planning, and adaptive decision-making. However, the absence of physical grounding in LLMs leads to hallucination and reproducibility problems in spatial reasoning and decision-making. To tackle these issues, we present Skypilot, an LLM-enhanced two-stage framework that grounds language models in physical reality by integrating monte carlo tree search (MCTS). In the first stage, we introduce a diversified action space that encompasses generate, regenerate, fine-tune, and evaluate operations, coupled with physics-informed reward functions to ensure trajectory feasibility. In the second stage, we fine-tune Qwen3-4B on 23,000 MCTS-generated samples, achieving substantial inference acceleration while maintaining solution quality. Extensive numerical simulations and real-world flight experiments validate the efficiency and superiority of our proposed approach. Detailed information and experimental results are accessible at https://sky-pilot.top.
Large Language Models (LLMs) have shown significant potential for improving recommendation systems through their inherent reasoning capabilities and extensive knowledge base. Yet, existing studies predominantly address warm-start scenarios with abundant user-item interaction data, leaving the more challenging cold-start scenarios, where sparse interactions hinder traditional collaborative filtering methods, underexplored. To address this limitation, we propose novel reasoning strategies designed for cold-start item recommendations within the Netflix domain. Our method utilizes the advanced reasoning capabilities of LLMs to effectively infer user preferences, particularly for newly introduced or rarely interacted items. We systematically evaluate supervised fine-tuning, reinforcement learning-based fine-tuning, and hybrid approaches that combine both methods to optimize recommendation performance. Extensive experiments on real-world data demonstrate significant improvements in both methodological efficacy and practical performance in cold-start recommendation contexts. Remarkably, our reasoning-based fine-tuned models outperform Netflix's production ranking model by up to 8% in certain cases.
Metamorphic Relations (MRs) serve as a foundational mechanism for generating semantically equivalent mutations. Software engineering has advanced significantly in recent years with the advent of Large Language Models (LLMs). However, the reliability of LLMs in software engineering is often compromised by ambiguities and inconsistencies due to improper user specification. To address this challenge, we present CodeMetaAgent (CMA), a metamorphic relation-driven LLM agent that systematically refines task specifications and generates semantically constrained test cases. Our proposed framework uses MRs with LLMs to improve generation consistency and reduce variability caused by specifications, unlike the traditional use of MRs as post validations. Our framework has been evaluated on the HumanEval-Pro, MBPP-Pro, and SWE-Bench_Lite datasets using the GPT-4o, Mistral Large, GPT-OSS, and Qwen3-Coder models. It improved code generation accuracy by up to 17% and achieved code coverage gains of up to 99.81%. These results show that metamorphic relations can be a simple but effective guide in assisting LLM-based software development.
Sensor data streams provide valuable information around activities and context for downstream applications, though integrating complementary information can be challenging. We show that large language models (LLMs) can be used for late fusion for activity classification from audio and motion time series data. We curated a subset of data for diverse activity recognition across contexts (e.g., household activities, sports) from the Ego4D dataset. Evaluated LLMs achieved 12-class zero- and one-shot classification F1-scores significantly above chance, with no task-specific training. Zero-shot classification via LLM-based fusion from modality-specific models can enable multimodal temporal applications where there is limited aligned training data for learning a shared embedding space. Additionally, LLM-based fusion can enable model deploying without requiring additional memory and computation for targeted application-specific multimodal models.
Public health agencies face critical challenges in identifying high-risk neighborhoods for childhood lead exposure with limited resources for outreach and intervention programs. To address this, we develop a Priority Score integrating untested children proportions, elevated blood lead prevalence, and public health coverage patterns to support optimized resource allocation decisions across 136 neighborhoods in Chicago, New York City, and Washington, D.C. We leverage these allocation tasks, which require integrating multiple vulnerability indicators and interpreting empirical evidence, to evaluate whether large language models (LLMs) with agentic reasoning and deep research capabilities can effectively allocate public health resources when presented with structured allocation scenarios. LLMs were tasked with distributing 1,000 test kits within each city based on neighborhood vulnerability indicators. Results reveal significant limitations: LLMs frequently overlooked neighborhoods with highest lead prevalence and largest proportions of untested children, such as West Englewood in Chicago, while allocating disproportionate resources to lower-priority areas like Hunts Point in New York City. Overall accuracy averaged 0.46, reaching a maximum of 0.66 with ChatGPT 5 Deep Research. Despite their marketed deep research capabilities, LLMs struggled with fundamental limitations in information retrieval and evidence-based reasoning, frequently citing outdated data and allowing non-empirical narratives about neighborhood conditions to override quantitative vulnerability indicators.
As the number of connected IoT devices continues to grow, securing these systems against cyber threats remains a major challenge, especially in environments with limited computational and energy resources. This paper presents an edge-centric Intrusion Detection System (IDS) framework that integrates lightweight machine learning (ML) based IDS models with pre-trained large language models (LLMs) to improve detection accuracy, semantic interpretability, and operational efficiency at the network edge. The system evaluates six ML-based IDS models: Decision Tree (DT), K-Nearest Neighbors (KNN), Random Forest (RF), Convolutional Neural Network (CNN), Long Short-Term Memory (LSTM), and a hybrid CNN-LSTM model on low-power edge gateways, achieving accuracy up to 98 percent under real-world cyberattacks. For anomaly detection, the system transmits a compact and secure telemetry snapshot (for example, CPU usage, memory usage, latency, and energy consumption) via low-bandwidth API calls to LLMs including GPT-4-turbo, DeepSeek V2, and LLaMA 3.5. These models use zero-shot, few-shot, and chain-of-thought reasoning to produce human-readable threat analyses and actionable mitigation recommendations. Evaluations across diverse attacks such as DoS, DDoS, brute force, and port scanning show that the system enhances interpretability while maintaining low latency (<1.5 s), minimal bandwidth usage (<1.2 kB per prompt), and energy efficiency (<75 J), demonstrating its practicality and scalability as an IDS solution for edge gateways.
The integration of Large Language Models (LLMs) with Evolutionary Computation (EC) has unlocked new frontiers in scientific discovery but remains shackled by a fundamental constraint: the reliance on an Oracle--an objective, machine-computable fitness function. This paper breaks this barrier by asking: Can evolution thrive in a purely subjective landscape governed solely by LLM judges? We introduce MADE (Multi-Agent Decomposed Evolution), a framework that tames the inherent noise of subjective evaluation through "Problem Specification." By decomposing vague instructions into specific, verifiable sub-requirements, MADE transforms high-variance LLM feedback into stable, precise selection pressure. The results are transformative: across complex benchmarks like DevAI and InfoBench, MADE outperforms strong baselines by over 50% in software requirement satisfaction (39.9% to 61.9%) and achieves a 95% perfect pass rate on complex instruction following. This work validates a fundamental paradigm shift: moving from optimizing "computable metrics" to "describable qualities," thereby unlocking evolutionary optimization for the vast open-ended domains where no ground truth exists.
Organization's communication is essential for public trust, but the rise of generative AI models has introduced significant challenges by generating persuasive content that can form competing narratives with official messages from government and commercial organizations at speed and scale. This has left agencies in a reactive position, often unaware of how these models construct their persuasive strategies, making it more difficult to sustain communication effectiveness. In this paper, we introduce a large LLM-generated persuasion attack dataset, which includes 134,136 attacks generated by GPT-4, Gemma 2, and Llama 3.1 on agency news. These attacks span 23 persuasive techniques from SemEval 2023 Task 3, directed toward 972 press releases from ten agencies. The generated attacks come in two mediums, press release statements and social media posts, covering both long-form and short-form communication strategies. We analyzed the moral resonance of these persuasion attacks to understand their attack vectors. GPT-4's attacks mainly focus on Care, with Authority and Loyalty also playing a role. Gemma 2 emphasizes Care and Authority, while Llama 3.1 centers on Loyalty and Care. Analyzing LLM-generated persuasive attacks across models will enable proactive defense, allow to create the reputation armor for organizations, and propel the development of both effective and resilient communications in the information ecosystem.
Large Language Models can break through knowledge and timeliness limitations by invoking external tools within the Model Context Protocol framework to achieve automated execution of complex tasks. However, with the rapid growth of enterprise-scale MCP services, efficiently and accurately matching target functionalities among thousands of heterogeneous tools has become a core challenge restricting system practicality. Existing approaches generally rely on full-prompt injection or static semantic retrieval, facing issues including semantic disconnection between user queries and tool descriptions, context inflation in LLM input, and high inference latency. To address these challenges, this paper proposes Z-Space, a data-generation-oriented multi-agent collaborative tool invocation framework Z-Space. The Z-Space framework establishes a multi-agent collaborative architecture and tool filtering algorithm: (1) A structured semantic understanding of user queries is achieved through an intent parsing model; (2) A tool filtering module (FSWW) based on fused subspace weighted algorithm realizes fine-grained semantic alignment between intents and tools without parameter tuning; (3) An inference execution agent is constructed to support dynamic planning and fault-tolerant execution for multi-step tasks. This framework has been deployed in the Eleme platform's technical division, serving large-scale test data generation scenarios across multiple business units including Taotian, Gaode, and Hema. Production data demonstrates that the system reduces average token consumption in tool inference by 96.26\% while achieving a 92\% tool invocation accuracy rate, significantly enhancing the efficiency and reliability of intelligent test data generation systems.
With the rapid rise of personalized AI, customized large language models (LLMs) equipped with Chain of Thought (COT) reasoning now power millions of AI agents. However, their complex reasoning processes introduce new and largely unexplored security vulnerabilities. We present DarkMind, a novel latent reasoning level backdoor attack that targets customized LLMs by manipulating internal COT steps without altering user queries. Unlike prior prompt based attacks, DarkMind activates covertly within the reasoning chain via latent triggers, enabling adversarial behaviors without modifying input prompts or requiring access to model parameters. To achieve stealth and reliability, we propose dual trigger types instant and retrospective and integrate them within a unified embedding template that governs trigger dependent activation, employ a stealth optimization algorithm to minimize semantic drift, and introduce an automated conversation starter for covert activation across domains. Comprehensive experiments on eight reasoning datasets spanning arithmetic, commonsense, and symbolic domains, using five LLMs, demonstrate that DarkMind consistently achieves high attack success rates. We further investigate defense strategies to mitigate these risks and reveal that reasoning level backdoors represent a significant yet underexplored threat, underscoring the need for robust, reasoning aware security mechanisms.
Fully Homomorphic Encryption (FHE), particularly the CKKS scheme, is a promising enabler for privacy-preserving MLaaS, but its practical deployment faces a prohibitive barrier: it heavily relies on domain expertise. Configuring CKKS involves a tightly coupled space of ring dimensions, modulus chains, and packing layouts. Without deep cryptographic knowledge to navigate these interactions, practitioners are restricted to compilers that rely on fixed heuristics. These "one-shot" tools often emit rigid configurations that are either severely over-provisioned in latency or fail to find a feasible solution entirely for deeper networks. We present FHE-Agent, an agentic framework that automates this expert reasoning process. By coupling a Large Language Model (LLM) controller with a deterministic tool suite, FHE-Agent decomposes the search into global parameter selection and layer-wise bottleneck repair. The agents operate within a multi-fidelity workflow, pruning invalid regimes using cheap static analysis and reserving expensive encrypted evaluations for the most promising candidates. We instantiate FHE-Agent on the Orion compiler and evaluate it on standard benchmarks (MLP, LeNet, LoLa) and deeper architectures (AlexNet). FHE-Agent consistently achieves better precision and lower latency than na茂ve search strategies. Crucially, it automatically discovers feasible, 128-bit secure configurations for complex models where baseline heuristics and one-shot prompts fail to produce a valid setup.
Large Language Models (LLMs) have shown impressive potential to simulate human behavior. We identify a fundamental challenge in using them to simulate experiments: when LLM-simulated subjects are blind to the experimental design (as is standard practice with human subjects), variations in treatment systematically affect unspecified variables that should remain constant, violating the unconfoundedness assumption. Using demand estimation as a context and an actual experiment with 40 different products as a benchmark, we show this can lead to implausible results. While confounding may in principle be addressed by controlling for covariates, this can compromise ecological validity in the context of LLM simulations: controlled covariates become artificially salient in the simulated decision process. We show formally that confoundness stems from ambiguous prompting strategies. Therefore, it can be addressed by developing unambiguous prompting strategies through unblinding, i.e., revealing the experiment design in LLM simulations. Our empirical results show that this strategy consistently enhances model performance across all tested models, including both out-of-box reasoning and non-reasoning models. We also show that it is a technique that complements fine-tuning: while fine-tuning can improve simulation performance, an unambiguous prompting strategy makes the predictions robust to the inclusion of irrelevant data in the fine-tuning process.
How Far Can LLMs Emulate Human Behavior?: A Strategic Analysis via the Buy-and-Sell Negotiation Game
With the rapid advancement of Large Language Models (LLMs), recent studies have drawn attention to their potential for handling not only simple question-answer tasks but also more complex conversational abilities and performing human-like behavioral imitations. In particular, there is considerable interest in how accurately LLMs can reproduce real human emotions and behaviors, as well as whether such reproductions can function effectively in real-world scenarios. However, existing benchmarks focus primarily on knowledge-based assessment and thus fall short of sufficiently reflecting social interactions and strategic dialogue capabilities. To address these limitations, this work proposes a methodology to quantitatively evaluate the human emotional and behavioral imitation and strategic decision-making capabilities of LLMs by employing a Buy and Sell negotiation simulation. Specifically, we assign different personas to multiple LLMs and conduct negotiations between a Buyer and a Seller, comprehensively analyzing outcomes such as win rates, transaction prices, and SHAP values. Our experimental results show that models with higher existing benchmark scores tend to achieve better negotiation performance overall, although some models exhibit diminished performance in scenarios emphasizing emotional or social contexts. Moreover, competitive and cunning traits prove more advantageous for negotiation outcomes than altruistic and cooperative traits, suggesting that the assigned persona can lead to significant variations in negotiation strategies and results. Consequently, this study introduces a new evaluation approach for LLMs' social behavior imitation and dialogue strategies, and demonstrates how negotiation simulations can serve as a meaningful complementary metric to measure real-world interaction capabilities-an aspect often overlooked in existing benchmarks.
Large language models perform surprisingly well on many zero-shot classification tasks, but are difficult to fairly compare to supervised classifiers due to the lack of a modifiable decision boundary. In this work, we propose and evaluate a method that transforms binary classification tasks into pairwise comparisons between instances within a dataset, using LLMs to produce relative rankings of those instances. Repeated pairwise comparisons can be used to score instances using the Elo rating system (used in chess and other competitions), inducing a confidence ordering over instances in a dataset. We evaluate scheduling algorithms for their ability to minimize comparisons, and show that our proposed algorithm leads to improved classification performance, while also providing more information than traditional zero-shot classification.
Large language models (LLMs) based on the Transformer architecture usually have their context length limited due to the high training cost. Recent advancements extend the context window by adjusting the scaling factors of RoPE and fine-tuning. However, suboptimal initialization of these factors results in increased fine-tuning costs and reduced performance at target length. To address these challenges, we propose a novel RoPE-based fine-tuning framework that diverges from conventional scaling factors search. Specifically, we present a \textbf{D}ivide-and-\textbf{C}onquer \textbf{I}ncremental \textbf{S}earch (DCIS) algorithm that strategically determines the better scaling factors. Further fine-tuning with the identified scaling factors effectively extends the context window of LLMs. Empirical results demonstrate that our methodology not only mitigates performance decay at extended target lengths but also allows the model to fine-tune on short contexts and generalize to long contexts, thereby reducing the cost of fine-tuning. The scaling factors obtained through DCIS can even perform effectively without fine-tuning. Further analysis of the search space reveals that DCIS achieves twice the search efficiency compared to other methods. We also examine the impact of the non-strictly increasing scaling factors utilized in DCIS and evaluate the general capabilities of LLMs across various context lengths.
Large language models (LLMs) show promise in automating clinical diagnosis, yet their non-transparent decision-making and limited alignment with diagnostic standards hinder trust and clinical adoption. We address this challenge by proposing a two-stage diagnostic framework that enhances transparency, trustworthiness, and reliability. First, we introduce Evidence-Guided Diagnostic Reasoning (EGDR), which guides LLMs to generate structured diagnostic hypotheses by interleaving evidence extraction with logical reasoning grounded in DSM-5 criteria. Second, we propose a Diagnosis Confidence Scoring (DCS) module that evaluates the factual accuracy and logical consistency of generated diagnoses through two interpretable metrics: the Knowledge Attribution Score (KAS) and the Logic Consistency Score (LCS). Evaluated on the D4 dataset with pseudo-labels, EGDR outperforms direct in-context prompting and Chain-of-Thought (CoT) across five LLMs. For instance, on OpenBioLLM, EGDR improves accuracy from 0.31 (Direct) to 0.76 and increases DCS from 0.50 to 0.67. On MedLlama, DCS rises from 0.58 (CoT) to 0.77. Overall, EGDR yields up to +45% accuracy and +36% DCS gains over baseline methods, offering a clinically grounded, interpretable foundation for trustworthy AI-assisted diagnosis.
Mixture-of-Experts (MoE) architectures are increasingly adopted in large language models (LLMs) for their scalability and efficiency. However, their modular structure introduces a unique vulnerability: adversaries can attempt to compress or repurpose models by pruning experts and cheaply fine-tuning the remainder, effectively bypassing licensing and security constraints. In this paper, we systematically study the prunability of MoE-LLMs under task-specific usage. We first develop an expert attribution framework that identifies the subset of experts most responsible for a given task, then evaluate the performance trade-offs of pruning and re-aligning these experts using active learning-driven fine-tuning. Our findings reveal a critical knowledge loss--recovery trade-off: while certain experts can be isolated to retain task accuracy, significant degradation occurs without targeted re-alignment. Based on this analysis, we propose defense strategies that aim to make MoE models harder to compress and fine-tune without authorization, including entangled expert training and selective fine-tuning protocols that resist unauthorized adaptation. By positioning expert pruning as both a threat vector and a defense target, this work highlights the dual-use nature of MoE modularity and provides the first systematic evaluation framework for secure specialization of MoE-LLMs.
Heterogeneous graphs are widely present in real-world complex networks, where the diversity of node and relation types leads to complex and rich semantics. Efforts for modeling complex relation semantics in heterogeneous graphs are restricted by the limitations of predefined semantic dependencies and the scarcity of supervised signals. The advanced pre-training and fine-tuning paradigm leverages graph structure to provide rich self-supervised signals, but introduces semantic gaps between tasks. Large Language Models (LLMs) offer significant potential to address the semantic issues of relations and tasks in heterogeneous graphs through their strong reasoning capabilities in textual modality, but their incorporation into heterogeneous graphs is largely limited by computational complexity. Therefore, in this paper, we propose an Efficient LLM-Aware (ELLA) framework for heterogeneous graphs, addressing the above issues. To capture complex relation semantics, we propose an LLM-aware Relation Tokenizer that leverages LLM to encode multi-hop, multi-type relations. To reduce computational complexity, we further employ a Hop-level Relation Graph Transformer, which help reduces the complexity of LLM-aware relation reasoning from exponential to linear. To bridge semantic gaps between pre-training and fine-tuning tasks, we introduce the fine-grained task-aware textual Chain-of-Thought (CoT) prompts. Extensive experiments on four heterogeneous graphs show that our proposed ELLA outperforms state-of-the-art methods in the performance and efficiency. In particular, ELLA scales up to 13b-parameter LLMs and achieves up to a 4x speedup compared with existing LLM-based methods. Our code is publicly available at https://github.com/l-wd/ELLA.