Papers

Multimodal large language models (MLLMs) are increasingly used to automate chart generation from data tables, improving efficiency but introducing new misuse risks. We present ChartAttack, a framework for evaluating how MLLMs use design misleaders to generate charts that induce incorrect interpretations. We also introduce AttackViz, a chart question-answering (QA) dataset labeled with effective misleaders and their induced incorrect answers. ChartAttack reduces MLLM QA accuracy by 17.2 points in-domain and 11.9 points cross-domain. Conditional deception rates show targeted effects: correct answers shift to attacker-intended answers 11.2% of the time in-domain and 11.7-14.9% cross-domain, while originally incorrect answers rarely change. A controlled human study shows that ChartAttack-generated charts also reduce human QA performance. Finally, fine-tuning on AttackViz improves in-domain MLLM robustness to misleading charts. Our findings highlight the need for secure, robust MLLM chart generation. Code and data are publicly available on the project website.

The pursuit of real-time agentic interaction has driven interest in Diffusion-based Large Language Models (dLLMs) as alternatives to auto-regressive backbones, promising to break the sequential latency bottleneck. However, does such efficiency gains translate into effective agentic behavior? In this work, we present a comprehensive evaluation of dLLMs (e.g., LLaDA, Dream) across two distinct agentic paradigms: Embodied Agents (requiring long-horizon planning) and Tool-Calling Agents (requiring precise formatting). Contrary to the efficiency hype, our results on Agentboard and BFCL reveal a "bitter lesson": current dLLMs fail to serve as reliable agentic backbones, frequently leading to systematically failure. (1) In Embodied settings, dLLMs suffer repeated attempts, failing to branch under temporal feedback. (2) In Tool-Calling settings, dLLMs fail to maintain symbolic precision (e.g. strict JSON schemas) under diffusion noise. To assess the potential of dLLMs in agentic workflows, we introduce DiffuAgent, a multi-agent evaluation framework that integrates dLLMs as plug-and-play cognitive cores. Our analysis shows that dLLMs are effective in non-causal roles (e.g., memory summarization and tool selection) but require the incorporation of causal, precise, and logically grounded reasoning mechanisms into the denoising process to be viable for agentic tasks.

The transition of Large Language Models (LLMs) from passive knowledge retrievers to autonomous clinical agents demands a shift in evaluation-from static accuracy to dynamic behavioral reliability. To explore this boundary in dentistry, a domain where high-quality AI advice uniquely empowers patient-participatory decision-making, we present the Standardized Clinical Management & Performance Evaluation (SCMPE) benchmark, which comprehensively assesses performance from knowledge-oriented evaluations (static objective tasks) to workflow-based simulations (multi-turn simulated patient interactions). Our analysis reveals that while models demonstrate high proficiency in static objective tasks, their performance precipitates in dynamic clinical dialogues, identifying that the primary bottleneck lies not in knowledge retention, but in the critical challenges of active information gathering and dynamic state tracking. Mapping "Guideline Adherence" versus "Decision Quality" reveals a prevalent "High Efficacy, Low Safety" risk in general models. Furthermore, we quantify the impact of Retrieval-Augmented Generation (RAG). While RAG mitigates hallucinations in static tasks, its efficacy in dynamic workflows is limited and heterogeneous, sometimes causing degradation. This underscores that external knowledge alone cannot bridge the reasoning gap without domain-adaptive pre-training. This study empirically charts the capability boundaries of dental LLMs, providing a roadmap for bridging the gap between standardized knowledge and safe, autonomous clinical practice.

Large Audio-Language Models (LALMs) have demonstrated strong performance in spoken question answering (QA), with existing evaluations primarily focusing on answer accuracy and robustness to acoustic perturbations. However, such evaluations implicitly assume that spoken inputs remain semantically answerable, an assumption that often fails in real-world interaction when essential information is missing. In this work, we introduce a repair-aware evaluation setting that explicitly distinguishes between answerable and unanswerable audio inputs. We define answerability as a property of the input itself and construct paired evaluation conditions using a semantic-acoustic masking protocol. Based on this setting, we propose the Evaluability Awareness and Repair (EAR) score, a non-compensatory metric that jointly evaluates task competence under answerable conditions and repair behavior under unanswerable conditions. Experiments on two spoken QA benchmarks across diverse LALMs reveal a consistent gap between answer accuracy and conversational reliability: while many models perform well when inputs are answerable, most fail to recognize semantic unanswerability and initiate appropriate conversational repair. These findings expose a limitation of prevailing accuracy-centric evaluation practices and motivate reliability assessments that treat unanswerable inputs as cues for repair and continued interaction.

The Lombard effect plays a key role in natural communication, particularly in noisy environments or when addressing hearing-impaired listeners. We present a controllable text-to-speech (TTS) system capable of synthesizing Lombard-like speech in a zero-shot manner without requiring Lombard-specific training data. Our approach extends F5-TTS with a learned style embedding representation and analyzes the resulting latent space using principal component analysis (PCA) to identify directions associated with Lombard-related attributes. By manipulating these directions, we obtain interpretable control over vocal effort and articulation and generate speech at different Lombard levels. Experimental results show that the proposed method preserves speaker identity and naturalness, improves intelligibility under noisy conditions, and generalizes to previously unseen speakers. These findings demonstrate that style-embedding manipulation provides an effective and scalable framework for controllable zero-shot Lombard speech synthesis.

Integrating deep learning into healthcare enables personalized care but raises trust issues due to model opacity. To improve transparency, we propose a system for dental age estimation from panoramic images that combines an opaque and a transparent method within a natural language generation (NLG) module. This module produces clinician-friendly textual explanations about the age estimations, designed with dental experts through a rule-based approach. Following the best practices in the field, the quality of the generated explanations was manually validated by dental experts using a questionnaire. The results showed a strong performance, since the experts rated 4.77+/-0.12 (out of 5) on average across the five dimensions considered. We also performed a trustworthy self-assessment procedure following the ALTAI checklist, in which it scored 4.40+/-0.27 (out of 5) across seven dimensions of the AI Trustworthiness Assessment List.

Generative artificial intelligence (AI) is rapidly populating medical records with synthetic content, creating a feedback loop where future models are increasingly at risk of training on uncurated AI-generated data. However, the clinical consequences of this AI-generated data contamination remain unexplored. Here, we show that in the absence of mandatory human verification, this self-referential cycle drives a rapid erosion of pathological variability and diagnostic reliability. By analysing more than 800,000 synthetic data points across clinical text generation, vision-language reporting, and medical image synthesis, we find that models progressively converge toward generic phenotypes regardless of the model architecture. Specifically, rare but critical findings, including pneumothorax and effusions, vanish from the synthetic content generated by AI models, while demographic representations skew heavily toward middle-aged male phenotypes. Crucially, this degradation is masked by false diagnostic confidence; models continue to issue reassuring reports while failing to detect life-threatening pathology, with false reassurance rates tripling to 40%. Blinded physician evaluation confirms that this decoupling of confidence and accuracy renders AI-generated documentation clinically useless after just two generations. We systematically evaluate three mitigation strategies, finding that while synthetic volume scaling fails to prevent collapse, mixing real data with quality-aware filtering effectively preserves diversity. Ultimately, our results suggest that without policy-mandated human oversight, the deployment of generative AI threatens to degrade the very healthcare data ecosystems it relies upon.

Large language models (LLMs) have become powerful and widely used systems for language understanding and generation, while multi-armed bandit (MAB) algorithms provide a principled framework for adaptive decision-making under uncertainty. This survey explores the potential at the intersection of these two fields. As we know, it is the first survey to systematically review the bidirectional interaction between large language models and multi-armed bandits at the component level. We highlight the bidirectional benefits: MAB algorithms address critical LLM challenges, spanning from pre-training to retrieval-augmented generation (RAG) and personalization. Conversely, LLMs enhance MAB systems by redefining core components such as arm definition and environment modeling, thereby improving decision-making in sequential tasks. We analyze existing LLM-enhanced bandit systems and bandit-enhanced LLM systems, providing insights into their design, methodologies, and performance. Key challenges and representative findings are identified to help guide future research. An accompanying GitHub repository that indexes relevant literature is available at https://github.com/bucky1119/Awesome-LLM-Bandit-Interaction.

Recently there have been intensifying efforts to improve the understanding of Indonesian cultures by large language models (LLMs). An attractive source of cultural knowledge that has been largely overlooked is local journals of social science, which likely contain substantial cultural studies from a native perspective. We present a novel text dataset of journal article passages, created from 151 open-source Indonesian social science journals, called IndoSoSci. We demonstrate an effective recipe for injecting Indonesian cultural knowledge therein into LLMs: extracting the facts related to Indonesian culture, and apply retrieval-augmented generation (RAG) with LLM-generated hypothetical documents as queries during retrieval. The proposed recipe yields strong performance gains over several strong baselines on the IndoCulture benchmark. Additionally, by combining IndoSoSci with Indonesian Wikipedia, we set a new state-of-the-art accuracy on the IndoCulture benchmark.

Discrepancies between scientific papers and their code undermine reproducibility, a concern that grows as automated research agents scale scientific output beyond human review capacity. Whether LLMs can reliably detect such discrepancies has not been systematically measured. To this end, we present SciCoQA, a dataset of 635 paper-code discrepancies (92 real, 543 synthetic) for this cross-modal verification task. Across 22 evaluated models, even the best-performing LLMs, Gemini 3.1 Pro and GPT-5 Mini, detect only 46.7% of real-world discrepancies, revealing a critical gap in automated scientific quality assurance. We construct SciCoQA from GitHub issues and reproducibility papers, and propose a synthetic generation pipeline to scale beyond AI to Physics, Quantitative Biology, and other computational sciences. We further introduce a taxonomy of discrepancy types and categories to characterize the occurring mismatches. Our analysis shows that models particularly struggle with omitted paper details, long-context inputs, and papers outside their pre-training corpus.

Long contexts challenge transformers: attention scores dilute across thousands of tokens, critical information is often lost in the middle, and models struggle to adapt to novel patterns at inference time. Recent work on test-time adaptation addresses this by maintaining a form of working memory – transient parameters updated on the current context – but existing approaches rely on uniform write policies that waste computation on low-utility regions and suffer from high gradient variance across semantically heterogeneous contexts. In this work, we reframe test-time adaptation as a budget-constrained memory consolidation problem, focusing on which parts of the context should be consolidated into working memory under limited computation. We propose Gdwm (Gated Differentiable Working Memory), a framework that introduces a write controller to gate the consolidation process. The controller estimates Contextual Utility, an information-theoretic measure of long-range contextual dependence, and allocates gradient steps accordingly while maintaining global coverage. Experiments on ZeroSCROLLS and LongBench v2 demonstrate that Gdwm achieves comparable or superior performance with 4$\times$ fewer gradient steps than uniform baselines, establishing a new efficiency-performance Pareto frontier for test-time adaptation.

Retrieval-Augmented Generation enhances Large Language Models by integrating external knowledge, which reduces hallucinations but increases prompt length. This increase leads to higher computational costs and longer Time to First Token (TTFT). To mitigate this issue, existing solutions aim to reuse the preprocessed KV cache of each retrieved chunk to accelerate RAG. However, the lack of cross-chunk contextual information leads to a significant drop in generation quality, leaving the potential benefits of KV cache reuse largely unfulfilled. The challenge lies in how to reuse the precomputed KV cache of chunks while preserving generation quality. We propose FusionRAG, a novel inference framework that optimizes both the preprocessing and reprocessing stages of RAG. In the offline preprocessing stage, we embed information from other related text chunks into each chunk, while in the online reprocessing stage, we recompute the KV cache for tokens that the model focuses on. As a result, we achieve a better trade-off between generation quality and efficiency. According to our experiments, FusionRAG significantly improves generation quality at the same recomputation ratio compared to previous state-of-the-art solutions. By recomputing fewer than 15% of the tokens, FusionRAG achieves up to 70% higher normalized F1 scores than baselines and reduces TTFT by 2.66x-9.39x compared to Full Attention.

Extracting sparse circuits from billion-parameter transformers is constrained by $O(2^n)$ search cost and pervasive feature reuse across co-active pathways. Hierarchical Attribution Graph Decomposition (HAGD) addresses this through four stages: cross-layer transcoder training, spectral coarsening of attribution graphs, graph-neural-network (GNN)-guided hierarchical traversal, and causal intervention verification, reducing worst-case complexity to $O(n^2 \log n)$. Per-layer transcoders trained on the RedPajama corpus yield monosemantic dictionaries; gradient-activation products form weighted attribution graphs; normalized-Laplacian spectral clustering builds multi-resolution hierarchies; an attention-based GNN assigns circuit-membership scores at successive coarsening stages. Evaluation spans GPT-2 (117M-774M), Pythia (1.4B-6.9B), and Llama (7B-70B) across modular arithmetic, parity computation, integer sorting, coreference resolution (WinoGrande), commonsense reasoning (HellaSwag), and factual recall. Behavioral preservation reaches 91% ($\pm$2.3%) on modular arithmetic with 49-347-node circuits, while ACDC exhausts memory beyond 1.4B parameters. Cross-architecture transfer coefficients span 0.38-0.82, with within-family pairs (Llama-7B $\to$ Llama-70B) attaining 0.82. Limitations include omitted attention-head circuits, 15-20% unexplained reconstruction variance, ablation-based validation circularity, and uncertain interpretability of circuits exceeding several hundred nodes.

Large language models (LLMs) increasingly operate in high-stakes settings where demographic attributes such as race and ethnicity may be explicitly stated or implicitly suggested through textual cues. However, existing studies primarily document outcome-level disparities, offering limited insight into internal mechanisms underlying these effects. We present a mechanistic study of how race and ethnicity cues are represented and operationalized within LLMs. Using two publicly available datasets spanning toxicity-related generation and clinical narrative understanding tasks, we analyze three open-source models with a reproducible interpretability pipeline combining probing, neuron-level attribution, and targeted intervention. We find that sensitivity to demographic cues is distributed across internal units and varies substantially across models. These units often align with entangled semantic facets, including explicit group labels, geography, language, culture, and associations related to stereotypes. Interventions on selected units can change some biased prediction patterns, but substantial residual effects remain, suggesting that effective mitigation requires understanding distributed, task-specific mechanisms rather than manipulating a small set of identified neurons alone. Code: https://github.com/LARK-NLP-Lab/LLM-Bias-Interpretability.

The TRAIMA project (TRaitement Automatique des Interactions Multimodales en Apprentissage), conducted between March 2019 and June 2020, investigates the potential of automatic processing of multimodal interactions in educational settings. The project addresses a central methodological challenge in educational and interactional research: the analysis of verbal, paraverbal, and non-verbal data is currently carried out manually, making it extremely time-consuming and difficult to scale. TRAIMA explores how machine learning approaches could contribute to the categorisation and classification of such interactions. The project focuses specifically on explanatory and collaborative sequences occurring in classroom interactions, particularly in French as a Foreign Language (FLE) and French as a First Language (FLM) contexts. These sequences are analysed as inherently multimodal phenomena, combining spoken language with prosody, gestures, posture, gaze, and spatial positioning. A key theoretical contribution of the project is the precise linguistic and interactional definition of explanatory discourse as a tripartite sequence (opening, explanatory core, closure), drawing on discourse analysis and interactional linguistics. A substantial part of the research is devoted to the methodological foundations of transcription, which constitute a critical bottleneck for any form of automation. The report provides a detailed state of the art of existing transcription conventions (ICOR, Mondada, GARS, VALIBEL, Ferré), highlighting their respective strengths and limitations when applied to multimodal classroom data. Through comparative analyses of manually transcribed sequences, the project demonstrates the inevitable variability and interpretative dimension of transcription practices, depending on theoretical positioning and analytical goals. Empirical work is based on several corpora, notably the INTER-EXPLIC corpus (approximately 30 hours of classroom interaction) and the EXPLIC-LEXIC corpus, which serve both as testing grounds for manual annotation and as reference datasets for future automation. Particular attention is paid to teacher gestures (kinésic and proxemic resources), prosodic features, and their functional role in meaning construction and learner comprehension. The project also highlights the strategic role of the TechnéLAB platform, which provides advanced multimodal data capture (multi-camera video, synchronized audio, eye-tracking, digital interaction traces) and constitutes both a research infrastructure and a test environment for the development of automated tools. In conclusion, TRAIMA does not aim to deliver a fully operational automated system, but rather to establish a rigorous methodological framework for the automatic processing of multimodal pedagogical interactions. The project identifies transcription conventions, annotation categories, and analytical units that are compatible with machine learning approaches, while emphasizing the need for theoretical explicitness and researcher reflexivity. TRAIMA thus lays the groundwork for future interdisciplinary research at the intersection of didactics, discourse analysis, multimodality, and artificial intelligence in education.

Legal Judgment Prediction (LJP) aims to predict the outcomes of legal cases based on factual descriptions, serving as a fundamental task to advance the development of legal systems. Traditional methods often rely on statistical analyses or role-based simulations but face challenges with multiple allegations, diverse evidence, and lack adaptability. In this paper, we introduce JurisMMA, a novel framework for LJP that effectively decomposes trial tasks, standardizes processes, and organizes them into distinct stages. Furthermore, we build JurisMM, a large dataset with over 100,000 recent Chinese judicial records, including both text and multimodal video-text data, enabling comprehensive evaluation. Experiments on JurisMM and the benchmark LawBench validate our framework's effectiveness. These results indicate that our framework is effective not only for LJP but also for a broader range of legal applications, offering new perspectives for the development of future legal methods and datasets.

Clinical Question-Answering (CQA) industry systems are increasingly rely on Large Language Models (LLMs), yet their deployment is often guided by the assumption that domain-specific fine-tuning is essential. Although specialised medical LLMs such as BioBERT, BioGPT, and PubMedBERT remain popular, they face practical limitations including narrow coverage, high retraining costs, and limited adaptability. Efforts based on Supervised Fine-Tuning (SFT) have attempted to address these assumptions but continue to reinforce what we term the SPECIALISATION FALLACY-the belief that specialised medical LLMs are inherently superior for CQA. To address this assumption, we introduce MEDASSESS-X, a deployment-industry-oriented CQA framework that applies alignment at inference time rather than through SFT. MEDASSESS-X uses lightweight steering vectors to guide model activations toward medically consistent reasoning without updating model weights or requiring domain-specific retraining. This inference-time alignment layer stabilises CQA performance across both general-purpose and specialised medical LLMs, thereby resolving the SPECIALISATION FALLACY. Empirically, MEDASSESS-X delivers consistent gains across all LLM families, improving Accuracy by up to +6%, Factual Consistency by +7%, and reducing Safety Error Rate by as much as 50%.

Large language models (LLMs) have shown growing promise in biomedical research, particularly for knowledge-driven interpretation tasks. However, their ability to reliably reason from gene-level knowledge to functional understanding, a core requirement for knowledge-enhanced cell atlas interpretation, remains largely underexplored. To address this gap, we introduce SciHorizon-GENE, a large-scale gene-centric benchmark constructed from authoritative biological databases. The benchmark integrates curated knowledge for over 190K human genes and comprises more than 540K questions covering diverse gene-to-function reasoning scenarios relevant to cell type annotation, functional interpretation, and mechanism-oriented analysis. Motivated by behavioral patterns observed in preliminary examinations, SciHorizon-GENE evaluates LLMs along four biologically critical perspectives: research attention sensitivity, hallucination tendency, answer completeness, and literature influence, explicitly targeting failure modes that limit the safe adoption of LLMs in biological interpretation pipelines. We systematically evaluate a wide range of state-of-the-art general-purpose and biomedical LLMs, revealing substantial heterogeneity in gene-level reasoning capabilities and persistent challenges in generating faithful, complete, and literature-grounded functional interpretations. Our benchmark establishes a systematic foundation for analyzing LLM behavior at the gene scale and offers insights for model selection and development, with direct relevance to knowledge-enhanced biological interpretation.

Humanoid robots are capable of performing various actions such as greeting, dancing and even backflipping. However, these motions are often hard-coded or specifically trained, which limits their versatility. In this work, we present FRoM-W1, an open-source framework designed to achieve general humanoid whole-body motion control using natural language. To universally understand natural language and generate corresponding motions, as well as enable various humanoid robots to stably execute these motions in the physical world under gravity, FRoM-W1 operates in two stages: (a) H-GPT: utilizing massive human data, a large-scale language-driven human whole-body motion generation model is trained to generate diverse natural behaviors. We further leverage the Chain-of-Thought technique to improve the model's generalization in instruction understanding. (b) H-ACT: After retargeting generated human whole-body motions into robot-specific actions, a motion controller that is pretrained and further fine-tuned through reinforcement learning in physical simulation enables humanoid robots to accurately and stably perform corresponding actions. It is then deployed on real robots via a modular simulation-to-reality module. We extensively evaluate FRoM-W1 on Unitree H1 and G1 robots. Results demonstrate superior performance on the HumanML3D-X benchmark for human whole-body motion generation, and our introduced reinforcement learning fine-tuning consistently improves both motion tracking accuracy and task success rates of these humanoid robots. We open-source the entire FRoM-W1 framework and hope it will advance the development of humanoid intelligence.

Given an input image and set of class names, panoptic segmentation aims to label each pixel in an image with class labels and instance labels. In comparison, Open Vocabulary Panoptic Segmentation aims to facilitate the segmentation of arbitrary classes according to user input. The challenge is that a panoptic segmentation system trained on a particular dataset typically does not generalize well to unseen classes beyond the training data. In this work, we propose RetCLIP, a retrieval-augmented panoptic segmentation method that improves the performance of unseen classes. In particular, we construct a masked segment feature database using paired image-text data. At inference time, we use masked segment features from the input image as query keys to retrieve similar features and associated class labels from the database. Classification scores for the masked segment are assigned based on the similarity between query features and retrieved features. The retrieval-based classification scores are combined with CLIP-based scores to produce the final output. We incorporate our solution with a previous SOTA method (FC-CLIP). When trained on COCO, the proposed method demonstrates 30.9 PQ, 19.3 mAP, 44.0 mIoU on the ADE20k dataset, achieving +4.5 PQ, +2.5 mAP, +10.0 mIoU absolute improvement over the baseline.

Large language models (LLMs) possess extensive world knowledge, yet methods for effectively eliciting this knowledge remain underexplored. Nationality and region prediction tasks require understanding of not only linguistic features but also cultural and historical background, making LLM world knowledge particularly valuable. However, conventional LLM prompting methods rely on direct reasoning approaches, which have limitations in applying abstract linguistic rules. We propose LLM Associative Memory Agents (LAMA), a novel framework that leverages LLM world knowledge as associative memory. Rather than directly inferring nationality from names, LAMA recalls famous individuals with the same name and aggregates their nationalities through indirect reasoning. A dual-agent architecture comprising a Person Agent and a Media Agent, specialized in different knowledge domains, recalls famous individuals in parallel, generating Top-1 predictions through voting and Top-K predictions through conditional completion. On a 99-country nationality prediction task, LAMA achieved 0.817 accuracy, substantially outperforming conventional LLM prompting methods and neural models. Our experiments reveal that LLMs exhibit higher reliability in recalling concrete examples than in abstract reasoning, that recall-based approaches are robust to low-frequency nationalities independent of data frequency distributions, and that the dual-agent architecture functions complementarily to produce synergistic effects. These results demonstrate the effectiveness of a new multi-agent system that retrieves and aggregates LLM knowledge rather than prompting reasoning.

As large language models are increasingly used in high-stakes domains, it is essential that their outputs reflect not average} human preference, rather range of varying perspectives. Achieving such pluralism, however, remains challenging. Existing approaches consider limited values or rely on prompt-level interventions, lacking value control and representation. To address this, we introduce VISPA, a training-free pluralistic alignment framework, that enables direct control over value expression by dynamic selection and internal model activation steering. Across extensive empirical studies spanning multiple models and evaluation settings, we show VISPA is performant across all pluralistic alignment modes in healthcare and beyond. Further analysis reveals VISPA is adaptable with different steering initiations, model, and/or values. These results suggest that pluralistic alignment can be achieved through internal activation mechanisms, offering a scalable path toward language models that serves all.

LLMs are increasingly used for mental health support, yet they can produce responses that are overly directive, inconsistent, or clinically misaligned. Most risk mitigation approaches rely on implicit alignment through training or prompting, offering limited runtime accountability. We introduce PAIR-SAFE, a paired-agent framework that integrates a Responder agent with a Judge agent grounded in the clinically validated Motivational Interviewing Treatment Integrity (MITI-4). The Judge audits each response and issues \allow{} or \revise{} decisions to guide real-time refinement. We evaluate interactions using a support-seeker simulator derived from human-annotated data. The Judge-supervised interactions show significant improvements in key MITI dimensions, including Partnership, Seek Collaboration, and overall Relational quality. Our quantitative findings are corroborated by expert evaluation, highlighting nuances of runtime supervision. Our study reveals that paired-agent auditing can provide clinically grounded, transparent oversight for AI-assisted mental health support.

Process Reward Models (PRMs) have achieved strong results in complex reasoning, but are bottlenecked by costly process-level supervision. A widely used alternative, Monte Carlo Estimation (MCE), defines process rewards as the probability that a policy model reaches the correct final answer from a given reasoning step. However, step correctness is an intrinsic property of the reasoning trajectory, and should be invariant to policy choice. Our empirical findings show that MCE producing policy-dependent rewards that induce label noise, including false positives that reward incorrect steps and false negatives that penalize correct ones. To address above challenges, we propose a two-stage framework to mitigate noisy supervision. In the labeling stage, we introduce a reflection-aware label correction mechanism that uses a large language model (LLM) as a judge to detect reflection and self-correction behaviors related to the current reasoning step, thereby suppressing overestimated rewards. In the training stage, we further propose a \underline{N}oise-\underline{A}ware \underline{I}terative \underline{T}raining framework that enables the PRM to progressively refine noisy labels based on its own confidence. Extensive Experiments show that our method substantially improves step-level correctness discrimination, achieving up to a 27% absolute gain in average F1 over PRMs trained with noisy supervision.

Predicting problem-difficulty in large language models (LLMs) refers to estimating how difficult a task is according to the model itself, typically by training linear probes on its internal representations. In this work, we study the multilingual geometry of problem-difficulty in LLMs by training linear probes using the AMC subset of the Easy2Hard benchmark, translated into 21 languages. We found that difficulty-related signals emerge at two distinct stages of the model internals, corresponding to shallow (early-layers) and deep (later-layers) internal representations, that exhibit functionally different behaviors. Probes trained on deep representations achieve high accuracy when evaluated on the same language but exhibit poor cross-lingual generalization. In contrast, probes trained on shallow representations generalize substantially better across languages, despite achieving lower within-language performance. Together, these results suggest that LLMs first form a language-agnostic representation of problem difficulty, which subsequently becomes language-specific. This closely aligns with existing findings in LLM interpretability showing that models tend to operate in an abstract conceptual space before producing language-specific outputs. We demonstrate that this two-stage representational process extends beyond semantic content to high-level meta-cognitive properties such as problem-difficulty estimation.

GPU code optimization is a key performance bottleneck for HPC workloads as well as large-model training and inference. Although compiler optimizations and hand-written kernels can partially alleviate this issue, achieving near-hardware-limit performance still relies heavily on manual code refactoring and parameter tuning. Recent progress in LLM-agent-based kernel generation and optimization has been reported, yet many approaches primarily focus on direct code rewriting, where parameter choices are often implicit and hard to control, or require human intervention, leading to unstable performance gains. This paper introduces a template-based rewriting layer on top of an agent-driven iterative loop: kernels are semantically refactored into explicitly parameterizable templates, and template parameters are then optimized via search-based autotuning, yielding more stable and higher-quality speedups. Experiments on a set of real-world kernels demonstrate speedups exceeding 3x in the best case. We extract representative CUDA kernels from SGLang as evaluation targets; the proposed agentic tuner iteratively performs templating, testing, analysis, and planning, and leverages profiling feedback to execute constrained parameter search under hardware resource limits. Compared to agent-only direct rewriting, the template-plus-search design significantly reduces the randomness of iterative optimization, making the process more interpretable and enabling a more systematic approach toward high-performance configurations. The proposed method can be further extended to OpenCL, HIP, and other backends to deliver automated performance optimization for real production workloads.

Current guardian models are predominantly Western-centric and optimized for high-resource languages, leaving low-resource African languages vulnerable to evolving harms, cross-lingual failures, and cultural misalignment. Moreover, most guardian models rely on rigid, predefined safety categories that fail to generalize across diverse linguistic and sociocultural contexts. Achieving robust safety requires flexible, runtime-enforceable policies and benchmarks that reflect local norms, harm scenarios, and cultural expectations. We introduce UbuntuGuard, the first policy-based safety benchmark for African languages built from adversarial queries authored by 155 domain experts across sensitive fields, including healthcare. From these expert-crafted queries, we derive context-specific safety policies and reference responses that capture culturally grounded risk signals, enabling policy-aligned evaluation of guardian models. We evaluate 15 models, comprising seven general-purpose LLMs and eight guardian models across three distinct variants: static, dynamic, and multilingual. Our findings reveal that existing English-centric benchmarks overestimate real-world multilingual safety, cross-lingual transfer provides partial but insufficient coverage, and dynamic models, while better equipped to leverage policies at inference time, still struggle to fully localize African-language contexts. These findings highlight the urgent need for multilingual, culturally grounded safety benchmarks to enable the development of reliable and equitable guardian models for low-resource languages.

Retrieval-Augmented Generation (RAG) has emerged as a powerful technique for enhancing the quality of responses in Question-Answering (QA) tasks. However, existing approaches often struggle with retrieving contextually relevant information, leading to incomplete or suboptimal answers. In this paper, we introduce Structured-Semantic RAG (SSRAG), a hybrid architecture that enhances QA quality by integrating query augmentation, agentic routing, and a structured retrieval mechanism combining vector and graph based techniques with context unification. By refining retrieval processes and improving contextual grounding, our approach improves both answer accuracy and informativeness. We conduct extensive evaluations on three popular QA datasets, TruthfulQA, SQuAD and WikiQA, across five Large Language Models (LLMs), demonstrating that our proposed approach consistently improves response quality over standard RAG implementations.

Procedural case logs are a core requirement in radiology training, yet they are time-consuming to complete and prone to inconsistency when authored manually. This study investigates whether large language models (LLMs) can automate procedural case log documentation directly from free-text radiology reports. We evaluate multiple local and commercial LLMs under instruction-based and chain-of-thought prompting to extract structured procedural information from 414 curated interventional radiology reports authored by nine residents between 2018 and 2024. Model performance is assessed using sensitivity, specificity, and F1-score, alongside inference latency and token efficiency to estimate operational cost. Results show that both local and commercial models achieve strong extraction performance, with best F1-scores approaching 0.87, while exhibiting different trade-offs between speed and cost. Automation using LLMs has the potential to substantially reduce clerical burden for trainees and improve consistency in case logging. These findings demonstrate the feasibility of AI-assisted documentation in medical education and highlight the need for further validation across institutions and clinical workflows.

Fine-tuning LLMs on benign data can still degrade alignment and adversarial robustness, yet direct analysis of the role of fine-tuning objectives in shaping these safety outcomes remain limited. We present a controlled comparison of six fine-tuning objectives – Supervised Fine-Tuning, Direct Preference Optimization, Conditional Fine-Tuning, Inoculation Prompting, Odds Ratio Preference Optimization, and KL-regularized fine-tuning – holding data, domain, architecture, and optimization fixed. Across closed-form reasoning and open-ended generation tasks, we find that objective choice induces systematic, scale-dependent shifts along the safety-capability frontier. At small training budgets, robustness is similar across objectives but capability differs. At larger budgets, objectives diverge sharply: supervised and preference-based tuning tightly couple capability gains to increased adversarial vulnerability and persona drift, while objectives that constrain learning signals – especially ORPO and KL-regularization – substantially mitigate both. Fine-tuning objectives therefore matter little for safety at small scales but become a primary driver of adversarial robustness and latent persona stability as training scale increases.