Large Language Models (LLMs) have been shown to be able to learn different tasks without explicit finetuning when given many input-output examples / demonstrations through In-Context Learning (ICL). Increasing the number of examples, called ``shots'', improves downstream task performance but incurs higher memory and computational costs. In this work, we study an approach to improve the memory and computational efficiency of ICL inference by compressing the many-shot prompts. Given many shots comprising t tokens, our goal is to generate a m soft-token summary, where m < t. We first show that existing prompt compression methods are ineffective for many-shot compression, and simply using fewer shots as a baseline is surprisingly strong. To achieve effective compression, we find that: (a) a stronger compressor model with more trainable parameters is necessary, and (b) compressing many-shot representations at each transformer layer enables more fine-grained compression by providing each layer with its own compressed representation. Based on these insights, we propose MemCom, a layer-wise compression method. We systematically evaluate various compressor models and training approaches across different model sizes (2B and 7B), architectures (Gemma and Mistral), many-shot sequence lengths (3k-6k tokens), and compression ratios (3x to 8x). MemCom outperforms strong baselines across all compression ratios on multiple classification tasks with large label sets. Notably, while baseline performance degrades sharply at higher compression ratios, often by over 20-30%, MemCom maintains high accuracy with minimal degradation, typically dropping by less than 10%.

This study quantifies how prompting strategies interact with large language models (LLMs) to automate the screening stage of systematic literature reviews (SLRs). We evaluate six LLMs (GPT-4o, GPT-4o-mini, DeepSeek-Chat-V3, Gemini-2.5-Flash, Claude-3.5-Haiku, Llama-4-Maverick) under five prompt types (zero-shot, few-shot, chain-of-thought (CoT), CoT-few-shot, self-reflection) across relevance classification and six Level-2 tasks, using accuracy, precision, recall, and F1. Results show pronounced model-prompt interaction effects: CoT-few-shot yields the most reliable precision-recall balance; zero-shot maximizes recall for high-sensitivity passes; and self-reflection underperforms due to over-inclusivity and instability across models. GPT-4o and DeepSeek provide robust overall performance, while GPT-4o-mini performs competitively at a substantially lower dollar cost. A cost-performance analysis for relevance classification (per 1,000 abstracts) reveals large absolute differences among model-prompt pairings; GPT-4o-mini remains low-cost across prompts, and structured prompts (CoT/CoT-few-shot) on GPT-4o-mini offer attractive F1 at a small incremental cost. We recommend a staged workflow that (1) deploys low-cost models with structured prompts for first-pass screening and (2) escalates only borderline cases to higher-capacity models. These findings highlight LLMs' uneven but promising potential to automate literature screening. By systematically analyzing prompt-model interactions, we provide a comparative benchmark and practical guidance for task-adaptive LLM deployment.

Large language models (LLMs) increasingly require mechanisms for continual adaptation without full retraining. However, sequential updates can lead to catastrophic forgetting, where new edits degrade previously acquired knowledge. This work presents STABLE, a gated continual self editing framework that constrains forgetting during sequential updates using parameter efficient fine tuning via Low Rank Adaptation (LoRA; see arXiv:2106.09685). Each candidate edit is evaluated against a stability budget using one of three metrics: (i) Exact Match (EM) drop, capturing factual accuracy loss; (ii) bits increase, reflecting reduced model confidence; and (iii) KL divergence, quantifying distributional drift between the base and adapted models. If a threshold is exceeded, the LoRA update is rescaled through a clipping procedure or rejected. Experiments on the Qwen-2.5-7B model show that gating effectively mitigates forgetting while preserving adaptability. EM based gating achieved the highest cumulative performance in short continual learning sequences. Our results show that different gating strategies can achieve comparable distribution shift (measured by KL divergence) while producing different accuracy outcomes, highlighting the importance of gating design in continual adaptation. This approach offers a principled method for continual model editing, enabling LLMs to integrate new knowledge while maintaining reliability. Code: https://github.com/Bhoy1/STABLE

Abstract--The 2022 World Mental Health Report calls for global mental health care reform, amid rising prevalence of issues like anxiety and depression that affect nearly one billion people worldwide. Traditional in-person therapy fails to meet this demand, and the situation is worsened by stigma. While general-purpose large language models (LLMs) offer efficiency for AI-driven mental health solutions, they underperform because they lack specialized fine-tuning. Existing LLM-based mental health chatbots can engage in empathetic conversations, but they overlook real-time user mental state assessment which is critical for professional counseling. This paper proposes MoPHES, a framework that integrates mental state evaluation, conversational support, and professional treatment recommendations. The agent developed under this framework uses two fine-tuned MiniCPM4-0.5B LLMs: one is fine-tuned on mental health conditions datasets to assess users' mental states and predict the severity of anxiety and depression; the other is fine-tuned on multi-turn dialogues to handle conversations with users. By leveraging insights into users' mental states, our agent provides more tailored support and professional treatment recommendations. Both models are also deployed directly on mobile devices to enhance user convenience and protect user privacy. Additionally, to evaluate the performance of MoPHES with other LLMs, we develop a benchmark for the automatic evaluation of mental state prediction and multi-turn counseling dialogues, which includes comprehensive evaluation metrics, datasets, and methods. ENT AL health issues are emerging as an increasingly severe threat to global public health, with their prevalence rising annually.

While clustering methods such as spectral clustering and K-means effectively group genes by expression similarity, downstream interpretation is typically performed using enrichment-based statistics. These approaches provide high-level functional summaries but often fail to yield cluster-specific mechanistic insight or explicit links to supporting literature. As a result, biological interpretation frequently relies on manual curation, limiting reproducibility and scalability. Large language models (LLMs) have recently emerged as powerful tools for biomedical text mining and knowledge synthesis. Although LLMs can generate fluent biological narratives, they are optimized for linguistic coherence rather than evidential accountability. When applied directly to transcriptomic interpretation, they may produce plausible but unverifiable statements, omit explicit citations, or hallucinate unsupported claims. While retrieval-augmented and agentic systems partially address this issue, systematic verification and critic-based validation remain underexplored. This limitation is particularly consequential for antimicrobial resistance research in Salmonella enterica, a major foodborne pathogen responsible for substantial global morbidity.

While Artificial Intelligence (AI) shows promise in healthcare applications, existing conversational systems often falter in complex and sensitive medical domains such as Sexual and Reproductive Health (SRH). These systems frequently struggle with hallucination and lack the specialized knowledge required, particularly for sensitive SRH topics. Furthermore, current AI approaches in healthcare tend to prioritize diagnostic capabilities over comprehensive patient care and education. Addressing these gaps, this work at the UNC School of Nursing introduces SARHAchat, a proof-of-concept Large Language Model (LLM)- based chatbot. SARHAchat is designed as a reliable, user-centered system integrating medical expertise with empathetic communication to enhance SRH care delivery. Our evaluation demonstrates SARHAchat's ability to provide accurate and contextually appropriate contraceptive counseling while maintaining a natural conversational flow. The demo is available at https://sarhachat.com/.

Emergency departments worldwide face rising patient volumes, workforce shortages, and variability in triage decisions that threaten the delivery of timely and accurate care. Current triage methods rely primarily on vital signs, routine laboratory values, and clinicians' judgment, which, while effective, often miss emerging biological signals that could improve risk prediction for infection typing or antibiotic administration in acute conditions. To address this challenge, we introduce TriAgent, a large language model (LLM)-based multi-agent framework that couples automated biomarker discovery with deep research for literature-grounded validation and novelty assessment. TriAgent employs a supervisor research agent to generate research topics and delegate targeted queries to specialized sub-agents for evidence retrieval from various data sources. Findings are synthesized to classify biomarkers as either grounded in existing knowledge or flagged as novel candidates, offering transparent justification and highlighting unexplored pathways in acute care risk stratification. Unlike prior frameworks limited to existing routine clinical biomarkers, TriAgent aims to deliver an end-to-end framework from data analysis to literature grounding to improve transparency, explainability and expand the frontier of potentially actionable clinical biomarkers. Given a user's clinical query and quantitative triage data, TriAgent achieved a topic adherence F1 score of 55.7 +/- 5.0%, surpassing the CoT-ReAct agent by over 10%, and a faithfulness score of 0.42 +/- 0.39, exceeding all baselines by more than 50%. Across experiments, TriAgent consistently outperformed state-of-the-art LLM-based agentic frameworks in biomarker justification and literature-grounded novelty assessment. We share our repo: https://github.com/CellFace/TriAgent.

Current Large Language Model (LLM) agents show strong performance in tool use, but lack the crucial capability to systematically learn from their own experiences. While existing frameworks mainly focus on mitigating external knowledge gaps, they fail to address a more fundamental limitation: the inability to itera-tively refine problem-solving strategies. In this work, we introduce EvolveR, a framework designed to enable agent to self-improve through a complete, closed-loop experience lifecycle. This lifecycle comprises two key stages: (1) Offline Self-Distillation, where the agent's interaction trajectories are synthesized into a structured repository of abstract, reusable strategic principles; (2) Online Interaction, where the agent interacts with tasks and actively retrieves distilled principles to guide its decision-making, accumulating a diverse set of behavioral trajectories. This loop employs a policy reinforcement mechanism to it-eratively update the agent based on its performance. We demonstrate the effectiveness of EvolveR on complex multi-hop question-answering benchmarks, where it achieves superior performance over strong agentic baselines. Our work presents a comprehensive blueprint for agents that learn not only from external data but also from the consequences of their own actions, paving the way for more autonomous and continuously improving systems. Code is available at https://github.com/Edaizi/EvolveR. Large Language Models (LLMs) have driven the development of autonomous agents capable of solving diverse tasks through advanced reasoning and tool use [1-3]. However, a significant limitation emerges when these agents engage in sequential tasks: each interaction is treated independently. They approach tasks as isolated episodes, suffering from operational amnesia and failing to learn from past successes or avoid prior mistakes[4].

We propose utilizing fast and slow thinking to enhance the capabilities of large language model-based agents on complex tasks such as program repair. In particular, we design an adaptive program repair method based on issue description response, called SIADAFIX. The proposed method utilizes slow thinking bug fix agent to complete complex program repair tasks, and employs fast thinking workflow decision components to optimize and classify issue descriptions, using issue description response results to guide the orchestration of bug fix agent workflows. SIADAFIX adaptively selects three repair modes, i.e., easy, middle and hard mode, based on problem complexity. It employs fast generalization for simple problems and test-time scaling techniques for complex problems. Experimental results on the SWE-bench Lite show that the proposed method achieves 60.67% pass@1 performance using the Claude-4 Sonnet model, reaching state-of-the-art levels among all open-source methods. SIADAFIX effectively balances repair efficiency and accuracy, providing new insights for automated program repair. Our code is available at https://github.com/liauto-siada/siada-cli.

Abstract--This study presents a novel multi-model fusion framework leveraging two state-of-the-art large language models (LLMs), ChatGPT and Claude, to enhance the reliability of chest X-ray interpretation on the CheXpert dataset. From the full CheXpert corpus of 224,316 chest radiographs, we randomly selected 234 radiologist-annotated studies to evaluate unimodal performance using image-only prompts. In this setting, ChatGPT and Claude achieved diagnostic accuracies of 62.8% and 76.9%, respectively. A similarity-based consensus approach, using a 95% output similarity threshold, improved accuracy to 77.6%. T o assess the impact of multimodal inputs, we then generated synthetic clinical notes following the MIMIC-CXR template and evaluated a separate subset of 50 randomly selected cases paired with both images and synthetic text. On this multimodal cohort, performance improved to 84% for ChatGPT and 76% for Claude, while consensus accuracy reached 91.3%. Across both experimental conditions, agreement-based fusion consistently outperformed individual models. These findings highlight the utility of integrating complementary modalities and using output-level consensus to improve the trustworthiness and clinical utility of AI-assisted radiological diagnosis, offering a practical path to reduce diagnostic errors with minimal computational overhead.