While Medical Large Language Models (MedLLMs) have demonstrated remarkable potential in clinical tasks, their ethical safety remains insufficiently explored. This paper introduces $\textbf{MedEthicsQA}$, a comprehensive benchmark comprising $\textbf{5,623}$ multiple-choice questions and $\textbf{5,351}$ open-ended questions for evaluation of medical ethics in LLMs. We systematically establish a hierarchical taxonomy integrating global medical ethical standards. The benchmark encompasses widely used medical datasets, authoritative question banks, and scenarios derived from PubMed literature. Rigorous quality control involving multi-stage filtering and multi-faceted expert validation ensures the reliability of the dataset with a low error rate ($2.72\%$). Evaluation of state-of-the-art MedLLMs exhibit declined performance in answering medical ethics questions compared to their foundation counterparts, elucidating the deficiencies of medical ethics alignment. The dataset, registered under CC BY-NC 4.0 license, is available at https://github.com/JianhuiWei7/MedEthicsQA.

Vision-language models (VLM) have demonstrated impressive performance in image recognition by leveraging self-supervised training on large datasets. Their performance can be further improved by adapting to the test sample using test-time prompt tuning (TPT). Unfortunately, the singular focus of TPT approaches on improving the accuracy suffers from tunnel vision, and leads to degradation in confidence calibration. This limits the applicability of TPT in critical applications. We make three contributions in this work. (1) We posit that random or naive initialization of prompts leads to overfitting on a particular test sample, and is the main reason for miscalibration of the VLM after TPT. To mitigate the problem, we propose careful initialization of test time prompt using prior knowledge about the target label attributes from a large language model (LLM); (2) To further maintain the quality of prompts during \tpt, we propose a novel regularization loss to reduce intraclass distance, and increase inter-class distance between the learnt Through extensive experiments on different CLIP architectures and 15 datasets, we show that our approach can effectively improve the calibration after TPT. We report an average expected calibration error (ECE) of 4.11 with our method, TCA, compared to 11.7 for vanilla TPT, 6.12 for C-TPT (ICLR'24), 6.78 for DiffTPT (CVPR'23), and 8.43 for PromptAlign (NeurIPS'23). The code is publicly accessible at: https://github.com/rhebbalaguppe/TCA_PromptWithoutPanic.

T o our knowledge, this is the first system to apply retrieval-based prompting for OpenMP pragma correctness without model fine-tuning or compiler instrumentation. P4OMP leverages Retrieval-Augmented Generation (RAG) with structured instructional knowledge from OpenMP tutorials to improve the reliability of prompt-driven code generation. By grounding generation in the retrieved context, P4OMP improves syntactic correctness compared to baseline prompting with GPT -3.5-T urbo. We evaluate P4OMP against a baseline--GPT -3.5-T urbo without retrieval--on a comprehensive benchmark of 108 real-world C++ programs drawn from Stack Overflow, PolyBench, and NAS benchmark suites. P4OMP achieves 100% compilation success on all parallelizable cases, while the baseline fails to compile in 20 out of 108 cases. Six cases that rely on non-random-access iterators or thread-unsafe constructs are excluded due to fundamental OpenMP limitations. A detailed analysis demonstrates how P4OMP consistently avoids scoping errors, syntactic misuse, and invalid directive combinations that commonly affect baseline-generated code. We further demonstrate strong runtime scaling across seven compute-intensive benchmarks on an HPC cluster . P4OMP offers a robust, modular pipeline that significantly improves the reliability and applicability of LLM-generated OpenMP code.

Multimodal embedding models, built upon causal Vision Language Models (VLMs), have shown promise in various tasks. However, current approaches face three key limitations: the use of causal attention in VLM backbones is suboptimal for embedding tasks; scalability issues due to reliance on high-quality labeled paired data for contrastive learning; and limited diversity in training objectives and data. To address these issues, we propose MoCa, a two-stage framework for transforming pre-trained VLMs into effective bidirectional multimodal embedding models. The first stage, Modality-aware Continual Pre-training, introduces a joint reconstruction objective that simultaneously denoises interleaved text and image inputs, enhancing bidirectional context-aware reasoning. The second stage, Heterogeneous Contrastive Fine-tuning, leverages diverse, semantically rich multimodal data beyond simple image-caption pairs to enhance generalization and alignment. Our method addresses the stated limitations by introducing bidirectional attention through continual pre-training, scaling effectively with massive unlabeled datasets via joint reconstruction objectives, and utilizing diverse multimodal data for enhanced representation robustness. Experiments demonstrate that MoCa consistently improves performance across MMEB and ViDoRe-v2 benchmarks, achieving new state-of-the-art results, and exhibits strong scalability with both model size and training data on MMEB.

Hierarchical planning is a powerful approach to model long sequences structurally. Aside from considering hierarchies in the temporal structure of music, this paper explores an even more important aspect: concept hierarchy, which involves generating music ideas, transforming them, and ultimately organizing them--across musical time and space--into a complete composition. To this end, we introduce TOMI (Transforming and Organizing Music Ideas) as a novel approach in deep music generation and develop a TOMI-based model via instruction-tuned foundation LLM. Formally, we represent a multi-track composition process via a sparse, four-dimensional space characterized by clips (short audio or MIDI segments), sections (temporal positions), tracks (instrument layers), and transformations (elaboration methods). Our model is capable of generating multi-track electronic music with full-song structure, and we further integrate the TOMI-based model with the REAPER digital audio workstation, enabling interactive human-AI co-creation. Experimental results demonstrate that our approach produces higher-quality electronic music with stronger structural coherence compared to baselines.

Peer-to-peer (P2P) trading is increasingly recognized as a key mechanism for decentralized market regulation, yet existing approaches often lack robust frameworks to ensure fairness. This paper presents FairMarket-RL, a novel hybrid framework that combines Large Language Models (LLMs) with Reinforcement Learning (RL) to enable fairness-aware trading agents. In a simulated P2P microgrid with multiple sellers and buyers, the LLM acts as a real-time fairness critic, evaluating each trading episode using two metrics: Fairness-To-Buyer (FTB) and Fairness-Between-Sellers (FBS). These fairness scores are integrated into agent rewards through scheduled λ-coefficients, forming an adaptive LLM-guided reward shaping loop that replaces brittle, rule-based fairness constraints. Agents are trained using Independent Proximal Policy Optimization (IPPO) and achieve equitable outcomes, fulfilling over 90% of buyer demand, maintaining fair seller margins, and consistently reaching FTB and FBS scores above 0.80. The training process demonstrates that fairness feedback improves convergence, reduces buyer shortfalls, and narrows profit disparities between sellers. With its language-based critic, the framework scales naturally, and its extension to a large power distribution system with household prosumers illustrates its practical applicability. FairMarket-RL thus offers a scalable, equity-driven solution for autonomous trading in decentralized energy systems.

We present a reproduction study of "Competition of Mechanisms: Tracing How Language Models Handle Facts and Counterfactuals" (Ortu et al., 2024), which investigates competition of mechanisms in language models between factual recall and counterfactual in-context repetition. Our study successfully reproduces their primary findings regarding the localization of factual and counterfactual information, the dominance of attention blocks in mechanism competition, and the specialization of attention heads in handling competing information. We reproduce their results on both GPT-2 (Radford et al., 2019) and Pythia 6.9B (Biderman et al., 2023). We extend their work in three significant directions. First, we explore the generalizability of these findings to even larger models by replicating the experiments on Llama 3.1 8B (Grattafiori et al., 2024), discovering greatly reduced attention head specialization. Second, we investigate the impact of prompt structure by introducing variations where we avoid repeating the counterfactual statement verbatim or we change the premise word, observing a marked decrease in the logit for the counterfactual token. Finally, we test the validity of the authors' claims for prompts of specific domains, discovering that certain categories of prompts skew the results by providing the factual prediction token as part of the subject of the sentence. Overall, we find that the attention head ablation proposed in Ortu et al. (2024) is ineffective for domains that are underrepresented in their dataset, and that the effectiveness varies based on model architecture, prompt structure, domain and task.

We present a new benchmark for evaluating Deep Search--a realistic and complex form of retrieval-augmented generation (RAG) that requires source-aware, multi-hop reasoning over diverse, sparsed, but related sources. These include documents, meeting transcripts, Slack messages, GitHub, and URLs, which vary in structure and often contain human-to-human interactions. We build it using a synthetic data pipeline that simulates business workflows across product planning, development, and support stages, generating interconnected content with realistic noise and multi-hop questions with guaranteed ground-truth answers. We release our benchmark with both answerable and unanswerable queries, and retrieval pool of 39,190 enterprise artifacts, enabling fine-grained evaluation of long-context LLM and RAG systems. Our experiments reveal that even the best-performing agentic RAG methods achieve an average performance score of 32.96 on our benchmark. With further analysis, we highlight retrieval as the main bottleneck: existing methods struggle to conduct deep searches and retrieve all necessary evidence. Consequently, they often reason over partial context, leading to significant performance degradation.

Recent advancements in Large Language Models (LLMs) have revealed a significant performance gap between closed-source and open-source models, particularly in tasks requiring complex reasoning and precise instruction following. This paper introduces ReasonBridge, a methodology that efficiently transfers reasoning capabilities from powerful closed-source to open-source models through a novel hierarchical knowledge distillation framework. We develop a tailored dataset Reason1K with only 1,000 carefully curated reasoning traces emphasizing difficulty, diversity, and quality. These traces are filtered from across multiple domains using a structured multi-criteria selection algorithm. Our transfer learning approach incorporates: (1) a hierarchical distillation process capturing both strategic abstraction and tactical implementation patterns, (2) a sparse reasoning-focused adapter architecture requiring only 0.3% additional trainable parameters, and (3) a test-time compute scaling mechanism using guided inference interventions. Comprehensive evaluations demonstrate that ReasonBridge improves reasoning capabilities in open-source models by up to 23% on benchmark tasks, significantly narrowing the gap with closed-source models. Notably, the enhanced Qwen2.5-14B outperforms Claude-Sonnet3.5 on MATH500 and matches its performance on competition-level AIME problems. Our methodology generalizes effectively across diverse reasoning domains and model architectures, establishing a sample-efficient approach to reasoning enhancement for instruction following.

In today's digital world, casual user-generated content often contains subtle cues that may inadvertently expose sensitive personal attributes. Such risks underscore the growing importance of effective text anonymization to safeguard individual privacy. However, existing methods either rely on rigid replacements that damage utility or cloud-based LLMs that are costly and pose privacy risks. To address these issues, we explore the use of locally deployed smaller-scale language models (SLMs) for anonymization. Yet training effective SLMs remains challenging due to limited high-quality supervision. To address the challenge, we propose AgentStealth, a self-reinforcing LLM anonymization framework.First, we introduce an adversarial anonymization workflow enhanced by In-context Contrastive Learning and Adaptive Utility-Aware Control. Second, we perform supervised adaptation of SLMs using high-quality data collected from the workflow, which includes both anonymization and attack signals. Finally, we apply online reinforcement learning where the model leverages its internal adversarial feedback to iteratively improve anonymization performance. Experiments on two datasets show that our method outperforms baselines in both anonymization effectiveness (+12.3%) and utility (+6.8%). Our lightweight design supports direct deployment on edge devices, avoiding cloud reliance and communication-based privacy risks. Our code is open-source at https://github.com/tsinghua-fib-lab/AgentStealth.