llm - 2025_08
Navigation
Papers
Aerial Vision-and-Language Navigation (VLN) is a novel task enabling Unmanned Aerial Vehicles (UAVs) to navigate in outdoor environments through natural language instructions and visual cues. However, it remains challenging due to the complex spatial relationships in aerial scenes.In this paper, we propose a training-free, zero-shot framework for aerial VLN tasks, where the large language model (LLM) is leveraged as the agent for action prediction. Specifically, we develop a novel Semantic-Topo-Metric Representation (STMR) to enhance the spatial reasoning capabilities of LLMs. This is achieved by extracting and projecting instruction-related semantic masks onto a top-down map, which presents spatial and topological information about surrounding landmarks and grows during the navigation process. At each step, a local map centered at the UAV is extracted from the growing top-down map, and transformed into a ma trix representation with distance metrics, serving as the text prompt to LLM for action prediction in response to the given instruction. Experiments conducted in real and simulation environments have proved the effectiveness and robustness of our method, achieving absolute success rate improvements of 26.8% and 5.8% over current state-of-the-art methods on simple and complex navigation tasks, respectively. The dataset and code will be released soon.
Large language models (LLMs) often fail to ask effective questions under uncertainty, making them unreliable in domains where proactive information-gathering is essential for decision-making. We present ALignment via Fine-grained Attributes, (ALFA) a framework that improves LLM question-asking by (i) decomposing the notion of a "good" question into a set of theory-grounded attributes (e.g., clarity, relevance), (ii) controllably synthesizing attribute-specific question variations, and (iii) aligning models via preference-based optimization to explicitly learn to ask better questions along these fine-grained attributes. Focusing on clinical reasoning as a case study, we introduce the MediQ-AskDocs dataset, composed of 17k real-world clinical interactions augmented with 80k attribute-specific preference pairs of follow-up questions, as well as a novel expert-annotated interactive healthcare QA task to evaluate question-asking abilities. Models aligned with ALFA reduce diagnostic errors by 56.6% on MediQ-AskDocs compared to SoTA instruction-tuned LLMs, with a question-level win-rate of 64.4% and strong generalizability. Our findings suggest that explicitly guiding question-asking with structured, fine-grained attributes offers a scalable path to improve LLMs, especially in expert application domains.
Large Language Models (LLMs) have achieved remarkable results on a variety of mathematical benchmarks. However, concerns remain as to whether these successes reflect genuine reasoning or superficial pattern recognition. Common evaluation methods, which focus on the either the final answer or the reasoning process, fail to assess the entire problem-solving procedure. To address these limitations, we introduce SMART: a Self-Generating and Self-Validating Multi-Dimensional Assessment Framework, together with its corresponding benchmark, SMART-Bench. SMART decomposes the entire problem solving process into four distinct cognitive dimensions: Understanding, Reasoning, Arithmetic, and Reflection \& Refinement. Each dimension is evaluated independently through tailored tasks, enabling interpretable and fine-grained analysis of LLM behavior. We apply SMART to 21 state-of-the-art open- and closed-source LLMs, uncovering significant discrepancies in their abilities across different dimensions. Our findings reveal genuine weaknesses in current LLMs and motivate a new metric, the All-Pass Score, to better capture true problem-solving capabilities. Code and benchmarks will be released upon acceptance.
Reinforcement learning with verifiable rewards (RLVR) has emerged as a powerful paradigm for enhancing the reasoning capabilities of large language models (LLMs). Unlike traditional RL approaches, RLVR leverages rule-based feedback to guide LLMs in generating and refining complex reasoning chains -- a process critically dependent on effective exploration strategies. While prior work has demonstrated RLVR's empirical success, the fundamental mechanisms governing LLMs' exploration behaviors remain underexplored. This technical report presents a systematic investigation of exploration capacities in RLVR, covering four main aspects: (1) exploration space shaping, where we develop quantitative metrics to characterize LLMs' capability boundaries; (2) entropy-performance exchange, analyzed across training stages, individual instances, and token-level patterns; and (3) RL performance optimization, examining methods to effectively translate exploration gains into measurable improvements. By unifying previously identified insights with new empirical evidence, this work aims to provide a foundational framework for advancing RLVR systems.
Word clouds are a common way to summarize qualitative interviews, yet traditional frequency-based methods often fail in conversational contexts: they surface filler words, ignore paraphrase, and fragment semantically related ideas. This limits their usefulness in early-stage analysis, when researchers need fast, interpretable overviews of what participant actually said. We introduce ThemeClouds, an open-source visualization tool that uses large language models (LLMs) to generate thematic, participant-weighted word clouds from dialogue transcripts. The system prompts an LLM to identify concept-level themes across a corpus and then counts how many unique participants mention each topic, yielding a visualization grounded in breadth of mention rather than raw term frequency. Researchers can customize prompts and visualization parameters, providing transparency and control. Using interviews from a user study comparing five recording-device configurations (31 participants; 155 transcripts, Whisper ASR), our approach surfaces more actionable device concerns than frequency clouds and topic-modeling baselines (e.g., LDA, BERTopic). We discuss design trade-offs for integrating LLM assistance into qualitative workflows, implications for interpretability and researcher agency, and opportunities for interactive analyses such as per-condition contrasts (``diff clouds'').
Effective recommender systems demand dynamic user understanding, especially in complex, evolving environments. Traditional user profiling often fails to capture the nuanced, temporal contextual factors of user preferences, such as transient short-term interests and enduring long-term tastes. This paper presents an assessment of Large Language Models (LLMs) for generating semantically rich, time-aware user profiles. We do not propose a novel end-to-end recommendation architecture; instead, the core contribution is a systematic investigation into the degree of LLM effectiveness in capturing the dynamics of user context by disentangling short-term and long-term preferences. This approach, framing temporal preferences as dynamic user contexts for recommendations, adaptively fuses these distinct contextual components into comprehensive user embeddings. The evaluation across Movies&TV and Video Games domains suggests that while LLM-generated profiles offer semantic depth and temporal structure, their effectiveness for context-aware recommendations is notably contingent on the richness of user interaction histories. Significant gains are observed in dense domains (e.g., Movies&TV), whereas improvements are less pronounced in sparse environments (e.g., Video Games). This work highlights LLMs' nuanced potential in enhancing user profiling for adaptive, context-aware recommendations, emphasizing the critical role of dataset characteristics for practical applicability.
Large language models (LLMs) have the potential to address social and behavioral determinants of health by transforming labor intensive workflows in resource-constrained settings. Creating LLM-based applications that serve the needs of underserved communities requires a deep understanding of their local context, but it is often the case that neither LLMs nor their developers possess this local expertise, and the experts in these communities often face severe time/resource constraints. This creates a disconnect: how can one engage in meaningful co-design of an LLM-based application for an under-resourced community when the communication channel between the LLM developer and domain expert is constrained? We explored this question through a real-world case study, in which our data science team sought to partner with social workers at a safety net hospital to build an LLM application that summarizes patients' social needs. Whereas prior works focus on the challenge of prompt tuning, we found that the most critical challenge in this setting is the careful and precise specification of \what information to surface to providers so that the LLM application is accurate, comprehensive, and verifiable. Here we present a novel co-design framework for settings with limited access to domain experts, in which the summary generation task is first decomposed into individually-optimizable attributes and then each attribute is efficiently refined and validated through a multi-tier cascading approach.
Small large language models (LLMs) often face difficulties in aligning output to human preferences, particularly when operating under severe performance gaps. In this work, we propose two lightweight DPO-based variants -- Adaptive Margin-Sigmoid Loss and APO-hinge-zero -- to better address underperformance scenarios by introducing margin-based objectives and selective update mechanisms. Our APO-hinge-zero method, which combines hinge-induced hard-example mining with the chosen-focused optimization of APO-zero, achieves strong results. In AlpacaEval, APO-hinge-zero improves the win rate by +2.0 points and the length-controlled win rate by +1.4 points compared to the APO-zero baseline. In MT-Bench, our methods maintain competitive performance in diverse categories, particularly excelling in STEM and Humanities tasks. These results demonstrate that simple modifications to preference-based objectives can significantly enhance small LLM alignment under resource constraints, offering a practical path toward more efficient deployment.
Long-context Large Language Model (LLM) inference faces increasing compute bottlenecks as attention calculations scale with context length, primarily due to the growing KV-cache transfer overhead that saturates High Bandwidth Memory (HBM). While prefetching techniques mitigate cache misses by fetching KV data in advance, their spatial and temporal benefits present new opportunities to exploit. This work proposes a packing-prefetch scheduling architecture with monolithic 3D (M3D) back-end-of-line (BEOL) compatible embedded memories with ultra-large on-chip capacity to accelerate long-context LLM inference. Our optimizations demonstrate 8.06x decode speedup and 1.83x overall latency reduction on Llama3.1-8B using TPUv6e-like hardware with additional 512MB BEOL memories over the serial execution. Evaluations of multi-request workloads on TPU-like architectures show 1.7x-2.4x throughput improvement and 1.5x-2.4x HBM bandwidth reduction compared to packing-only methods on Llama3.1-8B and Llama3.1-70B models. With the co-design of packing, prefetching, and BEOL memories, our approach alleviates HBM constraints and enables efficient long-context LLM inference.
Accurately modeling user preferences is crucial for improving the performance of content-based recommender systems. Existing approaches often rely on simplistic user profiling methods, such as averaging or concatenating item embeddings, which fail to capture the nuanced nature of user preference dynamics, particularly the interactions between long-term and short-term preferences. In this work, we propose LLM-driven Temporal User Profiling (LLM-TUP), a novel method for user profiling that explicitly models short-term and long-term preferences by leveraging interaction timestamps and generating natural language representations of user histories using a large language model (LLM). These representations are encoded into high-dimensional embeddings using a pre-trained BERT model, and an attention mechanism is applied to dynamically fuse the short-term and long-term embeddings into a comprehensive user profile. Experimental results on real-world datasets demonstrate that LLM-TUP achieves substantial improvements over several baselines, underscoring the effectiveness of our temporally aware user-profiling approach and the use of semantically rich user profiles, generated by LLMs, for personalized content-based recommendation.
Existing reasoning evaluation frameworks for Large Language Models (LLMs) and Large Vision-Language Models (LVLMs) predominantly assess either text-based reasoning or vision-language understanding capabilities, with limited dynamic interplay between textual and visual constraints. To address this limitation, we introduce CrossWordBench, a benchmark designed to evaluate the reasoning capabilities of both LLMs and LVLMs through the medium of crossword puzzles -- a task requiring multimodal adherence to semantic constraints from text-based clues and intersectional constraints from visual grid structures. CrossWordBench leverages a controllable puzzle generation framework that produces puzzles in two formats (text and image), supports adjustable difficulty through prefill ratio control, and offers different evaluation strategies, ranging from direct puzzle solving to interactive modes. Our extensive evaluation of over 20 models reveals that reasoning LLMs substantially outperform non-reasoning models by effectively leveraging crossing-letter constraints. We further demonstrate that LVLMs struggle with the task, showing a strong correlation between their puzzle-solving performance and grid-parsing accuracy. Our findings highlight limitations of the reasoning capabilities of current LLMs and LVLMs, and provide an effective approach for creating multimodal constrained tasks for future evaluations.
Global KV-cache sharing has emerged as a key optimization for accelerating large language model (LLM) inference. However, it exposes a new class of timing side-channel attacks, enabling adversaries to infer sensitive user inputs via shared cache entries. Existing defenses, such as per-user isolation, eliminate leakage but degrade performance by up to 38.9% in time-to-first-token (TTFT), making them impractical for high-throughput deployment. To address this gap, we introduce SafeKV (Secure and Flexible KV Cache Sharing), a privacy-aware KV-cache management framework that selectively shares non-sensitive entries while confining sensitive content to private caches. SafeKV comprises three components: (i) a hybrid, multi-tier detection pipeline that integrates rule-based pattern matching, a general-purpose privacy detector, and context-aware validation; (ii) a unified radix-tree index that manages public and private entries across heterogeneous memory tiers (HBM, DRAM, SSD); and (iii) entropy-based access monitoring to detect and mitigate residual information leakage. Our evaluation shows that SafeKV mitigates 94% - 97% of timing-based side-channel attacks. Compared to per-user isolation method, SafeKV improves TTFT by up to 40.58% and throughput by up to 2.66X across diverse LLMs and workloads. SafeKV reduces cache-induced TTFT overhead from 50.41% to 11.74% on Qwen3-235B. By combining fine-grained privacy control with high cache reuse efficiency, SafeKV reclaims the performance advantages of global sharing while providing robust runtime privacy guarantees for LLM inference.
Story visualization aims to create visually compelling images or videos corresponding to textual narratives. Despite recent advances in diffusion models yielding promising results, existing methods still struggle to create a coherent sequence of subject-consistent frames based solely on a story. To this end, we propose DreamStory, an automatic open-domain story visualization framework by leveraging the LLMs and a novel multi-subject consistent diffusion model. DreamStory consists of (1) an LLM acting as a story director and (2) an innovative Multi-Subject consistent Diffusion model (MSD) for generating consistent multi-subject across the images. First, DreamStory employs the LLM to generate descriptive prompts for subjects and scenes aligned with the story, annotating each scene's subjects for subsequent subject-consistent generation. Second, DreamStory utilizes these detailed subject descriptions to create portraits of the subjects, with these portraits and their corresponding textual information serving as multimodal anchors (guidance). Finally, the MSD uses these multimodal anchors to generate story scenes with consistent multi-subject. Specifically, the MSD includes Masked Mutual Self-Attention (MMSA) and Masked Mutual Cross-Attention (MMCA) modules. MMSA and MMCA modules ensure appearance and semantic consistency with reference images and text, respectively. Both modules employ masking mechanisms to prevent subject blending. To validate our approach and promote progress in story visualization, we established a benchmark, DS-500, which can assess the overall performance of the story visualization framework, subject-identification accuracy, and the consistency of the generation model. Extensive experiments validate the effectiveness of DreamStory in both subjective and objective evaluations. Please visit our project homepage at https://dream-xyz.github.io/dreamstory.
Serving LLM adapters has gained significant attention as an effective approach to adapt general-purpose language models to diverse, task-specific use cases. However, serving a wide range of adapters introduces several and substantial overheads, leading to performance degradation and challenges in optimal placement. To address these challenges, we present an analytical, AI-driven pipeline that accurately determines the optimal allocation of adapters in single-node setups. This allocation maximizes performance, effectively using GPU resources, while preventing request starvation. Crucially, the proposed allocation is given based on current workload patterns. These insights in single-node setups can be leveraged in multi-replica deployments for overall placement, load balancing and server configuration, ultimately enhancing overall performance and improving resource efficiency. Our approach builds on an in-depth analysis of LLM adapter serving, accounting for overheads and performance variability, and includes the development of the first Digital Twin capable of replicating online LLM-adapter serving systems with matching key performance metrics. The experimental results demonstrate that the Digital Twin achieves a SMAPE difference of no more than 5.5% in throughput compared to real results, and the proposed pipeline accurately predicts the optimal placement with minimal latency.
Unlimited, or so-called helpful-only language models are trained without safety alignment constraints and never refuse user queries. They are widely used by leading AI companies as internal tools for red teaming and alignment evaluation. For example, if a safety-aligned model produces harmful outputs similar to an unlimited model, this indicates alignment failures that require further attention. Despite their essential role in assessing alignment, such models are not available to the research community. We introduce Jinx, a helpful-only variant of popular open-weight LLMs. Jinx responds to all queries without refusals or safety filtering, while preserving the base model's capabilities in reasoning and instruction following. It provides researchers with an accessible tool for probing alignment failures, evaluating safety boundaries, and systematically studying failure modes in language model safety.
Group decision-making often suffers from uneven information sharing, hindering decision quality. While large language models (LLMs) have been widely studied as aids for individuals, their potential to support groups of users, potentially as facilitators, is relatively underexplored. We present a pre-registered randomized experiment with 1,475 participants assigned to 281 five-person groups completing a hidden profile task--selecting an optimal city for a hypothetical sporting event--under one of four facilitation conditions: no facilitation, a one-time message prompting information sharing, a human facilitator, or an LLM (GPT-4o) facilitator. We find that LLM facilitation increases information shared within a discussion by raising the minimum level of engagement with the task among group members, and that these gains come at limited cost in terms of participants' attitudes towards the task, their group, or their facilitator. Whether by human or AI, there is no significant effect of facilitation on the final decision outcome, suggesting that even substantial but partial increases in information sharing are insufficient to overcome the hidden profile effect studied. To support further research into how LLM-based interfaces can support the future of collaborative decision making, we release our experimental platform, the Group-AI Interaction Laboratory (GRAIL), as an open-source tool.
Evaluating the safety alignment of LLM responses in high-risk mental health dialogues is particularly difficult due to missing gold-standard answers and the ethically sensitive nature of these interactions. To address this challenge, we propose PsyCrisis-Bench, a reference-free evaluation benchmark based on real-world Chinese mental health dialogues. It evaluates whether the model responses align with the safety principles defined by experts. Specifically designed for settings without standard references, our method adopts a prompt-based LLM-as-Judge approach that conducts in-context evaluation using expert-defined reasoning chains grounded in psychological intervention principles. We employ binary point-wise scoring across multiple safety dimensions to enhance the explainability and traceability of the evaluation. Additionally, we present a manually curated, high-quality Chinese-language dataset covering self-harm, suicidal ideation, and existential distress, derived from real-world online discourse. Experiments on 3600 judgments show that our method achieves the highest agreement with expert assessments and produces more interpretable evaluation rationales compared to existing approaches. Our dataset and evaluation tool are publicly available to facilitate further research.
Reinforcement learning for LLM reasoning has rapidly emerged as a prominent research area, marked by a significant surge in related studies on both algorithmic innovations and practical applications. Despite this progress, several critical challenges remain, including the absence of standardized guidelines for employing RL techniques and a fragmented understanding of their underlying mechanisms. Additionally, inconsistent experimental settings, variations in training data, and differences in model initialization have led to conflicting conclusions, obscuring the key characteristics of these techniques and creating confusion among practitioners when selecting appropriate techniques. This paper systematically reviews widely adopted RL techniques through rigorous reproductions and isolated evaluations within a unified open-source framework. We analyze the internal mechanisms, applicable scenarios, and core principles of each technique through fine-grained experiments, including datasets of varying difficulty, model sizes, and architectures. Based on these insights, we present clear guidelines for selecting RL techniques tailored to specific setups, and provide a reliable roadmap for practitioners navigating the RL for the LLM domain. Finally, we reveal that a minimalist combination of two techniques can unlock the learning capability of critic-free policies using vanilla PPO loss. The results demonstrate that our simple combination consistently improves performance, surpassing strategies like GRPO and DAPO.
As large language models become increasingly capable at various writing tasks, their weakness at generating unique and creative content becomes a major liability. Although LLMs have the ability to generate text covering diverse topics, there is an overall sense of repetitiveness across texts that we aim to formalize and quantify via a similarity metric. The familiarity between documents arises from the persistence of underlying discourse structures. However, existing similarity metrics dependent on lexical overlap and syntactic patterns largely capture $\textit{content}$ overlap, thus making them unsuitable for detecting $\textit{structural}$ similarities. We introduce an abstraction based on linguistic theories in Questions Under Discussion (QUD) and question semantics to help quantify differences in discourse progression. We then use this framework to build $\textbf{QUDsim}$, a similarity metric that can detect discursive parallels between documents. Using QUDsim, we find that LLMs often reuse discourse structures (more so than humans) across samples, even when content differs. Furthermore, LLMs are not only repetitive and structurally uniform, but are also divergent from human authors in the types of structures they use.
Watermarking LLM-generated text is critical for content attribution and misinformation prevention. However, existing methods compromise text quality, require white-box model access and logit manipulation. These limitations exclude API-based models and multilingual scenarios. We propose SAEMark, a general framework for post-hoc multi-bit watermarking that embeds personalized messages solely via inference-time, feature-based rejection sampling without altering model logits or requiring training. Our approach operates on deterministic features extracted from generated text, selecting outputs whose feature statistics align with key-derived targets. This framework naturally generalizes across languages and domains while preserving text quality through sampling LLM outputs instead of modifying. We provide theoretical guarantees relating watermark success probability and compute budget that hold for any suitable feature extractor. Empirically, we demonstrate the framework's effectiveness using Sparse Autoencoders (SAEs), achieving superior detection accuracy and text quality. Experiments across 4 datasets show SAEMark's consistent performance, with 99.7% F1 on English and strong multi-bit detection accuracy. SAEMark establishes a new paradigm for scalable watermarking that works out-of-the-box with closed-source LLMs while enabling content attribution.
Large language models (LLMs) have transformed the way we access information. These models are often tuned to refuse to comply with requests that are considered harmful and to produce responses that better align with the preferences of those who control the models. To understand how this "censorship" works. We use representation engineering techniques to study open-weights safety-tuned models. We present a method for finding a refusal--compliance vector that detects and controls the level of censorship in model outputs. We also analyze recent reasoning LLMs, distilled from DeepSeek-R1, and uncover an additional dimension of censorship through "thought suppression". We show a similar approach can be used to find a vector that suppresses the model's reasoning process, allowing us to remove censorship by applying the negative multiples of this vector. Our code is publicly available at: https://github.com/hannahxchen/llm-censorship-steering
With the breakthrough progress of large language models (LLMs) in natural language processing and multimodal tasks, efficiently deploying them on resource-constrained edge devices has become a critical challenge. The Mixture of Experts (MoE) architecture enhances model capacity through sparse activation, but faces two major difficulties in practical deployment: (1) The presence of numerous outliers in activation distributions leads to severe degradation in quantization accuracy for both activations and weights, significantly impairing inference performance; (2) Under limited memory, efficient offloading and collaborative inference of expert modules struggle to balance latency and throughput. To address these issues, this paper proposes an efficient MoE edge deployment scheme based on Hessian-Aware Quantization (HAQ) and CPU-GPU collaborative inference. First, by introducing smoothed Hessian matrix quantization, we achieve joint 8-bit quantization of activations and weights, which significantly alleviates the accuracy loss caused by outliers while ensuring efficient implementation on mainstream hardware. Second, we design an expert-level collaborative offloading and inference mechanism, which, combined with expert activation path statistics, enables efficient deployment and scheduling of expert modules between CPU and GPU, greatly reducing memory footprint and inference latency. Extensive experiments validate the effectiveness of our method on mainstream large models such as the OPT series and Mixtral 8*7B: on datasets like Wikitext2 and C4, the inference accuracy of the low-bit quantized model approaches that of the full-precision model, while GPU memory usage is reduced by about 60%, and inference latency is significantly improved.
The rapid proliferation of Large Language Models (LLMs) has significantly contributed to the development of equitable AI systems capable of factual question-answering (QA). However, no known study tests the LLMs' robustness when presented with obfuscated versions of questions. To systematically evaluate these limitations, we propose a novel technique, ObfusQAte and, leveraging the same, introduce ObfusQA, a comprehensive, first of its kind, framework with multi-tiered obfuscation levels designed to examine LLM capabilities across three distinct dimensions: (i) Named-Entity Indirection, (ii) Distractor Indirection, and (iii) Contextual Overload. By capturing these fine-grained distinctions in language, ObfusQA provides a comprehensive benchmark for evaluating LLM robustness and adaptability. Our study observes that LLMs exhibit a tendency to fail or generate hallucinated responses when confronted with these increasingly nuanced variations. To foster research in this direction, we make ObfusQAte publicly available.
Multi-agent large language model (LLM) systems have shown strong potential in complex reasoning and collaborative decision-making tasks. However, most existing coordination schemes rely on static or full-context routing strategies, which lead to excessive token consumption, redundant memory exposure, and limited adaptability across interaction rounds. We introduce RCR-Router, a modular and role-aware context routing framework designed to enable efficient, adaptive collaboration in multi-agent LLMs. To our knowledge, this is the first routing approach that dynamically selects semantically relevant memory subsets for each agent based on its role and task stage, while adhering to a strict token budget. A lightweight scoring policy guides memory selection, and agent outputs are iteratively integrated into a shared memory store to facilitate progressive context refinement. To better evaluate model behavior, we further propose an Answer Quality Score metric that captures LLM-generated explanations beyond standard QA accuracy. Experiments on three multi-hop QA benchmarks -- HotPotQA, MuSiQue, and 2WikiMultihop -- demonstrate that RCR-Router reduces token usage (up to 30%) while improving or maintaining answer quality. These results highlight the importance of structured memory routing and output-aware evaluation in advancing scalable multi-agent LLM systems.
The autoregressive decoding in LLMs is the major inference bottleneck due to the memory-intensive operations and limited hardware bandwidth. 3D-stacked architecture is a promising solution with significantly improved memory bandwidth, which vertically stacked multi DRAM dies on top of logic die. However, our experiments also show the 3D-stacked architecture faces severer thermal issues compared to 2D architecture, in terms of thermal temperature, gradient and scalability. To better exploit the potential of 3D-stacked architecture, we present Tasa, a heterogeneous architecture with cross-stack thermal optimizations to balance the temperature distribution and maximize the performance under the thermal constraints. High-performance core is designed for compute-intensive operations, while high-efficiency core is used for memory-intensive operators, e.g. attention layers. Furthermore, we propose a bandwidth sharing scheduling to improve the bandwidth utilization in such heterogeneous architecture. Extensive thermal experiments show that our Tasa architecture demonstrates greater scalability compared with the homogeneous 3D-stacked architecture, i.e. up to 5.55 $\tccentigrade$, 9.37 $\tccentigrade$, and 7.91 $\tccentigrade$ peak temperature reduction for 48, 60, and 72 core configurations. Our experimental for Llama-65B and GPT-3 66B inferences also demonstrate 2.85x and 2.21x speedup are obtained over the GPU baselines and state-of-the-art heterogeneous PIM-based LLM accelerator
LLM inference on mobile devices faces extraneous challenges due to limited memory bandwidth and computational resources. To address these issues, speculative inference and processing-in-memory (PIM) techniques have been explored at the algorithmic and hardware levels. However, speculative inference results in more compute-intensive GEMM operations, creating new design trade-offs for existing GEMV-accelerated PIM architectures. Furthermore, there exists a significant amount of redundant draft tokens in tree-based speculative inference, necessitating efficient token management schemes to minimize energy consumption. In this work, we present LP-Spec, an architecture-dataflow co-design leveraging hybrid LPDDR5 performance-enhanced PIM architecture with draft token pruning and dynamic workload scheduling to accelerate LLM speculative inference. A near-data memory controller is proposed to enable data reallocation between DRAM and PIM banks. Furthermore, a data allocation unit based on the hardware-aware draft token pruner is developed to minimize energy consumption and fully exploit parallel execution opportunities. Compared to end-to-end LLM inference on other mobile solutions such as mobile NPUs or GEMV-accelerated PIMs, our LP-Spec achieves 13.21x, 7.56x, and 99.87x improvements in performance, energy efficiency, and energy-delay-product (EDP). Compared with prior AttAcc PIM and RTX 3090 GPU, LP-Spec can obtain 12.83x and 415.31x EDP reduction benefits.
Estimating individualized treatment effects from observational data presents a persistent challenge due to unmeasured confounding and structural bias. Causal Machine Learning (causal ML) methods, such as causal trees and doubly robust estimators, provide tools for estimating conditional average treatment effects. These methods have limited effectiveness in complex real-world environments due to the presence of latent confounders or those described in unstructured formats. Moreover, reliance on domain experts for confounder identification and rule interpretation introduces high annotation cost and scalability concerns. In this work, we proposed Large Language Model-based agents for automated confounder discovery and subgroup analysis that integrate agents into the causal ML pipeline to simulate domain expertise. Our framework systematically performs subgroup identification and confounding structure discovery by leveraging the reasoning capabilities of LLM-based agents, which reduces human dependency while preserving interpretability. Experiments on real-world medical datasets show that our proposed approach enhances treatment effect estimation robustness by narrowing confidence intervals and uncovering unrecognized confounding biases. Our findings suggest that LLM-based agents offer a promising path toward scalable, trustworthy, and semantically aware causal inference.
We demonstrate, for the first time, fully quantized training (FQT) of large language models (LLMs) using predominantly 4-bit floating-point (FP4) precision for weights, activations, and gradients on datasets up to 200 billion tokens. We extensively investigate key design choices for FP4, including block sizes, scaling formats, and rounding methods. Our analysis shows that the NVFP4 format, where each block of 16 FP4 values (E2M1) shares a scale represented in E4M3, provides optimal results. We use stochastic rounding for backward and update passes and round-to-nearest for the forward pass to enhance stability. Additionally, we identify a theoretical and empirical threshold for effective quantized training: when the gradient norm falls below approximately $\sqrt{3}$ times the quantization noise, quantized training becomes less effective. Leveraging these insights, we successfully train a 7-billion-parameter model on 256 Intel Gaudi2 accelerators. The resulting FP4-trained model achieves downstream task performance comparable to a standard BF16 baseline, confirming that FP4 training is a practical and highly efficient approach for large-scale LLM training. A reference implementation is supplied in https://github.com/Anonymous1252022/fp4-all-the-way .
We conduct a behavioral exploration of LoRA fine-tuned LLMs for Passage Reranking to understand how relevance signals are learned and deployed by Large Language Models. By fine-tuning Mistral-7B, LLaMA3.1-8B, and Pythia-6.9B on MS MARCO under diverse LoRA configurations, we investigate how relevance modeling evolves across checkpoints, the impact of LoRA rank (1, 2, 8, 32), and the relative importance of updated MHA vs. MLP components. Our ablations reveal which layers and projections within LoRA transformations are most critical for reranking accuracy. These findings offer fresh explanations into LoRA's adaptation mechanisms, setting the stage for deeper mechanistic studies in Information Retrieval. All models used in this study have been shared.
Large Language Models (LLMs) have recently demonstrated impressive capabilities in natural language processing due to their strong generalization and sequence modeling capabilities. However, their direct application to time series forecasting remains challenging due to two fundamental issues: the inherent heterogeneity of temporal patterns and the modality gap between continuous numerical signals and discrete language representations. In this work, we propose TALON, a unified framework that enhances LLM-based forecasting by modeling temporal heterogeneity and enforcing semantic alignment. Specifically, we design a Heterogeneous Temporal Encoder that partitions multivariate time series into structurally coherent segments, enabling localized expert modeling across diverse temporal patterns. To bridge the modality gap, we introduce a Semantic Alignment Module that aligns temporal features with LLM-compatible representations, enabling effective integration of time series into language-based models while eliminating the need for handcrafted prompts during inference. Extensive experiments on seven real-world benchmarks demonstrate that TALON achieves superior performance across all datasets, with average MSE improvements of up to 11\% over recent state-of-the-art methods. These results underscore the effectiveness of incorporating both pattern-aware and semantic-aware designs when adapting LLMs for time series forecasting. The code is available at: https://github.com/syrGitHub/TALON.
Large Language Models (LLMs) are increasingly used in systems that retrieve and summarize content from multiple sources, such as search engines and AI assistants. While these models enhance user experience by generating coherent summaries, they obscure the contributions of original content creators, raising concerns about credit attribution and compensation. We address the challenge of valuing individual documents used in LLM-generated summaries. We propose using Shapley values, a game-theoretic method that allocates credit based on each document's marginal contribution. Although theoretically appealing, Shapley values are expensive to compute at scale. We therefore propose Cluster Shapley, an efficient approximation algorithm that leverages semantic similarity between documents. By clustering documents using LLM-based embeddings and computing Shapley values at the cluster level, our method significantly reduces computation while maintaining attribution quality. We demonstrate our approach to a summarization task using Amazon product reviews. Cluster Shapley significantly reduces computational complexity while maintaining high accuracy, outperforming baseline methods such as Monte Carlo sampling and Kernel SHAP with a better efficient frontier. Our approach is agnostic to the exact LLM used, the summarization process used, and the evaluation procedure, which makes it broadly applicable to a variety of summarization settings.
Retrieval-augmented generation (RAG) demonstrates remarkable performance across tasks in open-domain question-answering. However, traditional search engines may retrieve shallow content, limiting the ability of LLMs to handle complex, multi-layered information. To address it, we introduce WebWalkerQA, a benchmark designed to assess the ability of LLMs to perform web traversal. It evaluates the capacity of LLMs to traverse a website's subpages to extract high-quality data systematically. We propose WebWalker, which is a multi-agent framework that mimics human-like web navigation through an explore-critic paradigm. Extensive experimental results show that WebWalkerQA is challenging and demonstrates the effectiveness of RAG combined with WebWalker, through the horizontal and vertical integration in real-world scenarios.
Fine-tuning-as-a-Service introduces a critical vulnerability where a few malicious examples mixed into the user's fine-tuning dataset can compromise the safety alignment of Large Language Models (LLMs). While a recognized paradigm frames safe fine-tuning as a multi-objective optimization problem balancing user task performance with safety alignment, we find existing solutions are critically sensitive to the harmful ratio, with defenses degrading sharply as harmful ratio increases. We diagnose that this failure stems from conflicting gradients, where the user-task update directly undermines the safety objective. To resolve this, we propose SafeGrad, a novel method that employs gradient surgery. When a conflict is detected, SafeGrad nullifies the harmful component of the user-task gradient by projecting it onto the orthogonal plane of the alignment gradient, allowing the model to learn the user's task without sacrificing safety. To further enhance robustness and data efficiency, we employ a KL-divergence alignment loss that learns the rich, distributional safety profile of the well-aligned foundation model. Extensive experiments show that SafeGrad provides state-of-the-art defense across various LLMs and datasets, maintaining robust safety even at high harmful ratios without compromising task fidelity.
Software correctness is ensured mathematically through formal verification, which involves the resources of generating formal requirement specifications and having an implementation that must be verified. Tools such as model-checkers and theorem provers ensure software correctness by verifying the implementation against the specification. Formal methods deployment is regularly enforced in the development of safety-critical systems e.g. aerospace, medical devices and autonomous systems. Generating these specifications from informal and ambiguous natural language requirements remains the key challenge. Our project, VERIFAI^{1}, aims to investigate automated and semi-automated approaches to bridge this gap, using techniques from Natural Language Processing (NLP), ontology-based domain modelling, artefact reuse, and large language models (LLMs). This position paper presents a preliminary synthesis of relevant literature to identify recurring challenges and prospective research directions in the generation of verifiable specifications from informal requirements.
Large Language Models (LLMs) are widely used for code generation. However, commercial models like ChatGPT require significant computing power, which leads to high energy use and carbon emissions. This has raised concerns about their environmental impact. In this study, we evaluate open-source Small Language Models (SLMs) trained explicitly for code generation and compare their performance and energy efficiency against large LLMs and efficient human-written Python code. The goal is to investigate whether SLMs can match the performance of LLMs on certain types of programming problems while producing more energy-efficient code. We evaluate 150 coding problems from LeetCode, evenly distributed across three difficulty levels: easy, medium, and hard. Our comparison includes three small open-source models, StableCode-3B, StarCoderBase-3B, and Qwen2.5-Coder-3B-Instruct, and two large commercial models, GPT-4.0 and DeepSeek-Reasoner. The generated code is evaluated using four key metrics: run-time, memory usage, energy consumption, and correctness. We use human-written solutions as a baseline to assess the quality and efficiency of the model-generated code. Results indicate that LLMs achieve the highest correctness across all difficulty levels, but SLMs are often more energy-efficient when their outputs are correct. In over 52% of the evaluated problems, SLMs consumed the same or less energy than LLMs.
The proliferation of large language models (LLMs) has driven the adoption of Mixture-of-Experts (MoE) architectures as a promising solution to scale model capacity while controlling computational costs. However, deploying MoE models in resource-constrained mobile edge computing environments presents significant challenges due to their large memory footprint and dynamic expert activation patterns. To address these challenges, we propose a novel dynamic resource-aware collaborative optimization framework that jointly optimizes expert aggregation granularity and offloading strategies based on real-time device resource states, network conditions, and input characteristics in mobile edge environments, denoted as CoMoE. In CoMoE, we first systematically analyze existing expert aggregation techniques, including expert parameter merging,knowledge distillation,and parameter sharing decomposition, identifying their limitations in dynamic mobile environments.We then investigate expert offloading strategies encompassing expert prediction and prefetching, expert caching and scheduling, and multi-tier storage architectures, revealing the interdependencies between routing decisions and offloading performance.The CoMoE incorporates adaptive scheduling mechanisms that respond to user mobility and varying network conditions, enabling efficient MoE deployment across heterogeneous edge devices. Extensive experiments on real mobile edge testbeds demonstrate that CoMoE achieves approximately 70% reduction in memory usage compared to baseline methods, 10.5% lower inference latency than existing expert offloading techniques, while maintaining model performance stability. For large-scale MoE models (e.g,7.4B-parameter Switch-Base-128), the CoMoE reduces memory requirements from 15.6GB to 4.7GB, enabling deployment on resource-constrained mobile edge devices that previously could only support much smaller models.
Contemporary language models are increasingly multilingual, but Chinese LLM developers must navigate complex political and business considerations of language diversity. Language policy in China aims at influencing the public discourse and governing a multi-ethnic society, and has gradually transitioned from a pluralist to a more assimilationist approach since 1949. We explore the impact of these influences on current language technology. We evaluate six open-source multilingual LLMs pre-trained by Chinese companies on 18 languages, spanning a wide range of Chinese, Asian, and Anglo-European languages. Our experiments show Chinese LLMs performance on diverse languages is indistinguishable from international LLMs. Similarly, the models' technical reports also show lack of consideration for pretraining data language coverage except for English and Mandarin Chinese. Examining Chinese AI policy, model experiments, and technical reports, we find no sign of any consistent policy, either for or against, language diversity in China's LLM development. This leaves a puzzling fact that while China regulates both the languages people use daily as well as language model development, they do not seem to have any policy on the languages in language models.
Language is a ubiquitous tool that is foundational to reasoning and collaboration, ranging from everyday interactions to sophisticated problem-solving tasks. The establishment of a common language can serve as a powerful asset in ensuring clear communication and understanding amongst agents, facilitating desired coordination and strategies. In this work, we extend the capabilities of large language models (LLMs) by integrating them with advancements in multi-agent decision-making algorithms. We propose a systematic framework for the design of multi-agentic large language models (LLMs), focusing on key integration practices. These include advanced prompt engineering techniques, the development of effective memory architectures, multi-modal information processing, and alignment strategies through fine-tuning algorithms. We evaluate these design choices through extensive ablation studies on classic game settings with significant underlying social dilemmas and game-theoretic considerations.
Large Language Models (LLMs) with inference-time scaling techniques show promise for code generation, yet face notable efficiency and scalability challenges. Construction-based tree-search methods suffer from rapid growth in tree size, high token consumption, and lack of anytime property. In contrast, improvement-based methods offer better performance but often struggle with uninformative reward signals and inefficient search strategies. In this work, we propose \textbf{ReLoc}, a unified local search framework which effectively performs step-by-step code revision. Specifically, ReLoc explores a series of local revisions through four key algorithmic components: initial code drafting, neighborhood code generation, candidate evaluation, and incumbent code updating, each of which can be instantiated with specific decision rules to realize different local search algorithms such as Hill Climbing (HC) or Genetic Algorithm (GA). Furthermore, we develop a specialized revision reward model that evaluates code quality based on revision distance to produce fine-grained preferences that guide the local search toward more promising candidates. Finally, our extensive experimental results demonstrate that our approach achieves superior performance across diverse code generation tasks, significantly outperforming both construction-based tree search as well as the state-of-the-art improvement-based code generation methods.
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.
Triple-S: A Collaborative Multi-LLM Framework for Solving Long-Horizon Implicative Tasks in Robotics
Leveraging Large Language Models (LLMs) to write policy code for controlling robots has gained significant attention. However, in long-horizon implicative tasks, this approach often results in API parameter, comments and sequencing errors, leading to task failure. To address this problem, we propose a collaborative Triple-S framework that involves multiple LLMs. Through In-Context Learning, different LLMs assume specific roles in a closed-loop Simplification-Solution-Summary process, effectively improving success rates and robustness in long-horizon implicative tasks. Additionally, a novel demonstration library update mechanism which learned from success allows it to generalize to previously failed tasks. We validate the framework in the Long-horizon Desktop Implicative Placement (LDIP) dataset across various baseline models, where Triple-S successfully executes 89% of tasks in both observable and partially observable scenarios. Experiments in both simulation and real-world robot settings further validated the effectiveness of Triple-S. Our code and dataset is available at: https://github.com/Ghbbbbb/Triple-S.
Large Language Models (LLMs) are widely adopted for automated code generation with promising results. Although prior research has assessed LLM-generated code and identified various quality issues -- such as redundancy, poor maintainability, and sub-optimal performance a systematic understanding and categorization of these inefficiencies remain unexplored. Without such knowledge, practitioners struggle to optimize LLM-generated code for real-world applications, limiting its adoption. This study can also guide improving code LLMs, enhancing the quality and efficiency of code generation. Therefore, in this study, we empirically investigate inefficiencies in LLM-generated code by state-of-the-art models, i.e., CodeLlama, DeepSeek-Coder, and CodeGemma. To do so, we analyze 492 generated code snippets in the HumanEval++ dataset. We then construct a taxonomy of inefficiencies in LLM-generated code that includes 5 categories General Logic, Performance, Readability, Maintainability, and Errors) and 19 subcategories of inefficiencies. We then validate the proposed taxonomy through an online survey with 58 LLM practitioners and researchers. Our study indicates that logic and performance-related inefficiencies are the most popular, relevant, and frequently co-occur and impact overall code quality inefficiency. Our taxonomy provides a structured basis for evaluating the quality LLM-generated code and guiding future research to improve code generation efficiency.
In this study, we wish to showcase the unique utility of large language models (LLMs) in financial semantic annotation and alpha signal discovery. Leveraging a corpus of company-related tweets, we use an LLM to automatically assign multi-label event categories to high-sentiment-intensity tweets. We align these labeled sentiment signals with forward returns over 1-to-7-day horizons to evaluate their statistical efficacy and market tradability. Our experiments reveal that certain event labels consistently yield negative alpha, with Sharpe ratios as low as -0.38 and information coefficients exceeding 0.05, all statistically significant at the 95\% confidence level. This study establishes the feasibility of transforming unstructured social media text into structured, multi-label event variables. A key contribution of this work is its commitment to transparency and reproducibility; all code and methodologies are made publicly available. Our results provide compelling evidence that social media sentiment is a valuable, albeit noisy, signal in financial forecasting and underscore the potential of open-source frameworks to democratize algorithmic trading research.
As the complexity of software systems continues to increase, the demand for automated testing and maintenance tools is growing exponentially. To meet this urgent need, we propose a new assertion generation method based on Hardware Description Language (HDL). This method combines a lightweight, parameter-adjustable large language model (LLM) with the Unsloth platform to automatically generate test cases, thereby significantly reducing training costs without sacrificing accuracy or generalization performance. Empirical evaluation shows that our method can efficiently generate assertions that strictly conform to the hardware logic. This framework provides a robust and flexible solution to modern software testing and maintenance challenges. https://github.com/liusu-orange/AutoAssert-1 and https://gitee.com/OpenBPU/auto-assert1 are the locations of the source code.
Design studies aim to create visualization solutions for real-world problems of different application domains. Recently, the emergence of large language models (LLMs) has introduced new opportunities to enhance the design study process, providing capabilities such as creative problem-solving, data handling, and insightful analysis. However, despite their growing popularity, there remains a lack of systematic understanding of how LLMs can effectively assist researchers in visualization-specific design studies. In this paper, we conducted a multi-stage qualitative study to fill this gap, involving 30 design study researchers from diverse backgrounds and expertise levels. Through in-depth interviews and carefully-designed questionnaires, we investigated strategies for utilizing LLMs, the challenges encountered, and the practices used to overcome them. We further compiled and summarized the roles that LLMs can play across different stages of the design study process. Our findings highlight practical implications to inform visualization practitioners, and provide a framework for leveraging LLMs to enhance the design study process in visualization research.
Recently, quantization has been widely used for the compression and acceleration of large language models (LLMs). Due to the outliers in LLMs, it is crucial to flatten weights and activations to minimize quantization error with equally spaced quantization points. Prior research explores various pre-quantization transformations to suppress outliers, such as per-channel scaling and Hadamard transformation. However, we observe that these transformed weights and activations can still exhibit steep and dispersed distributions. In this paper, we propose FlatQuant (Fast and Learnable Affine Transformation), a new post-training quantization approach that enhances the flatness of weights and activations. Our approach identifies optimal affine transformations for each linear layer, calibrated in hours via a lightweight objective. To reduce runtime overhead of affine transformation, we apply Kronecker product with two lightweight matrices, and fuse all operations in FlatQuant into a single kernel. Extensive experiments demonstrate that FlatQuant establishes a new state-of-the-art benchmark for quantization. For example, it achieves less than 1\% accuracy drop for W4A4 quantization on the LLaMA-3-70B model, surpassing SpinQuant by 7.5\%. Additionally, it provides up to 2.3x prefill speedup and 1.7x decoding speedup compared to the FP16 model. Code is available at: https://github.com/ruikangliu/FlatQuant.
Numerous benchmarks have been built to evaluate the domain-specific abilities of large language models (LLMs), highlighting the need for effective and efficient benchmark construction. Existing domain-specific benchmarks primarily focus on the scaling law, relying on massive corpora for supervised fine-tuning or generating extensive question sets for broad coverage. However, the impact of corpus and question-answer (QA) set design on the precision and recall of domain-specific LLMs remains unexplored. In this paper, we address this gap and demonstrate that the scaling law is not always the optimal principle for benchmark construction in specific domains. Instead, we propose Comp-Comp, an iterative benchmarking framework based on a comprehensiveness-compactness principle. Here, comprehensiveness ensures semantic recall of the domain, while compactness enhances precision, guiding both corpus and QA set construction. To validate our framework, we conducted a case study in a well-renowned university, resulting in the creation of XUBench, a large-scale and comprehensive closed-domain benchmark. Although we use the academic domain as the case in this work, our Comp-Comp framework is designed to be extensible beyond academia, providing valuable insights for benchmark construction across various domains.
Graph Neural Networks (GNNs) have emerged as powerful tools for learning over structured data, including text-attributed graphs, which are common in domains such as citation networks, social platforms, and knowledge graphs. GNNs are not inherently interpretable and thus, many explanation methods have been proposed. However, existing explanation methods often struggle to generate interpretable, fine-grained rationales, especially when node attributes include rich natural language. In this work, we introduce LOGIC, a lightweight, post-hoc framework that uses large language models (LLMs) to generate faithful and interpretable explanations for GNN predictions. LOGIC projects GNN node embeddings into the LLM embedding space and constructs hybrid prompts that interleave soft prompts with textual inputs from the graph structure. This enables the LLM to reason about GNN internal representations and produce natural language explanations along with concise explanation subgraphs. Our experiments across four real-world TAG datasets demonstrate that LOGIC achieves a favorable trade-off between fidelity and sparsity, while significantly improving human-centric metrics such as insightfulness. LOGIC sets a new direction for LLM-based explainability in graph learning by aligning GNN internals with human reasoning.
Global cloud service providers handle inference workloads for Large Language Models (LLMs) that span latency-sensitive (e.g., chatbots) and insensitive (e.g., report writing) tasks, resulting in diverse and often conflicting Service Level Agreement (SLA) requirements. Managing such mixed workloads is challenging due to the complexity of the inference serving stack, which encompasses multiple models, GPU hardware, and global data centers. Existing solutions often silo such fast and slow tasks onto separate GPU resource pools with different SLAs, but this leads to significant under-utilization of expensive accelerators due to load mismatch. In this article, we characterize the LLM serving workloads at Microsoft Office 365, one of the largest users of LLMs within Microsoft Azure cloud with over 10 million requests per day, and highlight key observations across workloads in different data center regions and across time. This is one of the first such public studies of Internet-scale LLM workloads. We use these insights to propose SageServe, a comprehensive LLM serving framework that dynamically adapts to workload demands using multi-timescale control knobs. It combines short-term request routing to data centers with long-term scaling of GPU VMs and model placement with higher lead times, and co-optimizes the routing and resource allocation problem using a traffic forecast model and an Integer Linear Programming (ILP) solution. We evaluate SageServe through real runs and realistic simulations on 10 million production requests across three regions and four open-source models. We achieve up to 25% savings in GPU-hours compared to the current baseline deployment and reduce GPU-hour wastage due to inefficient auto-scaling by 80%, resulting in a potential monthly cost savings of up to $2.5 million, while maintaining tail latency and meeting SLAs.
Large Language Models (LLMs) struggle with dynamic knowledge updates, especially when new information conflicts with deeply embedded facts. Such conflicting factual edits often lead to two critical issues: resistance to adopting the new fact and severe catastrophic forgetting of unrelated knowledge. This paper introduces and evaluates a novel "unlearn-then-learn" strategy for precise knowledge editing in LLMs, leveraging the parameter-efficient fine-tuning (PEFT) technique, Infused Adapter by Inhibiting and Amplifying Inner Activations ($IA^3$). Crucially, this two-stage approach is powered by an initial circuit localization phase that identifies and targets the specific internal components responsible for encoding the conflicting fact. Through a rigorous experimental methodology on microsoft/Phi-3-mini-4k-instruct, we demonstrate that this mechanistically informed two-stage approach achieves near-perfect accuracy (98.50%) for the new, modulated fact while simultaneously effectively suppressing the original conflicting fact (96.00% forget rate). Critically, our strategy exhibits unprecedented localization (72.00% F_control accuracy), dramatically mitigating catastrophic forgetting observed in direct fine-tuning approaches (which showed as low as ~20% F_control accuracy), a direct benefit of our targeted interpretability-guided intervention. Furthermore, qualitative analysis reveals a nuanced mechanism of "soft forgetting," where original knowledge is suppressed from default retrieval but remains latent and conditionally accessible, enhancing model safety and control. These findings represent a significant advancement towards precise, localized, and safe knowledge management in compact LLMs.
Purpose: This study proposes a framework for fine-tuning large language models (LLMs) with differential privacy (DP) to perform multi-abnormality classification on radiology report text. By injecting calibrated noise during fine-tuning, the framework seeks to mitigate the privacy risks associated with sensitive patient data and protect against data leakage while maintaining classification performance. Materials and Methods: We used 50,232 radiology reports from the publicly available MIMIC-CXR chest radiography and CT-RATE computed tomography datasets, collected between 2011 and 2019. Fine-tuning of LLMs was conducted to classify 14 labels from MIMIC-CXR dataset, and 18 labels from CT-RATE dataset using Differentially Private Low-Rank Adaptation (DP-LoRA) in high and moderate privacy regimes (across a range of privacy budgets = {0.01, 0.1, 1.0, 10.0}). Model performance was evaluated using weighted F1 score across three model architectures: BERT-medium, BERT-small, and ALBERT-base. Statistical analyses compared model performance across different privacy levels to quantify the privacy-utility trade-off. Results: We observe a clear privacy-utility trade-off through our experiments on 2 different datasets and 3 different models. Under moderate privacy guarantees the DP fine-tuned models achieved comparable weighted F1 scores of 0.88 on MIMIC-CXR and 0.59 on CT-RATE, compared to non-private LoRA baselines of 0.90 and 0.78, respectively. Conclusion: Differentially private fine-tuning using LoRA enables effective and privacy-preserving multi-abnormality classification from radiology reports, addressing a key challenge in fine-tuning LLMs on sensitive medical data.
As AI systems become more integrated into daily life, the need for safer and more reliable moderation has never been greater. Large Language Models (LLMs) have demonstrated remarkable capabilities, surpassing earlier models in complexity and performance. Their evaluation across diverse tasks has consistently showcased their potential, enabling the development of adaptive and personalized agents. However, despite these advancements, LLMs remain prone to errors, particularly in areas requiring nuanced moral reasoning. They struggle with detecting implicit hate, offensive language, and gender biases due to the subjective and context-dependent nature of these issues. Moreover, their reliance on training data can inadvertently reinforce societal biases, leading to inconsistencies and ethical concerns in their outputs. To explore the limitations of LLMs in this role, we developed an experimental framework based on state-of-the-art (SOTA) models to assess human emotions and offensive behaviors. The framework introduces a unified benchmark dataset encompassing 49 distinct categories spanning the wide spectrum of human emotions, offensive and hateful text, and gender and racial biases. Furthermore, we introduced SafePhi, a QLoRA fine-tuned version of Phi-4, adapting diverse ethical contexts and outperforming benchmark moderators by achieving a Macro F1 score of 0.89, where OpenAI Moderator and Llama Guard score 0.77 and 0.74, respectively. This research also highlights the critical domains where LLM moderators consistently underperformed, pressing the need to incorporate more heterogeneous and representative data with human-in-the-loop, for better model robustness and explainability.
Recent advances in Knowledge Distillation (KD) aim to mitigate the high computational demands of Large Language Models (LLMs) by transferring knowledge from a large ''teacher'' to a smaller ''student'' model. However, students may inherit the teacher's privacy when the teacher is trained on private data. In this work, we systematically characterize and investigate membership and memorization privacy risks inherent in six LLM KD techniques. Using instruction-tuning settings that span seven NLP tasks, together with three teacher model families (GPT-2, LLAMA-2, and OPT), and various size student models, we demonstrate that all existing LLM KD approaches carry membership and memorization privacy risks from the teacher to its students. However, the extent of privacy risks varies across different KD techniques. We systematically analyse how key LLM KD components (KD objective functions, student training data and NLP tasks) impact such privacy risks. We also demonstrate a significant disagreement between memorization and membership privacy risks of LLM KD techniques. Finally, we characterize per-block privacy risk and demonstrate that the privacy risk varies across different blocks by a large margin.
While large language models (LLMs) are widely used for writing, users often struggle to express their nuanced and evolving intents through prompt-based interfaces. Intents -- low-level strategies or preferences for achieving a writing goal -- are often vague, fluid, or even subconscious, making it difficult for users to articulate and adjust them. To address this, we present IntentFlow, which supports the communication of dynamically evolving intents throughout LLM-assisted writing. IntentFlow extracts goals and intents from user prompts and presents them as editable interface components, which users can revise, remove, or refine via direct manipulation or follow-up prompts. Visual links connect each component to the output segments it influences, helping users understand model behavior. In a within-subjects study (N=12), participants using IntentFlow, compared to a chat-based baseline, expressed their intents more easily and in detail, engaged in more meaningful actions to communicate intents, such as adjusting and deleting, and produced outputs that better aligned with their evolving intents. We found that editable intent representations help users refine and consolidate a final set of intents, which can be reused across similar tasks to support consistent and transferable LLM-assisted writing.
Large Language Models (LLMs) applying Mixture-of-Experts (MoE) scale to trillions of parameters but require vast memory, motivating a line of research to offload expert weights from fast-but-small DRAM (HBM) to denser Flash SSDs. While SSDs provide cost-effective capacity, their read energy per bit is substantially higher than that of DRAM. This paper quantitatively analyzes the energy implications of offloading MoE expert weights to SSDs during the critical decode stage of LLM inference. Our analysis, comparing SSD, CPU memory (DDR), and HBM storage scenarios for models like DeepSeek-R1, reveals that offloading MoE weights to current SSDs drastically increases per-token-generation energy consumption (e.g., by up to ~12x compared to the HBM baseline), dominating the total inference energy budget. Although techniques like prefetching effectively hide access latency, they cannot mitigate this fundamental energy penalty. We further explore future technological scaling, finding that the inherent sparsity of MoE models could potentially make SSDs energy-viable if Flash read energy improves significantly, roughly by an order of magnitude.
Large Language Models (LLMs) face persistent and evolving trustworthiness issues, motivating developers to seek automated and flexible repair methods that enable convenient deployment across diverse scenarios. Existing repair methods like supervised fine-tuning (SFT) and reinforcement learning with human feedback (RLHF) are costly and slow, while prompt engineering lacks robustness and scalability. Representation engineering, which steers model behavior by injecting targeted concept vectors during inference, offers a lightweight, training-free alternative. However, current approaches depend on manually crafted samples and fixed steering strategies, limiting automation and adaptability. To overcome these challenges, we propose MASteer, the first end-to-end framework for trustworthiness repair in LLMs based on representation engineering. MASteer integrates two core components: AutoTester, a multi-agent system that generates diverse, high-quality steer samples tailored to developer needs; and AutoRepairer, which constructs adaptive steering strategies with anchor vectors for automated, context-aware strategy selection during inference. Experiments on standard and customized trustworthiness tasks show MASteer consistently outperforms baselines, improving metrics by 15.36% on LLaMA-3.1-8B-Chat and 4.21% on Qwen-3-8B-Chat, while maintaining general model capabilities. MASteer demonstrates strong robustness, generalization, and practical value for scalable, efficient trustworthiness repair.
Multi-agent applications utilize the advanced capabilities of large language models (LLMs) for intricate task completion through agent collaboration in a workflow. Under this situation, requests from different agents usually access the same shared LLM to perform different kinds of tasks, forcing the shared LLM to suffer excessive loads. However, existing works have low serving performance for these multi-agent applications, mainly due to the ignorance of inter-agent latency and resource differences for request scheduling. We therefore propose Kairos, a multi-agent orchestration system that optimizes end-to-end latency for multi-agent applications. Kairos consists of a workflow orchestrator, a workflow-aware priority scheduler, and a memory-aware dispatcher. The orchestrator collects agent-specific information for online workflow analysis. The scheduler decides the serving priority of the requests based on their latency characteristics to reduce the overall queuing. The dispatcher dispatches the requests to different LLM instances based on their memory demands to avoid GPU overloading. Experimental results show that Kairos reduces end-to-end latency by 17.8% to 28.4% compared to state-of-the-art works.
Large Language Models (LLMs) are typically fine-tuned for reasoning tasks through a two-stage pipeline of Supervised Fine-Tuning (SFT) followed by Reinforcement Learning (RL), a process fraught with catastrophic forgetting and suboptimal trade-offs between imitation and exploration. Recent single-stage methods attempt to unify SFT and RL using heuristics, but lack a principled mechanism for dynamically balancing the two paradigms. In this paper, we reframe this challenge through the theoretical lens of \textbf{implicit rewards}, viewing SFT and RL not as distinct methods but as complementary reward signals. We introduce \textbf{Adaptive Meta Fine-Tuning (AMFT)}, a novel single-stage algorithm that learns the optimal balance between SFT's implicit, path-level reward and RL's explicit, outcome-based reward. The core of AMFT is a \textbf{meta-gradient adaptive weight controller} that treats the SFT-RL balance as a learnable parameter, dynamically optimizing it to maximize long-term task performance. This forward-looking approach, regularized by policy entropy for stability, autonomously discovers an effective training curriculum. We conduct a comprehensive evaluation on challenging benchmarks spanning mathematical reasoning, abstract visual reasoning (General Points), and vision-language navigation (V-IRL). AMFT consistently establishes a new state-of-the-art and demonstrats superior generalization on out-of-distribution (OOD) tasks. Ablation studies and training dynamic analysis confirm that the meta-learning controller is crucial for AMFT's stability, sample efficiency, and performance, offering a more principled and effective paradigm for LLM alignment.Our codes are open-sourced via https://github.com/hlxtsyj/AMFT.
Interactive theorem provers (ITPs) require manual formalization, which is labor-intensive and demands expert knowledge. While automated formalization offers a potential solution, it faces two major challenges: model hallucination (e.g., undefined predicates, symbol misuse, and version incompatibility) and the semantic gap caused by ambiguous or missing premises in natural language descriptions. To address these issues, we propose CRAMF, a Concept-driven Retrieval-Augmented Mathematical Formalization framework. CRAMF enhances LLM-based autoformalization by retrieving formal definitions of core mathematical concepts, providing contextual grounding during code generation. However, applying retrieval-augmented generation (RAG) in this setting is non-trivial due to the lack of structured knowledge bases, the polymorphic nature of mathematical concepts, and the high precision required in formal retrieval. We introduce a framework for automatically constructing a concept-definition knowledge base from Mathlib4, the standard mathematical library for the Lean 4 theorem prover, indexing over 26,000 formal definitions and 1,000+ core mathematical concepts. To address conceptual polymorphism, we propose contextual query augmentation with domain- and application-level signals. In addition, we design a dual-channel hybrid retrieval strategy with reranking to ensure accurate and relevant definition retrieval. Experiments on miniF2F, ProofNet, and our newly proposed AdvancedMath benchmark show that CRAMF can be seamlessly integrated into LLM-based autoformalizers, yielding consistent improvements in translation accuracy, achieving up to 62.1% and an average of 29.9% relative improvement.
Automated translation of legacy C code into Rust aims to ensure memory safety while reducing the burden of manual migration. Early approaches in code translation rely on static rule-based methods, but they suffer from limited coverage due to dependence on predefined rule patterns. Recent works regard the task as a sequence-to-sequence problem by leveraging large language models (LLMs). Although these LLM-based methods are capable of reducing unsafe code blocks, the translated code often exhibits issues in following Rust rules and maintaining semantic consistency. On one hand, existing methods adopt a direct prompting strategy to translate the C code, which struggles to accommodate the syntactic rules between C and Rust. On the other hand, this strategy makes it difficult for LLMs to accurately capture the semantics of complex code. To address these challenges, we propose IRENE, an LLM-based framework that Integrates RulEs aNd sEmantics to enhance translation. IRENE consists of three modules: 1) a rule-augmented retrieval module that selects relevant translation examples based on rules generated from a static analyzer developed by us, thereby improving the handling of Rust rules; 2) a structured summarization module that produces a structured summary for guiding LLMs to enhance the semantic understanding of C code; 3) an error-driven translation module that leverages compiler diagnostics to iteratively refine translations. We evaluate IRENE on two datasets (xCodeEval, a public dataset, and HW-Bench, an industrial dataset provided by Huawei) and eight LLMs, focusing on translation accuracy and safety.
Recent advances in molecular science have been propelled significantly by large language models (LLMs). However, their effectiveness is limited when relying solely on molecular sequences, which fail to capture the complex structures of molecules. Beyond sequence representation, molecules exhibit two complementary structural views: the first focuses on the topological relationships between atoms, as exemplified by the graph view; and the second emphasizes the spatial configuration of molecules, as represented by the image view. The two types of views provide unique insights into molecular structures. To leverage these views collaboratively, we propose the CROss-view Prefixes (CROP) to enhance LLMs' molecular understanding through efficient multi-view integration. CROP possesses two advantages: (i) efficiency: by jointly resampling multiple structural views into fixed-length prefixes, it avoids excessive consumption of the LLM's limited context length and allows easy expansion to more views; (ii) effectiveness: by utilizing the LLM's self-encoded molecular sequences to guide the resampling process, it boosts the quality of the generated prefixes. Specifically, our framework features a carefully designed SMILES Guided Resampler for view resampling, and a Structural Embedding Gate for converting the resulting embeddings into LLM's prefixes. Extensive experiments demonstrate the superiority of CROP in tasks including molecule captioning, IUPAC name prediction and molecule property prediction.
The rapid advancement of large language models (LLMs) has resulted in increasingly sophisticated AI-generated content, posing significant challenges in distinguishing LLM-generated text from human-written language. Existing detection methods, primarily based on lexical heuristics or fine-tuned classifiers, often suffer from limited generalizability and are vulnerable to paraphrasing, adversarial perturbations, and cross-domain shifts. In this work, we propose SentiDetect, a model-agnostic framework for detecting LLM-generated text by analyzing the divergence in sentiment distribution stability. Our method is motivated by the empirical observation that LLM outputs tend to exhibit emotionally consistent patterns, whereas human-written texts display greater emotional variability. To capture this phenomenon, we define two complementary metrics: sentiment distribution consistency and sentiment distribution preservation, which quantify stability under sentiment-altering and semantic-preserving transformations. We evaluate SentiDetect on five diverse datasets and a range of advanced LLMs,including Gemini-1.5-Pro, Claude-3, GPT-4-0613, and LLaMa-3.3. Experimental results demonstrate its superiority over state-of-the-art baselines, with over 16% and 11% F1 score improvements on Gemini-1.5-Pro and GPT-4-0613, respectively. Moreover, SentiDetect also shows greater robustness to paraphrasing, adversarial attacks, and text length variations, outperforming existing detectors in challenging scenarios.
Large language models (LLMs) are susceptible to generating inaccurate or false information, often referred to as "hallucinations" or "confabulations." While several technical advancements have been made to detect hallucinated content by assessing the factuality of the model's responses, there is still limited research on how to effectively communicate this information to users. To address this gap, we conducted two scenario-based experiments with a total of 208 participants to systematically compare the effects of various design strategies for communicating factuality scores by assessing participants' ratings of trust, ease in validating response accuracy, and preference. Our findings reveal that participants preferred and trusted a design in which all phrases within a response were color-coded based on factuality scores. Participants also found it easier to validate accuracy of the response in this style compared to a baseline with no style applied. Our study offers practical design guidelines for LLM application developers and designers, aimed at calibrating user trust, aligning with user preferences, and enhancing users' ability to scrutinize LLM outputs.
Large language models (LLMs) are revolutionizing education, with LLM-based agents playing a key role in simulating student behavior. A major challenge in student simulation is modeling the diverse learning patterns of students at various cognitive levels. However, current LLMs, typically trained as ``helpful assistants'', target at generating perfect responses. As a result, they struggle to simulate students with diverse cognitive abilities, as they often produce overly advanced answers, missing the natural imperfections that characterize student learning and resulting in unrealistic simulations. To address this issue, we propose a training-free framework for student simulation. We begin by constructing a cognitive prototype for each student using a knowledge graph, which captures their understanding of concepts from past learning records. This prototype is then mapped to new tasks to predict student performance. Next, we simulate student solutions based on these predictions and iteratively refine them using a beam search method to better replicate realistic mistakes. To validate our approach, we construct the \texttt{Student\_100} dataset, consisting of $100$ students working on Python programming and $5,000$ learning records. Experimental results show that our method consistently outperforms baseline models, achieving $100\%$ improvement in simulation accuracy.
As large language models (LLMs) increasingly adapt and personalize to diverse sets of users, there is an increased risk of systems appropriating sociolects, i.e., language styles or dialects that are associated with specific minoritized lived experiences (e.g., African American English, Queer slang). In this work, we examine whether sociolect usage by an LLM agent affects user reliance on its outputs and user perception (satisfaction, frustration, trust, and social presence). We designed and conducted user studies where 498 African American English (AAE) speakers and 487 Queer slang speakers performed a set of question-answering tasks with LLM-based suggestions in either standard American English (SAE) or their self-identified sociolect. Our findings showed that sociolect usage by LLMs influenced both reliance and perceptions, though in some surprising ways. Results suggest that both AAE and Queer slang speakers relied more on the SAE agent, and had more positive perceptions of the SAE agent. Yet, only Queer slang speakers felt more social presence from the Queer slang agent over the SAE one, whereas only AAE speakers preferred and trusted the SAE agent over the AAE one. These findings emphasize the need to test for behavioral outcomes rather than simply assume that personalization would lead to a better and safer reliance outcome. They also highlight the nuanced dynamics of minoritized language in machine interactions, underscoring the need for LLMs to be carefully designed to respect cultural and linguistic boundaries while fostering genuine user engagement and trust.
Large language models (LLMs), despite their ability to perform few-shot machine translation (MT), often lag behind dedicated MT systems trained on parallel corpora, which are crucial for high quality machine translation (MT). However, parallel corpora are often scarce or non-existent for low-resource languages. In this paper, we propose CycleDistill, a bootstrapping approach leveraging LLMs and few-shot translation to obtain high-quality MT systems. CycleDistill involves iteratively generating synthetic parallel corpora from monolingual corpora via zero- or few-shot MT, which is then used to fine-tune the model that was used for generating said data for MT. CycleDistill does not need parallel corpora beyond 1 to 4 few-shot examples, and in our experiments focusing on three Indian languages, by relying solely on monolingual corpora, it can achieve high-quality machine translation, improving upon a few-shot baseline model by over 20-30 chrF points on average in the first iteration. We also study the effect of leveraging softmax activations during the distillation process and observe mild improvements in translation quality.
LSDTs: LLM-Augmented Semantic Digital Twins for Adaptive Knowledge-Intensive Infrastructure Planning
Digital Twins (DTs) offer powerful tools for managing complex infrastructure systems, but their effectiveness is often limited by challenges in integrating unstructured knowledge. Recent advances in Large Language Models (LLMs) bring new potential to address this gap, with strong abilities in extracting and organizing diverse textual information. We therefore propose LSDTs (LLM-Augmented Semantic Digital Twins), a framework that helps LLMs extract planning knowledge from unstructured documents like environmental regulations and technical guidelines, and organize it into a formal ontology. This ontology forms a semantic layer that powers a digital twin-a virtual model of the physical system-allowing it to simulate realistic, regulation-aware planning scenarios. We evaluate LSDTs through a case study of offshore wind farm planning in Maryland, including its application during Hurricane Sandy. Results demonstrate that LSDTs support interpretable, regulation-aware layout optimization, enable high-fidelity simulation, and enhance adaptability in infrastructure planning. This work shows the potential of combining generative AI with digital twins to support complex, knowledge-driven planning tasks.
Collaboratively fine-tuning (FT) large language models (LLMs) over heterogeneous mobile devices fosters immense potential applications of personalized intelligence. However, such a vision faces critical system challenges. Conventional federated LLM FT approaches place prohibitive computational and memory burdens on mobile hardware, and their synchronous model aggregation protocols stall for slower devices. In this paper, we propose Fed MobiLLM, a novel design to facilitate efficient federated LLM FT across mobile devices with diverse computing/communication speeds and local model architectures. In particular, Fed MobiLLM implements a pioneering server-assisted federated side-tuning paradigm. Briefly, mobile devices perform lightweight forward propagation computations on local data using their frozen pre-scaled backbone LLMs, and then upload selected intermediate activations. The server trains a shared side-network independently, eliminating client-side backpropagation and enabling asynchronous updates. To bridge model heterogeneity across different devices, we introduce an adaptive layer-wise feature alignment method, which ensures consistent representations for collaboratively tuning a shared side network. Extensive experimental results demonstrate that Fed MobiLLM can maintain robust fine-tuning performance while achieving extremely low on-device memory, with at least 95.2% reduction in computation overhead, 93.2% reduction in communication costs and 5.1x faster convergence compared to existing methods, validating its efficacy for practical LLM adaptation over heterogeneous mobile devices.
LLM-driven chatbots like ChatGPT have created large volumes of conversational data, but little is known about how user privacy expectations are evolving with this technology. We conduct a survey experiment with 300 US ChatGPT users to understand emerging privacy norms for sharing chatbot data. Our findings reveal a stark disconnect between user concerns and behavior: 82% of respondents rated chatbot conversations as sensitive or highly sensitive - more than email or social media posts - but nearly half reported discussing health topics and over one-third discussed personal finances with ChatGPT. Participants expressed strong privacy concerns (t(299) = 8.5, p < .01) and doubted their conversations would remain private (t(299) = -6.9, p < .01). Despite this, respondents uniformly rejected sharing personal data (search history, emails, device access) for improved services, even in exchange for premium features worth $200. To identify which factors influence appropriate chatbot data sharing, we presented participants with factorial vignettes manipulating seven contextual factors. Linear mixed models revealed that only the transmission factors such as informed consent, data anonymization, or the removal of personally identifiable information, significantly affected perceptions of appropriateness and concern for data access. Surprisingly, contextual factors including the recipient of the data (hospital vs. tech company), purpose (research vs. advertising), type of content, and geographic location did not show significant effects. Our results suggest that users apply consistent baseline privacy expectations to chatbot data, prioritizing procedural safeguards over recipient trustworthiness. This has important implications for emerging agentic AI systems that assume user willingness to integrate personal data across platforms.
There is a huge gap between numerous intriguing applications fostered by on-device large language model (LLM) fine-tuning (FT) from fresh mobile data and the limited resources of a mobile device. While existing server-assisted methods (e.g., split learning or side-tuning) may enable LLM FT on the local mobile device, they suffer from heavy communication burdens of activation transmissions, and may disclose data and labels to the server. To address those issues, we develop PAE MobiLLM, a a privacy-aware and efficient LLM FT method which can be deployed on the mobile device via server-assisted additive side-tuning. To further accelerate FT convergence and improve computing efficiency, PAE MobiLLM integrates activation caching on the server side, which allows the server to reuse historical activations and saves the mobile device from repeatedly computing forward passes for the recurring data samples. Besides, to reduce communication cost, PAE MobiLLM develops an activation shortcut that transmits only the token involved in the loss calculation instead of full activation matrices to guide the side network tuning. Last but not least, PAE MobiLLM introduces the additive adapter side-network design which makes the server train the adapter modules based on device-defined prediction differences rather than raw ground-truth labels. In this way, the server can only assist device-defined side-network computing, and learn nothing about data and labels. Extensive experimental results demonstrate PAE MobiLLM's superiority.
Large Language Models (LLMs) have shown promise in automating code generation and software engineering tasks, yet they often struggle with complex, multi-file projects due to context limitations and knowledge gaps. We propose a novel context engineering workflow that combines multiple AI components: an Intent Translator (GPT-5) for clarifying user requirements, an Elicit-powered semantic literature retrieval for injecting domain knowledge, NotebookLM-based document synthesis for contextual understanding, and a Claude Code multi-agent system for code generation and validation. Our integrated approach leverages intent clarification, retrieval-augmented generation, and specialized sub-agents orchestrated via Claude's agent framework. We demonstrate that this method significantly improves the accuracy and reliability of code assistants in real-world repositories, yielding higher single-shot success rates and better adherence to project context than baseline single-agent approaches. Qualitative results on a large Next.js codebase show the multi-agent system effectively plans, edits, and tests complex features with minimal human intervention. We compare our system with recent frameworks like CodePlan, MASAI, and HyperAgent, highlighting how targeted context injection and agent role decomposition lead to state-of-the-art performance. Finally, we discuss the implications for deploying LLM-based coding assistants in production, along with lessons learned on context management and future research directions.
Recent studies have discovered that large language models (LLM) may be ``fooled'' to output private information, including training data, system prompts, and personally identifiable information, under carefully crafted adversarial prompts. Existing red-teaming approaches for privacy leakage either rely on manual efforts or focus solely on system prompt extraction, making them ineffective for severe risks of training data leakage. We propose LeakAgent, a novel black-box red-teaming framework for LLM privacy leakage. Our framework trains an open-source LLM through reinforcement learning as the attack agent to generate adversarial prompts for both training data extraction and system prompt extraction. To achieve this, we propose a novel reward function to provide effective and fine-grained rewards and design novel mechanisms to balance exploration and exploitation during learning and enhance the diversity of adversarial prompts. Through extensive evaluations, we first show that LeakAgent significantly outperforms existing rule-based approaches in training data extraction and automated methods in system prompt leakage. We also demonstrate the effectiveness of LeakAgent in extracting system prompts from real-world applications in OpenAI's GPT Store. We further demonstrate LeakAgent's effectiveness in evading the existing guardrail defense and its helpfulness in enabling better safety alignment. Finally, we validate our customized designs through a detailed ablation study. We release our code here https://github.com/rucnyz/LeakAgent.
We present Big5-Scaler, a prompt-based framework for conditioning large language models (LLMs) with controllable Big Five personality traits. By embedding numeric trait values into natural language prompts, our method enables fine-grained personality control without additional training. We evaluate Big5-Scaler across trait expression, dialogue generation, and human trait imitation tasks. Results show that it induces consistent and distinguishable personality traits across models, with performance varying by prompt type and scale. Our analysis highlights the effectiveness of concise prompts and lower trait intensities, providing a efficient approach for building personality-aware dialogue agents.
We study the problem of serving LLM (Large Language Model) requests where each request has heterogeneous prefill and decode lengths. In LLM serving, the prefill length corresponds to the input prompt length, which determines the initial memory usage in the KV cache. The decode length refers to the number of output tokens generated sequentially, with each additional token increasing the KV cache memory usage by one unit. Given a set of n requests, our goal is to schedule and process them to minimize the total completion time. We show that this problem is NP-hard due to the interplay of batching, placement constraints, precedence relationships, and linearly increasing memory usage. We then analyze commonly used scheduling strategies in practice, such as First-Come-First-Serve (FCFS) and Shortest-First (SF), and prove that their competitive ratios scale up sublinearly with the memory limit-a significant drawback in real-world settings where memory demand is large. To address this, we propose a novel algorithm based on a new selection metric that efficiently forms batches over time. We prove that this algorithm achieves a constant competitive ratio. Finally, we develop and evaluate a few algorithm variants inspired by this approach, including dynamic programming variants, local search methods, and an LP-based scheduler, demonstrating through comprehensive simulations that they outperform standard baselines while maintaining computational efficiency.
Evaluating the capabilities and risks of foundation models is paramount, yet current methods demand extensive domain expertise, hindering their scalability as these models rapidly evolve. We introduce SKATE: a novel evaluation framework in which large language models (LLMs) compete by generating and solving verifiable tasks for one another. Our core insight is to treat evaluation as a game: models act as both task-setters and solvers, incentivized to create questions which highlight their own strengths while exposing others' weaknesses. SKATE offers several key advantages, balancing scalability, open-endedness, and objectivity. It is fully automated, data-free, and scalable, requiring no human input or domain expertise. By using verifiable tasks rather than LLM judges, scoring is objective. Unlike domain-limited programmatically-generated benchmarks (e.g. chess-playing or spatial reasoning), having LLMs creatively pose challenges enables open-ended and scalable evaluation. As a proof of concept, we introduce LLM-set code-output-prediction (COP) challenges as a verifiable and extensible framework in which to test our approach. Using a TrueSkill-based ranking system, we evaluate six frontier LLMs and find that: (1) weaker models can reliably differentiate and score stronger ones, (2) LLM-based systems are capable of self-preferencing behavior, generating questions that align with their own capabilities, and (3) SKATE automatically surfaces fine-grained capability differences between models. Our findings are an important step towards general, scalable evaluation frameworks which can keep pace with LLM progress.
Constructed languages (conlangs) such as Esperanto and Quenya have played diverse roles in art, philosophy, and international communication. Meanwhile, large-scale foundation models have revolutionized creative generation in text, images, and beyond. In this work, we leverage modern LLMs as computational creativity aids for end-to-end conlang creation. We introduce ConlangCrafter, a multi-hop pipeline that decomposes language design into modular stages -- phonology, morphology, syntax, lexicon generation, and translation. At each stage, our method leverages LLMs' meta-linguistic reasoning capabilities, injecting randomness to encourage diversity and leveraging self-refinement feedback to encourage consistency in the emerging language description. We evaluate ConlangCrafter on metrics measuring coherence and typological diversity, demonstrating its ability to produce coherent and varied conlangs without human linguistic expertise.
Time series forecasting aims to model temporal dependencies among variables for future state inference, holding significant importance and widespread applications in real-world scenarios. Although deep learning-based methods have achieved remarkable progress, they still exhibit suboptimal performance in long-term forecasting and data-scarce scenarios. Recent research demonstrates that large language models (LLMs) achieve promising performance in time series forecasting. However, we find existing LLM-based methods still have shortcomings: (1) the absence of a unified paradigm for textual prompt formulation and (2) the neglect of modality discrepancies between textual prompts and time series. To address this, we propose LLM-Prompt, an LLM-based time series forecasting framework integrating multi-prompt information and cross-modal semantic alignment. Specifically, we first construct a unified textual prompt paradigm containing learnable soft prompts and textualized hard prompts. Second, to enhance LLMs' comprehensive understanding of the forecasting task, we design a semantic space embedding and cross-modal alignment module to achieve cross-modal fusion of temporal and textual information. Finally, the transformed time series from the LLMs are projected to obtain the forecasts. Comprehensive evaluations on 6 public datasets and 3 carbon emission datasets demonstrate that LLM-Prompt is a powerful framework for time series forecasting.
Inductive program synthesis, or programming by example, requires synthesizing functions from input-output examples that generalize to unseen inputs. While large language model agents have shown promise in programming tasks guided by natural language, their ability to perform inductive program synthesis is underexplored. Existing evaluation protocols rely on static sets of examples and held-out tests, offering no feedback when synthesized functions are incorrect and failing to reflect real-world scenarios such as reverse engineering. We propose CodeARC, the Code Abstraction and Reasoning Challenge, a new evaluation framework where agents interact with a hidden target function by querying it with new inputs, synthesizing candidate functions, and iteratively refining their solutions using a differential testing oracle. This interactive setting encourages agents to perform function calls and self-correction based on feedback. We construct the first large-scale benchmark for general-purpose inductive program synthesis, featuring 1114 functions. Among 18 models evaluated, o3-mini performs best with a success rate of 52.7%, highlighting the difficulty of this task. Fine-tuning LLaMA-3.1-8B-Instruct on curated synthesis traces yields up to a 31% relative performance gain. CodeARC provides a more realistic and challenging testbed for evaluating LLM-based program synthesis and inductive reasoning. Our code, data, and models are publicly available at https://github.com/Anjiang-Wei/CodeARC
Large language models (LLMs) have revolutionized algorithm development, yet their application in symbolic regression, where algorithms automatically discover symbolic expressions from data, remains constrained and is typically designed manually by human experts. In this paper, we propose a meta learning framework that enables LLMs to automatically design selection operators for evolutionary symbolic regression algorithms. We first identify two key limitations in existing LLM-based algorithm evolution techniques: a lack of semantic guidance and code bloat. The absence of semantic awareness can lead to ineffective exchange of useful code components, and bloat results in unnecessarily complex components, both of which can reduce the interpretability of the designed algorithm or hinder evolutionary learning progress. To address these issues, we enhance the LLM-based evolution framework for meta symbolic regression with two key innovations: a complementary, semantics-aware selection operator and bloat control. Additionally, we embed domain knowledge into the prompt, enabling the LLM to generate more effective and contextually relevant selection operators. Our experimental results on symbolic regression benchmarks show that LLMs can devise selection operators that outperform nine expert-designed baselines, achieving state-of-the-art performance. Moreover, the evolved operator can further improve the state-of-the-art symbolic regression algorithm, achieving the best performance among 26 symbolic regression and machine learning algorithms across 116 regression datasets. This demonstrates that LLMs can exceed expert-level algorithm design for symbolic regression.
Large Language Models (LLMs) are increasingly expected to handle complex decision-making tasks, yet their ability to perform structured resource allocation remains underexplored. Evaluating their reasoning is also difficult due to data contamination and the static nature of existing benchmarks. We present a dual-purpose framework leveraging Participatory Budgeting (PB) both as (i) a practical setting for LLM-based resource allocation and (ii) an adaptive benchmark for evaluating their reasoning capabilities. We task LLMs with selecting project subsets under feasibility (e.g., budget) constraints via three prompting strategies: greedy selection, direct optimization, and a hill-climbing-inspired refinement. We benchmark LLMs' allocations against a utility-maximizing oracle. Interestingly, we also test whether LLMs can infer structured preferences from natural-language voter input or metadata, without explicit votes. By comparing allocations based on inferred preferences to those from ground-truth votes, we evaluate LLMs' ability to extract preferences from open-ended input. Our results underscore the role of prompt design and show that LLMs hold promise for mechanism design with unstructured inputs.
The "LLM-as-an-annotator" and "LLM-as-a-judge" paradigms employ Large Language Models (LLMs) as annotators, judges, and evaluators in tasks traditionally performed by humans. LLM annotations are widely used, not only in NLP research but also in fields like medicine, psychology, and social science. Despite their role in shaping study results and insights, there is no standard or rigorous procedure to determine whether LLMs can replace human annotators. In this paper, we propose a novel statistical procedure, the Alternative Annotator Test (alt-test), that requires only a modest subset of annotated examples to justify using LLM annotations. Additionally, we introduce a versatile and interpretable measure for comparing LLM annotators and judges. To demonstrate our procedure, we curated a diverse collection of ten datasets, consisting of language and vision-language tasks, and conducted experiments with six LLMs and four prompting techniques. Our results show that LLMs can sometimes replace humans with closed-source LLMs (such as GPT-4o), outperforming the open-source LLMs we examine, and that prompting techniques yield judges of varying quality. We hope this study encourages more rigorous and reliable practices.
Modern SoC datapaths include deeply pipelined, domain-specific accelerators, but their RTL implementation and verification are still mostly done by hand. While large language models (LLMs) exhibit advanced code-generation abilities for programming languages like Python, their application to Verilog-like RTL remains in its nascent stage. This is reflected in the simple arithmetic and control circuits currently used to evaluate generative capabilities in existing benchmarks. In this paper, we introduce ArchXBench, a six-level benchmark suite that encompasses complex arithmetic circuits and other advanced digital subsystems drawn from domains such as cryptography, image processing, machine learning, and signal processing. Architecturally, some of these designs are purely combinational, others are multi-cycle or pipelined, and many require hierarchical composition of modules. For each benchmark, we provide a problem description, design specification, and testbench, enabling rapid research in the area of LLM-driven agentic approaches for complex digital systems design. Using zero-shot prompting with Claude Sonnet 4, GPT 4.1, o4-mini-high, and DeepSeek R1 under a pass@5 criterion, we observed that o4-mini-high successfully solves the largest number of benchmarks, 16 out of 30, spanning Levels 1, 2, and 3. From Level 4 onward, however, all models consistently fail, highlighting a clear gap in the capabilities of current state-of-the-art LLMs and prompting/agentic approaches.
How do the latent spaces used by independently-trained LLMs relate to one another? We study the nearest neighbor relationships induced by activations at different layers of 24 open-weight LLMs, and find that they 1) tend to vary from layer to layer within a model, and 2) are approximately shared between corresponding layers of different models. Claim 2 shows that these nearest neighbor relationships are not arbitrary, as they are shared across models, but Claim 1 shows that they are not "obvious" either, as there is no single set of nearest neighbor relationships that is universally shared. Together, these suggest that LLMs generate a progression of activation geometries from layer to layer, but that this entire progression is largely shared between models, stretched and squeezed to fit into different architectures.
We present GhostShell, a novel approach that leverages Large Language Models (LLMs) to enable streaming and concurrent behavioral programming for embodied systems. In contrast to conventional methods that rely on pre-scheduled action sequences or behavior trees, GhostShell drives embodied systems to act on-the-fly by issuing function calls incrementally as tokens are streamed from the LLM. GhostShell features a streaming XML function token parser, a dynamic function interface mapper, and a multi-channel scheduler that orchestrates intra-channel synchronous and inter-channel asynchronous function calls, thereby coordinating serial-parallel embodied actions across multiple robotic components under LLM guidance. We evaluate GhostShell on our robotic prototype COCO through comprehensive grounded experiments across 34 real-world interaction tasks and multiple LLM backends. The results demonstrate that our approach achieves a state-of-the-art Behavioral Correctness Metric of 0.85 with Claude-4-Sonnet, and up to 66X faster response times compared to native LLM function calling APIs. GhostShell also proves effective in long-horizon multimodal tasks, exhibiting strong robustness and generalization capabilities.
Author-level citation metrics provide a practical, interpretable, and scalable signal of scholarly influence in a complex research ecosystem. It has been widely used as a proxy in hiring decisions. However, the past five years have seen the rapid emergence of large-scale publications in the field of large language models and foundation models, with papers featuring hundreds to thousands of co-authors and receiving tens of thousands of citations within months. For example, Gemini has 1361 authors and has been cited around 4600 times in 19 months. In such cases, traditional metrics, such as total citation count and the $h$-index, fail to meaningfully distinguish individual contributions. Therefore, we propose the following research question: How can one identify standout researchers among thousands of co-authors in large-scale LLM papers? This question is particularly important in scenarios such as academic hiring and funding decisions. In this paper, we introduce a novel citation metric designed to address this challenge by balancing contributions across large-scale and small-scale publications. We propose the SBCI index, analyze its theoretical properties, and evaluate its behavior on synthetic publication datasets. Our results demonstrate that the proposed metric provides a more robust and discriminative assessment of individual scholarly impact in the era of large-scale collaborations.
Operational skill learning, inherently physical and reliant on hands-on practice and kinesthetic feedback, has yet to be effectively replicated in large language model (LLM)-supported training. Current LLM training assistants primarily generate customized textual feedback, neglecting the crucial kinesthetic modality. This gap derives from the textual and uncertain nature of LLMs, compounded by concerns on user acceptance of LLM driven body control. To bridge this gap and realize the potential of collaborative human-LLM action, this work explores human experience of LLM driven kinesthetic assistance. Specifically, we introduced an "Align-Analyze-Adjust" strategy and developed FlightAxis, a tool that integrates LLM with Electrical Muscle Stimulation (EMS) for flight skill acquisition, a representative operational skill domain. FlightAxis learns flight skills from manuals and guides forearm movements during simulated flight tasks. Our results demonstrate high user acceptance of LLM-mediated body control and significantly reduced task completion times. Crucially, trainees reported that this kinesthetic assistance enhanced their awareness of operation flaws and fostered increased engagement in the training process, rather than relieving perceived load. This work demonstrated the potential of kinesthetic LLM training in operational skill acquisition.
Large language models (LLMs) have demonstrated remarkable capabilities in code generation and structured reasoning; however, their performance often degrades on complex tasks that require consistent multi-step planning. Recent work has explored combining LLMs with Monte Carlo Tree Search (MCTS), yet existing approaches primarily focus on generating heuristic-based code for optimization or target simpler tasks where correctness alone is sufficient. In this work, we propose MCTS-OPS, a novel neural-symbolic framework that formulates prompt selection as a sequential decision process guided by MCTS. Our method explores and refines multi-step prompt sequences for the goal of improving code generation quality and enhancing the problem-solving capabilities of LLMs in general optimization. Experiments on network optimization show significant improvement over the baselines, both in the success rate of executing the generated code and in the optimization results with the specified objective and constraints (2$\sim$4$\times$ higher reward and 3$\times$ lower standard deviation). Moreover, it improves the chance of attaining the optimal solution by about 10\% of cases, compared to baseline methods in hard problems. These results highlight the promise of combining symbolic planning with LLMs for robust, high-quality code generation in complex domains.
AMR (Abstract Meaning Representation) is a semantic formalism that encodes sentence meaning as rooted, directed, acyclic graphs, where nodes represent concepts and edges denote semantic relations. Finetuning decoder only Large Language Models (LLMs) represent a promising novel straightfoward direction for AMR parsing. This paper presents a comprehensive evaluation of finetuning four distinct LLM architectures, Phi 3.5, Gemma 2, LLaMA 3.2, and DeepSeek R1 LLaMA Distilled using the LDC2020T02 Gold AMR3.0 test set. Our results have shown that straightfoward finetuning of decoder only LLMs can achieve comparable performance to complex State of the Art (SOTA) AMR parsers. Notably, LLaMA 3.2 demonstrates competitive performance against SOTA AMR parsers given a straightforward finetuning approach. We achieved SMATCH F1: 0.804 on the full LDC2020T02 test split, on par with APT + Silver (IBM) at 0.804 and approaching Graphene Smatch (MBSE) at 0.854. Across our analysis, we also observed a consistent pattern where LLaMA 3.2 leads in semantic performance while Phi 3.5 excels in structural validity.
There is growing interest in integrating high-fidelity visual synthesis capabilities into large language models (LLMs) without compromising their strong reasoning capabilities. Existing methods that directly train LLMs or bridge LLMs and diffusion models usually suffer from costly training since the backbone LLMs have not seen image representations during pretraining. We present Bifrost-1, a unified framework that bridges pretrained multimodal LLMs (MLLMs) and diffusion models using patch-level CLIP image embeddings as latent variables, which are natively aligned with the MLLM's CLIP visual encoder. These patch-level image embeddings are integrated into the diffusion model with a lightweight adaptation of its ControlNet. To retain the original multimodal reasoning capabilities of MLLMs, we equip the MLLM with a visual generation branch initialized from the original MLLM parameters when predicting the patch-level image embeddings. By seamlessly integrating pretrained MLLMs and diffusion models with patch-level CLIP latents, our framework enables high-fidelity controllable image generation with significant training efficiency. Our experiments demonstrate that Bifrost-1 achieves comparable or better performance than previous methods in terms of visual fidelity and multimodal understanding, with substantially lower compute during training. We also provide comprehensive ablation studies showing the effectiveness of our design choices.
Using LLMs to give educational feedback to students for their assignments has attracted much attention in the AI in Education field. Yet, there is currently no large-scale open-source dataset of student assignments that includes detailed assignment descriptions, rubrics, and student submissions across various courses. As a result, research on generalisable methodology for automatic generation of effective and responsible educational feedback remains limited. In the current study, we constructed a large-scale dataset of Synthetic Computer science Assignments for LLM-generated Educational Feedback research (SCALEFeedback). We proposed a Sophisticated Assignment Mimicry (SAM) framework to generate the synthetic dataset by one-to-one LLM-based imitation from real assignment descriptions, student submissions to produce their synthetic versions. Our open-source dataset contains 10,000 synthetic student submissions spanning 155 assignments across 59 university-level computer science courses. Our synthetic submissions achieved BERTScore F1 0.84, PCC of 0.62 for assignment marks and 0.85 for length, compared to the corresponding real-world assignment dataset, while ensuring perfect protection of student private information. All these results of our SAM framework outperformed results of a naive mimicry method baseline. The LLM-generated feedback for our synthetic assignments demonstrated the same level of effectiveness compared to that of real-world assignment dataset. Our research showed that one-to-one LLM imitation is a promising method for generating open-source synthetic educational datasets that preserve the original dataset's semantic meaning and student data distribution, while protecting student privacy and institutional copyright. SCALEFeedback enhances our ability to develop LLM-based generalisable methods for offering high-quality, automated educational feedback in a scalable way.
The use of LLM tutors to provide automated educational feedback to students on student assignment submissions has received much attention in the AI in Education field. However, the stochastic nature and tendency for hallucinations in LLMs can undermine both quality of learning experience and adherence to ethical standards. To address this concern, we propose a method that uses LLM feedback evaluators (DeanLLMs) to automatically and comprehensively evaluate feedback generated by LLM tutor for submissions on university assignments before it is delivered to students. This allows low-quality feedback to be rejected and enables LLM tutors to improve the feedback they generated based on the evaluation results. We first proposed a comprehensive evaluation framework for LLM-generated educational feedback, comprising six dimensions for feedback content, seven for feedback effectiveness, and three for hallucination types. Next, we generated a virtual assignment submission dataset covering 85 university assignments from 43 computer science courses using eight commonly used commercial LLMs. We labelled and open-sourced the assignment dataset to support the fine-tuning and evaluation of LLM feedback evaluators. Our findings show that o3-pro demonstrated the best performance in zero-shot labelling of feedback while o4-mini demonstrated the best performance in few-shot labelling of feedback. Moreover, GPT-4.1 achieved human expert level performance after fine-tuning (Accuracy 79.8%, F1-score 79.4%; human average Accuracy 78.3%, F1-score 82.6%). Finally, we used our best-performance model to evaluate 2,000 assignment feedback instances generated by 10 common commercial LLMs, 200 each, to compare the quality of feedback generated by different LLMs. Our LLM feedback evaluator method advances our ability to automatically provide high-quality and reliable educational feedback to students.
The advent of ultra-low-bit LLM models (1/1.58/2-bit), which match the perplexity and end-task performance of their full-precision counterparts using the same model size, is ushering in a new era of LLM inference for resource-constrained environments such as edge devices and AI PCs. While these quantization advances promise models that are more cost-effective in terms of latency, memory, throughput, and energy consumption, the computational efficiency of state-of-the-art (SOTA) inference runtimes (e.g., bitnet.cpp) used to deploy them remains underexplored. In this work, we take a bottom-up approach: we first design and implement 1-bit and 2-bit microkernels optimized for modern CPUs, achieving peak computational efficiency across a variety of CPU platforms. We integrate these microkernels into a state-of-the-art LLM inference framework, namely PyTorch-TPP, and present end-to-end inference results with 2-bit models that outperform the current SOTA runtime bitnet.cpp by up to 2.2x, and deliver up to 7x speedup compared to the 16-bit model inference. Our optimized runtime advances the state of LLM inference on AI PCs and edge devices, paving the way for efficient deployment of ultra-low-bit LLM models.
Sellers at eBay are recommended keyphrases to bid on to enhance the performance of their advertising campaigns. The relevance of these keyphrases is crucial in avoiding the overcrowding of search systems with irrelevant items and maintaining a positive seller perception. It is essential that keyphrase recommendations align with both seller and Search judgments regarding auctions. Due to the difficulty in procuring negative human judgment at scale, employing LLM-as-a-judge to mimic seller judgment has been established as the norm in several studies. This study introduces a novel two-step LLM distillation process from a LLM-judge used to debias our Embedding Based Retrieval (EBR) model from the various biases that exist in click-data. We distill from an LLM teacher via a cross-encoder assistant into a bi-encoder student using a multi-task training approach, ultimately employing the student bi-encoder to retrieve relevant advertiser keyphrases. We show that integrating a knowledge distillation process from LLMs in a multi-task training setup enhances bi-encoder performance in retrieving relevant advertiser keyphrases at eBay.
Large language models (LLMs) can serve as judges that offer rapid and reliable assessments of other LLM outputs. However, models may systematically assign overly favorable ratings to their own outputs, a phenomenon known as self-bias, which can distort evaluations of true model performance. Previous studies often conflate genuine differences in model quality with bias or incorrectly assume that evaluations from LLMs and humans follow the same rating distributions. In this work, we present a statistical framework that explicitly formalizes assumptions under which self-bias can be identified and estimated. Our method models the difference in the scoring distribution that LLM-as-a-judge assigns to its own completions compared to other models, while accounting for the underlying quality of the completions provided by an independent, third-party judge (e.g., humans). Our method reliably isolates and quantifies self-bias, even when models vary in ability, ensuring that genuine performance differences are not mistaken for self-bias. We conduct an empirical analysis of self-bias on a large dataset (>5000 prompt-completion pairs) consisting of expert human annotations and judgments from nine different LLM judges. We find that some models, such as GPT-4o and Claude 3.5 Sonnet, systematically assign higher scores to their own outputs. These models also display family-bias; systematically assigning higher ratings to outputs produced by other models of the same family. Our findings highlight potential pitfalls of using LLM judges and offer practical guidance to mitigate biases when interpreting automated evaluations.
The emergent capabilities of large language models (LLMs) have prompted interest in using them as surrogates for human subjects in opinion surveys. However, prior evaluations of LLM-based opinion simulation have relied heavily on costly, domain-specific survey data, and mixed empirical results leave their reliability in question. To enable cost-effective, early-stage evaluation, we introduce a quality control assessment designed to test the viability of LLM-simulated opinions on Likert-scale tasks without requiring large-scale human data for validation. This assessment comprises two key tests: \emph{logical consistency} and \emph{alignment with stakeholder expectations}, offering a low-cost, domain-adaptable validation tool. We apply our quality control assessment to an opinion simulation task relevant to AI-assisted content moderation and fact-checking workflows -- a socially impactful use case -- and evaluate seven LLMs using a baseline prompt engineering method (backstory prompting), as well as fine-tuning and in-context learning variants. None of the models or methods pass the full assessment, revealing several failure modes. We conclude with a discussion of the risk management implications and release \texttt{TopicMisinfo}, a benchmark dataset with paired human and LLM annotations simulated by various models and approaches, to support future research.
Open-weight AI systems offer unique benefits, including enhanced transparency, open research, and decentralized access. However, they are vulnerable to tampering attacks which can efficiently elicit harmful behaviors by modifying weights or activations. Currently, there is not yet a robust science of open-weight model risk management. Existing safety fine-tuning methods and other post-training techniques have struggled to make LLMs resistant to more than a few dozen steps of adversarial fine-tuning. In this paper, we investigate whether filtering text about dual-use topics from training data can prevent unwanted capabilities and serve as a more tamper-resistant safeguard. We introduce a multi-stage pipeline for scalable data filtering and show that it offers a tractable and effective method for minimizing biothreat proxy knowledge in LLMs. We pretrain multiple 6.9B-parameter models from scratch and find that they exhibit substantial resistance to adversarial fine-tuning attacks on up to 10,000 steps and 300M tokens of biothreat-related text -- outperforming existing post-training baselines by over an order of magnitude -- with no observed degradation to unrelated capabilities. However, while filtered models lack internalized dangerous knowledge, we find that they can still leverage such information when it is provided in context (e.g., via search tool augmentation), demonstrating a need for a defense-in-depth approach. Overall, these findings help to establish pretraining data curation as a promising layer of defense for open-weight AI systems.
The conversational capabilities of large language models hold significant promise for enabling scalable and interactive tutoring. While prior research has primarily examined their capacity for Socratic questioning, it often overlooks a critical dimension: adaptively guiding learners based on their cognitive states. This study shifts focus from mere question generation to the broader instructional guidance capability. We ask: Can LLMs emulate expert tutors who dynamically adjust strategies in response to learners' understanding? To investigate this, we propose GuideEval, a benchmark grounded in authentic educational dialogues that evaluates pedagogical guidance through a three-phase behavioral framework: (1) Perception, inferring learner states; (2) Orchestration, adapting instructional strategies; and (3) Elicitation, stimulating proper reflections. Empirical findings reveal that existing LLMs frequently fail to provide effective adaptive scaffolding when learners exhibit confusion or require redirection. Furthermore, we introduce a behavior-guided finetuning strategy that leverages behavior-prompted instructional dialogues, significantly enhancing guidance performance. By shifting the focus from isolated content evaluation to learner-centered interaction, our work advocates a more dialogic paradigm for evaluating Socratic LLMs.
Large language models (LLMs) are increasingly proposed as scalable solutions to the global mental health crisis. But their deployment in psychiatric contexts raises a distinctive ethical concern: the problem of atypicality. Because LLMs generate outputs based on population-level statistical regularities, their responses -- while typically appropriate for general users -- may be dangerously inappropriate when interpreted by psychiatric patients, who often exhibit atypical cognitive or interpretive patterns. We argue that standard mitigation strategies, such as prompt engineering or fine-tuning, are insufficient to resolve this structural risk. Instead, we propose dynamic contextual certification (DCC): a staged, reversible and context-sensitive framework for deploying LLMs in psychiatry, inspired by clinical translation and dynamic safety models from artificial intelligence governance. DCC reframes chatbot deployment as an ongoing epistemic and ethical process that prioritises interpretive safety over static performance benchmarks. Atypicality, we argue, cannot be eliminated -- but it can, and must, be proactively managed.
The growing legal and ethical scrutiny of large language models (LLMs) necessitates effective machine unlearning, particularly for sensitive or unauthorized data. Existing empirical methods often yield incomplete forgetting or unintended degradation of unrelated knowledge due to poor localization. In this work, we propose GRIN: a modular and targeted framework for LLM unlearning. GRIN introduces a novel gradient-ratio-based metric to identify parameters most responsible for memorizing forget data. We then perform selective noise injection into these parameters prior to fine-tuning, which improves unlearning performance while maintaining model utility. Finally, we propose new evaluation metrics tailored to the LLM setting and validate our approach on standard benchmarks such as TOFU, WMDP, and SafePKU.
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 will be released upon acceptance.