Reward models (RMs) play a crucial role in Reinforcement Learning from Human Feedback by serving as proxies for human preferences in aligning large language models. In this paper, we identify a model preference bias in RMs, where they systematically assign disproportionately high scores to responses from certain policy models. This bias distorts ranking evaluations and leads to unfair judgments. To address this issue, we propose a calibration method named CHatbot Arena calibrated Reward Modeling (CHARM) that leverages Elo scores from the Chatbot Arena leaderboard to mitigate RM overvaluation. We also introduce a Mismatch Degree metric to measure this preference bias. Our approach is computationally efficient, requiring only a small preference dataset for continued training of the RM. We conduct extensive experiments on reward model benchmarks and human preference alignment. Results demonstrate that our calibrated RMs (1) achieve improved evaluation accuracy on RM-Bench and the Chat-Hard domain of RewardBench, and (2) exhibit a stronger correlation with human preferences by producing scores more closely aligned with Elo rankings. By mitigating model preference bias, our method provides a generalizable and efficient solution for building fairer and more reliable reward models.

The emergence of medical generalist foundation models has revolutionized conventional task-specific model development paradigms, aiming to better handle multiple tasks through joint training on large-scale medical datasets. However, recent advances prioritize simple data scaling or architectural component enhancement, while neglecting to re-examine multi-task learning from a data-centric perspective. Critically, simply aggregating existing data resources leads to decentralized image-task alignment, which fails to cultivate comprehensive image understanding or align with clinical needs for multi-dimensional image interpretation. In this paper, we introduce the image-centric multi-annotation X-ray dataset (IMAX), the first attempt to enhance the multi-task learning capabilities of medical multi-modal large language models (MLLMs) from the data construction level. To be specific, IMAX is featured from the following attributes: 1) High-quality data curation. A comprehensive collection of more than 354K entries applicable to seven different medical tasks. 2) Image-centric dense annotation. Each X-ray image is associated with an average of 4.10 tasks and 7.46 training entries, ensuring multi-task representation richness per image. Compared to the general decentralized multi-annotation X-ray dataset (DMAX), IMAX consistently demonstrates significant multi-task average performance gains ranging from 3.20% to 21.05% across seven open-source state-of-the-art medical MLLMs. Moreover, we investigate differences in statistical patterns exhibited by IMAX and DMAX training processes, exploring potential correlations between optimization dynamics and multi-task performance. Finally, leveraging the core concept of IMAX data construction, we propose an optimized DMAX-based training strategy to alleviate the dilemma of obtaining high-quality IMAX data in practical scenarios.

As Large Language Models (LLMs) become integral to scientific workflows, concerns over the confidentiality and ethical handling of confidential data have emerged. This paper explores data exposure risks through LLM-powered scientific tools, which can inadvertently leak confidential information, including intellectual property and proprietary data, from scientists' perspectives. We propose "DataShield", a framework designed to detect confidential data leaks, summarize privacy policies, and visualize data flow, ensuring alignment with organizational policies and procedures. Our approach aims to inform scientists about data handling practices, enabling them to make informed decisions and protect sensitive information. Ongoing user studies with scientists are underway to evaluate the framework's usability, trustworthiness, and effectiveness in tackling real-world privacy challenges.

Retrieval-Augmented Generation (RAG) is a promising technique for applying LLMs to proprietary domains. However, retrieved documents may contain sensitive knowledge, posing risks of privacy leakage in generative results. Thus, effectively erasing private information from retrieved documents is a key challenge for RAG. Unlike traditional text anonymization, RAG should consider: (1) the inherent multi-document reasoning may face de-anonymization attacks; (2) private knowledge varies by scenarios, so users should be allowed to customize which information to erase; (3) preserving sufficient publicly available knowledge for generation tasks. This paper introduces the privacy erasure task for RAG and proposes Eraser4RAG, a private knowledge eraser which effectively removes user-defined private knowledge from documents while preserving sufficient public knowledge for generation. Specifically, we first construct a global knowledge graph to identify potential knowledge across documents, aiming to defend against de-anonymization attacks. Then we randomly split it into private and public sub-graphs, and fine-tune Flan-T5 to rewrite the retrieved documents excluding private triples. Finally, PPO algorithm optimizes the rewriting model to minimize private triples and maximize public triples retention. Experiments on four QA datasets demonstrate that Eraser4RAG achieves superior erase performance than GPT-4o.

Legal writing demands clarity, formality, and domain-specific precision-qualities often lacking in documents authored by individuals without legal training. To bridge this gap, this paper explores the task of legal text refinement that transforms informal, conversational inputs into persuasive legal arguments. We introduce FinDR, a Chinese dataset of financial dispute records, annotated with official judgments on claim reasonableness. Our proposed method, Multi-Scale Model Interaction (MSMI), leverages a lightweight classifier to evaluate outputs and guide iterative refinement by Large Language Models (LLMs). Experimental results demonstrate that MSMI significantly outperforms single-pass prompting strategies. Additionally, we validate the generalizability of MSMI on several short-text benchmarks, showing improved adversarial robustness. Our findings reveal the potential of multi-model collaboration for enhancing legal document generation and broader text refinement tasks.

Sentiment analysis is a crucial task in natural language processing (NLP) that enables the extraction of meaningful insights from textual data, particularly from dynamic platforms like Twitter and IMDB. This study explores a hybrid framework combining transformer-based models, specifically BERT, GPT-2, RoBERTa, XLNet, and DistilBERT, to improve sentiment classification accuracy and robustness. The framework addresses challenges such as noisy data, contextual ambiguity, and generalization across diverse datasets by leveraging the unique strengths of these models. BERT captures bidirectional context, GPT-2 enhances generative capabilities, RoBERTa optimizes contextual understanding with larger corpora and dynamic masking, XLNet models dependency through permutation-based learning, and DistilBERT offers efficiency with reduced computational overhead while maintaining high accuracy. We demonstrate text cleaning, tokenization, and feature extraction using Term Frequency Inverse Document Frequency (TF-IDF) and Bag of Words (BoW), ensure high-quality input data for the models. The hybrid approach was evaluated on benchmark datasets Sentiment140 and IMDB, achieving superior accuracy rates of 94\% and 95\%, respectively, outperforming standalone models. The results validate the effectiveness of combining multiple transformer models in ensemble-like setups to address the limitations of individual architectures. This research highlights its applicability to real-world tasks such as social media monitoring, customer sentiment analysis, and public opinion tracking which offers a pathway for future advancements in hybrid NLP frameworks.

--T ask execution for object rearrangement could be challenged by T ask-Level Perturbations (TLP)--unexpected object additions, removals, and displacements--that can disrupt underlying visual policies and fundamentally compromise task feasibility and progress. T o address these challenges, we present LangPert, a language-based framework designed to detect and mitigate TLP situations in tabletop rearrangement tasks. Lang-Pert integrates a Visual Language Model (VLM) to comprehensively monitor policy's skill execution and environmental TLP, while leveraging the Hierarchical Chain-of-Thought (HCoT) reasoning mechanism to enhance the Large Language Model (LLM)'s contextual understanding and generate adaptive, corrective skill-execution plans. Our experimental results demonstrate that LangPert handles diverse TLP situations more effectively than baseline methods, achieving higher task completion rates, improved execution efficiency, and potential generalization to unseen scenarios. Large Language Models (LLMs) have demonstrated significant potential as task planners [1], [13], [14], effectively converting a user's task description into robot skill instructions. When combined with visual perception capabilities [12] and trained on robot-specific datasets [3], [19], LLMs can serve as both task planners and environment perceivers. This enables robots to detect execution failures and re-plan skills in real-time [9], [18], [21], making LLMs a promising approach to enhance robotic autonomy in unstructured environments [20]. However, most approaches focus on task completion in static environments, overlooking the dynamic nature of real-world scenarios where unexpected workspace changes could occur due to factors such as human intervention [6]. These unpredictable disturbances [11] can significantly impact task progress and ultimately compromise task completability.

This paper leverages past sentence processing studies to investigate whether monolingual and multilingual LLMs show human-like preferences when presented with examples of relative clause attachment ambiguities in Italian and English. Furthermore, we test whether these preferences can be modulated by lexical factors (the type of verb/noun in the matrix clause) which have been shown to be tied to subtle constraints on syntactic and semantic relations. Our results overall showcase how LLM behavior varies interestingly across models, but also general failings of these models in correctly capturing human-like preferences. In light of these results, we argue that RC attachment is the ideal benchmark for cross-linguistic investigations of LLMs' linguistic knowledge and biases.

Recent LLMs have significantly improved reasoning capabilities, primarily by including an explicit, lengthy Thinking process as part of generation. In this paper, we question whether this explicit thinking is necessary. Using the state-of-the-art DeepSeek-R1-Distill-Qwen, we find that bypassing the thinking process via simple prompting, denoted as NoThinking, can be surprisingly effective. When controlling for the number of tokens, NoThinking outperforms Thinking across a diverse set of seven challenging reasoning datasets--including mathematical problem solving, formal theorem proving, and coding--especially in low-budget settings, e.g., 51.3 vs. 28.9 on ACM 23 with 700 tokens. Notably, the performance of NoThinking becomes more competitive with pass@k as k increases. Building on this observation, we demonstrate that a parallel scaling approach that uses NoThinking to generate N outputs independently and aggregates them is highly effective. For aggregation, we use task-specific verifiers when available, or we apply simple best-of-N strategies such as confidence-based selection. Our method outperforms a range of baselines with similar latency using Thinking, and is comparable to Thinking with significantly longer latency (up to 9x). Together, our research encourages a reconsideration of the necessity of lengthy thinking processes, while also establishing a competitive reference for achieving strong reasoning performance in low-budget settings or at low latency using parallel scaling.

Vaccine hesitancy and misinformation are significant barriers to achieving widespread vaccination coverage. Smaller public health departments may lack the expertise or resources to craft effective vaccine messaging. This paper explores the potential of ChatGPT-augmented messaging to promote confidence in vaccination uptake. We conducted a survey in which participants chose between pairs of vaccination messages and assessed which was more persuasive and to what extent. In each pair, one message was the original, and the other was augmented by ChatGPT. At the end of the survey, participants were informed that half of the messages had been generated by ChatGPT. They were then asked to provide both quantitative and qualitative responses regarding how knowledge of a message's ChatGPT origin affected their impressions. Overall, ChatGPT-augmented messages were rated slightly higher than the original messages. These messages generally scored better when they were longer. Respondents did not express major concerns about ChatGPT-generated content, nor was there a significant relationship between participants' views on ChatGPT and their message ratings. Notably, there was a correlation between whether a message appeared first or second in a pair and its score. These results point to the potential of ChatGPT to enhance vaccine messaging, suggesting a promising direction for future research on human-AI collaboration in public health communication.