Papers

The influence of gender diversity on the success of scientific teams is of great interest to academia. However, prior findings remain inconsistent, and most studies operationalize diversity in aggregate terms, overlooking internal role differentiation. This limitation obscures a more nuanced understanding of how gender diversity shapes team impact. In particular, the effect of gender diversity across different team roles remains poorly understood. To this end, we define a scientific team as all coauthors of a paper and measure team impact through five-year citation counts. Using author contribution statements, we classified members into leadership and support roles. Drawing on more than 130,000 papers from PLOS journals, most of which are in biomedical-related disciplines, we employed multivariable regression to examine the association between gender diversity in these roles and team impact. Furthermore, we apply a threshold regression model to investigate how team size moderates this relationship. The results show that (1) the relationship between gender diversity and team impact follows an inverted U-shape for both leadership and support groups; (2) teams with an all-female leadership group and an all-male support group achieve higher impact than other team types. Interestingly, (3) the effect of leadership-group gender diversity is significantly negative for small teams but becomes positive and statistically insignificant in large teams. In contrast, the estimates for support-group gender diversity remain significant and positive, regardless of team size.

As access to high-quality, domain-specific data grows increasingly scarce, multi-epoch training has become a practical strategy for adapting large language models (LLMs). However, autoregressive models often suffer from performance degradation under repeated data exposure, where overfitting leads to a marked decline in model capability. Through empirical analysis, we trace this degradation to an imbalance in learning dynamics: predictable, low-entropy tokens are learned quickly and come to dominate optimization, while the model's ability to generalize on high-entropy tokens deteriorates with continued training. To address this, we introduce EntroDrop, an entropy-guided token dropout method that functions as structured data regularization. EntroDrop selectively masks low-entropy tokens during training and employs a curriculum schedule to adjust regularization strength in alignment with training progress. Experiments across model scales from 0.6B to 8B parameters show that EntroDrop consistently outperforms standard regularization baselines and maintains robust performance throughout extended multi-epoch training. These findings underscore the importance of aligning regularization with token-level learning dynamics when training on limited data. Our approach offers a promising pathway toward more effective adaptation of LLMs in data-constrained domains.

Neural scaling laws have become foundational for optimizing large language model (LLM) training, yet they typically assume a single dense model output. This limitation effectively overlooks "Familial models, a transformative paradigm essential for realizing ubiquitous intelligence across heterogeneous device-edge-cloud hierarchies. Transcending static architectures, familial models integrate early exits with relay-style inference to spawn G deployable sub-models from a single shared backbone. In this work, we theoretically and empirically extend the scaling law to capture this "one-run, many-models" paradigm by introducing Granularity (G) as a fundamental scaling variable alongside model size (N) and training tokens (D). To rigorously quantify this relationship, we propose a unified functional form L(N, D, G) and parameterize it using large-scale empirical runs. Specifically, we employ a rigorous IsoFLOP experimental design to strictly isolate architectural impact from computational scale. Across fixed budgets, we systematically sweep model sizes (N) and granularities (G) while dynamically adjusting tokens (D). This approach effectively decouples the marginal cost of granularity from the benefits of scale, ensuring high-fidelity parameterization of our unified scaling law. Our results reveal that the granularity penalty follows a multiplicative power law with an extremely small exponent. Theoretically, this bridges fixed-compute training with dynamic architectures. Practically, it validates the "train once, deploy many" paradigm, demonstrating that deployment flexibility is achievable without compromising the compute-optimality of dense baselines.

Large Language Models (LLMs) have achieved strong performance across a wide range of natural language processing tasks in recent years, including machine translation, text generation, and question answering. As their applications extend to increasingly complex scenarios, however, LLMs continue to face challenges in tasks that require deep reasoning and logical inference. In particular, models trained on large scale textual corpora may incorporate noisy or irrelevant information during generation, which can lead to incorrect predictions or outputs that are inconsistent with factual knowledge. To address this limitation, we propose a stepwise reasoning enhancement framework for LLMs based on external subgraph generation, termed SGR. The proposed framework dynamically constructs query relevant subgraphs from external knowledge bases and leverages their semantic structure to guide the reasoning process. By performing reasoning in a step by step manner over structured subgraphs, SGR reduces the influence of noisy information and improves reasoning accuracy. Specifically, the framework first generates an external subgraph tailored to the input query, then guides the model to conduct multi step reasoning grounded in the subgraph, and finally integrates multiple reasoning paths to produce the final answer. Experimental results on multiple benchmark datasets demonstrate that SGR consistently outperforms strong baselines, indicating its effectiveness in enhancing the reasoning capabilities of LLMs.

Memory serves as the pivotal nexus bridging past and future, providing both humans and AI systems with invaluable concepts and experience to navigate complex tasks. Recent research on autonomous agents has increasingly focused on designing efficient memory workflows by drawing on cognitive neuroscience. However, constrained by interdisciplinary barriers, existing works struggle to assimilate the essence of human memory mechanisms. To bridge this gap, we systematically synthesizes interdisciplinary knowledge of memory, connecting insights from cognitive neuroscience with LLM-driven agents. Specifically, we first elucidate the definition and function of memory along a progressive trajectory from cognitive neuroscience through LLMs to agents. We then provide a comparative analysis of memory taxonomy, storage mechanisms, and the complete management lifecycle from both biological and artificial perspectives. Subsequently, we review the mainstream benchmarks for evaluating agent memory. Additionally, we explore memory security from dual perspectives of attack and defense. Finally, we envision future research directions, with a focus on multimodal memory systems and skill acquisition.

Large Language Model (LLM) agents, while proficient in the digital realm, face a significant gap in physical-world deployment due to the challenge of forming and maintaining a robust spatial mental model. We identify three core cognitive challenges hindering this transition: spatial reasoning, long-horizon state tracking via mental simulation, and active exploration under partial observation. To isolate and evaluate these faculties, we introduce CubeBench, a novel generative benchmark centered on the Rubik's Cube. CubeBench uses a three-tiered diagnostic framework that progressively assesses agent capabilities, from foundational state tracking with full symbolic information to active exploration with only partial visual data. Our experiments on leading LLMs reveal critical limitations, including a uniform 0.00% pass rate on all long-horizon tasks, exposing a fundamental failure in long-term planning. We also propose a diagnostic framework to isolate these cognitive bottlenecks by providing external solver tools. By analyzing the failure modes, we provide key insights to guide the development of more physically-grounded intelligent agents.

Audiobook interpretations are attracting increasing attention, as they provide accessible and in-depth analyses of books that offer readers practical insights and intellectual inspiration. However, their manual creation process remains time-consuming and resource-intensive. To address this challenge, we propose AI4Reading, a multi-agent collaboration system leveraging large language models (LLMs) and speech synthesis technology to generate podcast, like audiobook interpretations. The system is designed to meet three key objectives: accurate content preservation, enhanced comprehensibility, and a logical narrative structure. To achieve these goals, we develop a framework composed of 11 specialized agents,including topic analysts, case analysts, editors, a narrator, and proofreaders that work in concert to explore themes, extract real world cases, refine content organization, and synthesize natural spoken language. By comparing expert interpretations with our system's output, the results show that although AI4Reading still has a gap in speech generation quality, the generated interpretative scripts are simpler and more accurate.

E-commerce live streaming in China, particularly on platforms like Douyin, has become a major sales channel, but hosts often use morphs to evade scrutiny and engage in false advertising. This study introduces the Live Auditory Morph Resolution (LiveAMR) task to detect such violations. Unlike previous morph research focused on text-based evasion in social media and underground industries, LiveAMR targets pronunciation-based evasion in health and medical live streams. We constructed the first LiveAMR dataset with 86,790 samples and developed a method to transform the task into a text-to-text generation problem. By leveraging large language models (LLMs) to generate additional training data, we improved performance and demonstrated that morph resolution significantly enhances live streaming regulation.

Safety alignment – training large language models (LLMs) to refuse harmful requests while remaining helpful – is critical for responsible deployment. Prior work established that safety behaviors are governed by low-rank structures, suggesting parameter-efficient fine-tuning (PEFT) should be well-suited for alignment. However, Low-Rank Adaptation (LoRA) consistently underperforms full fine-tuning and reinforcement learning on safety benchmarks. We attribute this gap to semantic entanglement: safety-relevant directions are intertwined with unrelated concepts due to polysemanticity, impeding implicit subspace identification. To address this, we propose SAILS (Safety Alignment via Interpretable Low-rank Subspace), which leverages Sparse Autoencoders (SAEs) to disentangle representations into monosemantic features, constructs an interpretable safety subspace from SAE decoder directions, and uses it to initialize LoRA adapters. Theoretically, we prove that SAE-based identification achieves arbitrarily small recovery error under monosemanticity assumptions, while direct identification suffers an irreducible error floor. Empirically, SAILS achieves up to 99.6% safety rate on Gemma-2-9B – exceeding full fine-tuning by 7.4 points and matching RLHF-based models – while updating only 0.19% of parameters and providing interpretability.

Large Language Models (LLMs) have demonstrated remarkable capabilities in code generation, yet they exhibit systematic errors on complex, multi-step programming tasks. We hypothesize that these errors stem from the flexibility of general-purpose languages, which permits multiple valid approaches and requires implicit state management. To test this hypothesis, we introduce Anka, a domain-specific language (DSL) for data transformation pipelines designed with explicit, constrained syntax that reduces ambiguity in code generation. Despite having zero prior training exposure to Anka, Claude 3.5 Haiku achieves 99.9% parse success and 95.8% overall task accuracy across 100 benchmark problems. Critically, Anka demonstrates a 40 percentage point accuracy advantage over Python on multi-step pipeline tasks (100% vs. 60%), where Python's flexible syntax leads to frequent errors in operation sequencing and variable management. Cross-model validation with GPT-4o-mini confirms this advantage (+26.7 percentage points on multi-step tasks). Our results demonstrate that: (1) LLMs can learn novel DSLs entirely from in-context prompts, achieving near-native accuracy; (2) constrained syntax significantly reduces errors on complex tasks; and (3) domain-specific languages purposefully designed for LLM generation can outperform general-purpose languages on which the LLM has extensive training. We release the complete language implementation, benchmark suite, and evaluation framework to facilitate further research.

We propose LLM-PeerReview, an unsupervised LLM Ensemble method that selects the most ideal response from multiple LLM-generated candidates for each query, harnessing the collective wisdom of multiple models with diverse strengths. LLM-PeerReview is built on a novel, peer-review-inspired framework that offers a transparent and interpretable mechanism, while remaining fully unsupervised for flexible adaptability and generalization. Specifically, it operates in three stages: For scoring, we use the emerging LLM-as-a-Judge technique to evaluate each response by reusing multiple LLMs at hand; For reasoning, we can apply a straightforward averaging strategy or a principled graphical model-based truth inference algorithm to aggregate multiple scores to produce a final score for each response; Finally, the highest-scoring response is selected as the best ensemble output. LLM-PeerReview is conceptually simple and empirically powerful. Our results across four datasets show that the two variants of the proposed approach outperform the advanced model Smoothie-Global by 6.9% and 7.3% points, cross diverse task types including factual recall QA, math reasoning, and instruction following. Notably, we also establish a carefully curated benchmark suite for LLM Ensemble, integrating 12 methods across four classic datasets and three task families, all evaluated under a rigorous and consistent protocol. We hope this repository will help researchers reproduce the LLM Ensemble baselines.

High-quality scientific extreme summary (TLDR) facilitates effective science communication. How do large language models (LLMs) perform in generating them? How are LLM-generated summaries different from those written by human experts? However, the lack of a comprehensive, high-quality scientific TLDR dataset hinders both the development and evaluation of LLMs' summarization ability. To address these, we propose a novel dataset, BiomedTLDR, containing a large sample of researcher-authored summaries from scientific papers, which leverages the common practice of including authors' comments alongside bibliography items. We then test popular open-weight LLMs for generating TLDRs based on abstracts. Our analysis reveals that, although some of them successfully produce humanoid summaries, LLMs generally exhibit a greater affinity for the original text's lexical choices and rhetorical structures, hence tend to be more extractive rather than abstractive in general, compared to humans. Our code and datasets are available at https://github.com/netknowledge/LLM_summarization (Lyu and Ke, 2025).

Large language models (LLMs) have achieved state-of-the-art performance in natural language processing; however, their high computational cost remains a major bottleneck. In this study, we target computational efficiency by focusing on a matrix multiplication free language model (MatMul-free LM) and further reducing the training cost through an architecture inspired by reservoir computing. Specifically, we partially fix and share the weights of selected layers in the MatMul-free LM and insert reservoir layers to obtain rich dynamic representations without additional training overhead. Additionally, several operations are combined to reduce memory accesses. Experimental results show that the proposed architecture reduces the number of parameters by up to 19%, training time by 9.9%, and inference time by 8.0%, while maintaining comparable performance to the baseline model.

This paper develops AI agents that help job seekers write effective requests for job referrals in a professional online community. The basic workflow consists of an improver agent that rewrites the referral request and an evaluator agent that measures the quality of revisions using a model trained to predict the probability of receiving referrals from other users. Revisions suggested by the LLM (large language model) increase predicted success rates for weaker requests while reducing them for stronger requests. Enhancing the LLM with Retrieval-Augmented Generation (RAG) prevents edits that worsen stronger requests while it amplifies improvements for weaker requests. Overall, using LLM revisions with RAG increases the predicted success rate for weaker requests by 14% without degrading performance on stronger requests. Although improvements in model-predicted success do not guarantee more referrals in the real world, they provide low-cost signals for promising features before running higher-stakes experiments on real users.

Clinical document metadata, such as document type, structure, author role, medical specialty, and encounter setting, is essential for accurate interpretation of information captured in clinical documents. However, vast documentation heterogeneity and drift over time challenge harmonization of document metadata. Automated extraction methods have emerged to coalesce metadata from disparate practices into target schema. This scoping review aims to catalog research on clinical document metadata extraction, identify methodological trends and applications, and highlight gaps. We followed the PRISMA-ScR (Preferred Reporting Items for Systematic Reviews and Meta-Analyses Extension for Scoping Reviews) guidelines to identify articles that perform clinical document metadata extraction. We initially found and screened 266 articles published between January 2011 and August 2025, then comprehensively reviewed 67 we deemed relevant to our study. Among the articles included, 45 were methodological, 17 used document metadata as features in a downstream application, and 5 analyzed document metadata composition. We observe myriad purposes for methodological study and application types. Available labelled public data remains sparse except for structural section datasets. Methods for extracting document metadata have progressed from largely rule-based and traditional machine learning with ample feature engineering to transformer-based architectures with minimal feature engineering. The emergence of large language models has enabled broader exploration of generalizability across tasks and datasets, allowing the possibility of advanced clinical text processing systems. We anticipate that research will continue to expand into richer document metadata representations and integrate further into clinical applications and workflows.

Financial reports and earnings communications contain large volumes of structured and semi structured information, making detailed manual analysis inefficient. Earnings conference calls provide valuable evidence about a firm's performance, outlook, and strategic priorities. The manual analysis of lengthy call transcripts requires substantial effort and is susceptible to interpretive bias and unintentional error. In this work, we present a hybrid summarization framework that combines extractive and abstractive techniques to produce concise and factually reliable Reuters-style summaries from the ECTSum dataset. The proposed two stage pipeline first applies the LexRank algorithm to identify salient sentences, which are subsequently summarized using fine-tuned variants of BART and PEGASUS designed for resource constrained settings. In parallel, we fine-tune a Longformer Encoder-Decoder (LED) model to directly capture long-range contextual dependencies in financial documents. Model performance is evaluated using standard automatic metrics, including ROUGE, METEOR, MoverScore, and BERTScore, along with domain-specific variants such as SciBERTScore and FinBERTScore. To assess factual accuracy, we further employ entity-level measures based on source-precision and F1-target. The results highlight complementary trade offs between approaches, long context models yield the strongest overall performance, while the hybrid framework achieves competitive results with improved factual consistency under computational constraints. These findings support the development of practical summarization systems for efficiently distilling lengthy financial texts into usable business insights.

In recent years, using predefined agentic workflows to guide large language models (LLMs) for literature classification and review has become a research focus. However, writing research introductions is more challenging. It requires rigorous logic, coherent structure, and abstract summarization. Existing workflows often suffer from long reasoning chains, error accumulation, and reduced textual coherence. To address these limitations, we propose eliminating external agentic workflows. Instead, we directly parameterize their logical structure into the LLM. This allows the generation of a complete introduction in a single inference. To this end, we introduce the Stage Token for Introduction Generation (STIG). STIG converts the multiple stages of the original workflow into explicit stage signals. These signals guide the model to follow different logical roles and functions during generation. Through instruction tuning, the model learns the mapping between stage tokens and text functions. It also learns the logical order and transition patterns between stages, encoding this knowledge into the model parameters. Experimental results show that STIG can generate multi-stage text in a single inference. It does not require explicit workflow calls. STIG outperforms traditional agentic workflows and other baselines on metrics of semantic similarity and sentence-level structural rationality. The code is provided in the Supplementary Materials.

Cross-domain scientific synthesis requires connecting mechanistic explanations across fragmented literature, a capability that remains challenging for both retrieval-based systems and unconstrained language models. While recent work has applied large language models to scientific summarization and question answering, these approaches provide limited control over reasoning depth and structural grounding. We frame mechanistic synthesis as a graph-constrained multi-hop reasoning problem over literature-derived concept graphs. Given a scientific query and a compact, query-local corpus, SciNets constructs a directed concept graph and synthesizes mechanistic explanations by identifying multi-hop reasoning paths that connect concepts that rarely co-occur within individual papers. We systematically compare shortest-path reasoning, k-shortest paths with diversity constraints, stochastic random walks, and a retrieval-augmented language model baseline. Rather than evaluating correctness, which is often indeterminate when synthesizing connections across distributed sources, we introduce a behavioral framework that measures symbolic reasoning depth, mechanistic diversity, and grounding stability. Across machine learning, biology, and climate science tasks, explicit graph constraints enable controllable multi-hop reasoning while revealing a consistent trade-off: deeper and more diverse symbolic reasoning increases grounding instability, whereas shortest-path reasoning remains highly stable but structurally conservative. These findings provide a systematic behavioral characterization of the limits and capabilities of current graph-LLM integration for scientific synthesis.

Large Language Models (LLMs) are often evaluated against ideals of perfect Bayesian inference, yet growing evidence suggests that their in-context reasoning exhibits systematic forgetting of past information. Rather than viewing this behavior as a limitation, we reinterpret forgetting as a functional cognitive mechanism. Drawing inspiration from human memory dynamics, we model LLM inference as a probabilistic memory process governed by exponential decay. We introduce a benchmark suite that evaluates temporal reasoning, concept drift adaptation, and associative recall, enabling direct comparison between model behavior and human cognitive patterns. Our empirical results reveal that LLMs demonstrate forgetting rates analogous to human memory efficiency trade-offs between stability and adaptability. Building on these observations, we propose probabilistic memory prompting, a lightweight strategy that shapes evidence integration to mimic human-like memory decay, leading to improved long-horizon reasoning performance. Our findings position forgetting not as a failure mode, but as a principled mechanism for adaptive intelligence.

Neural Machine Translation (NMT) systems rely heavily on explicit punctuation cues to resolve semantic ambiguities in a source sentence. Inputting user-generated sentences, which are likely to contain missing or incorrect punctuation, results in fluent but semantically disastrous translations. This work attempts to highlight and address the problem of punctuation robustness of NMT systems through an English-to-Marathi translation. First, we introduce \textbf{Viram}, a human-curated diagnostic benchmark of 54 punctuation-ambiguous English-Marathi sentence pairs to stress-test existing NMT systems. Second, we evaluate two simple remediation strategies: cascade-based restore-then-translate and direct fine-tuning. Our experimental results and analysis demonstrate that both strategies yield substantial NMT performance improvements. Furthermore, we find that current Large Language Models (LLMs) exhibit relatively poorer robustness in translating such sentences than these task-specific strategies, thus necessitating further research in this area. The code and dataset are available at https://github.com/KaustubhShejole/Viram_Marathi.

MixRx uses Large Language Models (LLMs) to classify drug combination interactions as Additive, Synergistic, or Antagonistic, given a multi-drug patient history. We evaluate the performance of 4 models, GPT-2, Mistral Instruct 2.0, and the fine-tuned counterparts. Our results showed a potential for such an application, with the Mistral Instruct 2.0 Fine-Tuned model providing an average accuracy score on standard and perturbed datasets of 81.5%. This paper aims to further develop an upcoming area of research that evaluates if LLMs can be used for biological prediction tasks.

We present a unified framework for Large Language Model (LLM) fine-tuning that integrates Imitation Learning and Reinforcement Learning. By analyzing the gradient of a composite objective combining trajectory-level KL divergence with task rewards, we derive a natural decomposition into two components: (1) an analytically computable Dense Gradient for token-level imitation, and (2) a Monte Carlo estimated Sparse Gradient for long-horizon reward optimization. The Dense Gradient admits a closed-form logit-level formula, enabling efficient GPU implementation.

The introduction of BERT established encoder-only transformer models as a foundational paradigm in natural language processing. Encoder-only models remain the standard tool for classification, tagging and retrieval, where contextual representations and low inference cost matter more than text generation, yet Turkish lacks a monolingual encoder trained from scratch with the advances consolidated in ModernBERT (rotary positional embeddings, FlashAttention, refined normalization). We introduce TabiBERT, a monolingual Turkish encoder based on the ModernBERT architecture, pretrained from scratch for one trillion tokens sampled from an 86.58B-token multi-domain corpus of web text (72%), scientific publications (19%), source code (6%) and mathematical content (0.3%). The model supports a context length of 8,192 tokens, sixteen times that of existing Turkish BERT models, and inherits the ModernBERT architecture's efficiency at long context. For rigorous and reproducible evaluation we introduce TabiBench, a benchmark of 27 datasets across eight task categories with standardized splits and evaluation protocols, summarized as a GLUE-style macro-average on a 0-100 scale. TabiBERT leads the Turkish models in five of eight categories and BERTurk, the previous best, in six of eight; the gains concentrate on question answering (+9.55 F1) and code retrieval (+2.41 NDCG@10), while the four short-text categories are near saturation. Its average of 77.28 exceeds BERTurk's 75.66; the multilingual mmBERT reaches 78.98 with twice the parameters and three times the training tokens, at 41% more tokens per Turkish input. We release model weights, training configurations and evaluation code as a transparent and reproducible foundation for future Turkish encoder research.

Large language models are increasingly deployed in multi-agent workflows. We introduce Prompt Choreography, a framework that efficiently executes LLM workflows by maintaining a dynamic, global KV cache. Each LLM call can attend to an arbitrary, reordered subset of previously encoded messages. Parallel calls are supported. Though caching messages' encodings sometimes gives different results from re-encoding them in a new context, we show in diverse settings that fine-tuning the LLM to work with the cache can help it mimic the original results. Prompt Choreography significantly reduces per-message latency (2.0–6.2$\times$ faster time-to-first-token) and achieves substantial end-to-end speedups ($>$2.2$\times$) in some workflows dominated by redundant computation.

Recent work, using the Biasing Features metric, labels a CoT as unfaithful if it omits a prompt-injected hint that affected the prediction. We argue this metric adopts a narrow notion of faithfulness and confuses unfaithfulness with incompleteness, the lossy compression needed to turn distributed transformer computation into a linear natural language narrative. On multi-hop reasoning tasks with instruct-tuned and reasoning models, many CoTs flagged as unfaithful by Biasing Features are judged faithful by other metrics, exceeding 50% in some models. With a new faithful@k metric, we show that larger inference-time budgets greatly increase hint verbalization (up to 90% in some settings), suggesting much apparent unfaithfulness is due to tight token limits. Using Causal Mediation Analysis, we further show that even non-verbalized hints can causally mediate prediction changes through the CoT. We therefore caution against relying solely on hint-based evaluations and advocate a broader interpretability toolkit, including causal mediation and corruption-based metrics. We do not claim all CoTs are faithful, only that the absence of hint words alone does not prove unfaithfulness.

Multimodal health sensing offers rich behavioral signals for assessing mental health, yet translating these numerical time-series measurements into natural language remains challenging. Current LLMs cannot natively ingest long-duration sensor streams, and paired sensor-text datasets are scarce. To address these challenges, we introduce LENS, a framework that aligns multimodal sensing data with language models to generate clinically grounded mental-health narratives. LENS first constructs a large-scale dataset by transforming Ecological Momentary Assessment (EMA) responses related to depression and anxiety symptoms into natural-language descriptions, yielding over 100,000 sensor-text QA pairs from 258 participants. To enable native time-series integration, we train a patch-level encoder that projects raw sensor signals directly into an LLM's representation space. Our results show that LENS outperforms strong baselines on standard NLP metrics and task-specific measures of symptom-severity accuracy. A user study with 13 mental-health professionals further indicates that LENS-produced narratives are comprehensive and clinically meaningful. Ultimately, our approach advances LLMs as interfaces for health sensing, providing a scalable path toward models that can reason over raw behavioral signals and support downstream clinical decision-making.

Large Language Models (LLMs) demonstrate impressive performance across natural language tasks but incur substantial computational and storage costs due to their scale. Post-training structured pruning offers an efficient solution. However, when few-shot calibration sets fail to adequately reflect the pretraining data distribution, existing methods exhibit limited generalization to downstream tasks. To address this issue, we propose Function-Aware Neuron Grouping (FANG), a post-training pruning framework that alleviates calibration bias by identifying and preserving neurons critical to specific function. FANG groups neurons with similar function based on the type of semantic context they process and prunes each group independently. During importance estimation within each group, tokens that strongly correlate with the functional role of the neuron group are given higher weighting. Additionally, FANG also preserves neurons that contribute across multiple context types. To achieve a better trade-off between sparsity and performance, it allocates sparsity to each block adaptively based on its functional complexity. Experiments show that FANG improves downstream accuracy while preserving language modeling performance. It achieves the state-of-the-art (SOTA) results when combined with FLAP and OBC, two representative pruning methods. Specifically, FANG outperforms FLAP and OBC by 1.5%–8.5% in average accuracy under 30% and 40% sparsity.

Large Language Models (LLMs) have demonstrated promise in medical knowledge assessments, yet their practical utility in real-world clinical decision-making remains underexplored. In this study, we evaluated the performance of three state-of-the-art LLMs-ChatGPT-4o, Gemini 1.5 Pro, and LIama 3.3 70B-in clinical decision support across the entire clinical reasoning workflow of a typical patient encounter. Using 36 case studies, we first assessed LLM's out-of-the-box performance across five key sequential clinical decision-making tasks under two temperature settings (default vs. zero): differential diagnosis, essential immediate steps, relevant diagnostic testing, final diagnosis, and treatment recommendation. All models showed high variability by task, achieving near-perfect accuracy in final diagnosis, poor performance in relevant diagnostic testing, and moderate performance in remaining tasks. Furthermore, ChatGPT performed better under the zero temperature, whereas LIama showed stronger performance under the default temperature. Next, we assessed whether prompt engineering could enhance LLM performance by applying variations of the MedPrompt framework, incorporating targeted and random dynamic few-shot learning. The results demonstrate that prompt engineering is not a one-size-fit-all solution. While it significantly improved the performance on the task with lowest baseline accuracy (relevant diagnostic testing), it was counterproductive for others. Another key finding was that the targeted dynamic few-shot prompting did not consistently outperform random selection, indicating that the presumed benefits of closely matched examples may be counterbalanced by loss of broader contextual diversity. These findings suggest that the impact of prompt engineering is highly model and task-dependent, highlighting the need for tailored, context-aware strategies for integrating LLMs into healthcare.

Recent advancements have shown that reinforcement learning (RL) can substantially improve the reasoning abilities of large language models (LLMs). The effectiveness of such RL training, however, depends critically on the exploration space defined by the pre-trained model's token-output distribution. In this paper, we revisit the standard cross-entropy loss, interpreting it as a specific instance of policy gradient optimization applied within a single-step episode. To systematically study how the pre-trained distribution shapes the exploration potential for subsequent RL, we propose a generalized pre-training objective that adapts on-policy RL principles to supervised learning. By framing next-token prediction as a stochastic decision process, we introduce a reward-shaping strategy that explicitly balances diversity and precision. Our method employs a positive reward scaling factor to control probability concentration on ground-truth tokens and a rank-aware mechanism that treats high-ranking and low-ranking negative tokens asymmetrically. This allows us to reshape the pre-trained token-output distribution and investigate how to provide a more favorable exploration space for RL, ultimately enhancing end-to-end reasoning performance. Contrary to the intuition that higher distribution entropy facilitates effective exploration, we find that imposing a precision-oriented prior yields a superior exploration space for RL.

Multimodal misinformation increasingly leverages visual persuasion, where repurposed or manipulated images strengthen misleading text. We introduce RW-Post, a post-aligned text–image benchmark for real-world multimodal fact-checking with auditable annotations: each instance links the original social-media post with reasoning traces and explicitly linked evidence items derived from human fact-check articles via an LLM-assisted extraction-and-auditing pipeline. RW-Post supports controlled evaluation across closed-book, evidence-bounded, and open-web regimes, enabling systematic diagnosis of visual grounding and evidence utilization. We provide AgentFact as a reference verification baseline and benchmark strong open-source LVLMs under unified protocols. Experiments show substantial headroom: current models struggle with faithful evidence grounding, while evidence-bounded evaluation improves both accuracy and faithfulness.