There has been a growing interest in translating C code to Rust due to Rust’s robust memory and thread safety guarantees. Tools such as C2Rust enable syntax-guided transpilation from C to semantically equivalent Rust code. However, the resulting Rust programs often rely heavily on unsafe constructs, particularly raw pointers, which undermines Rust’s safety guarantees. This paper aims to improve the memory safety of Rust programs generated by C2Rust by eliminating raw pointers. Specifically, we propose a raw pointer rewriting technique that lifts raw pointers in individual functions to appropriate Rust data structures. Technically, PR2 employs decision-tree-based prompting to guide the pointer lifting process. It also leverages code change analysis to guide the repair of errors introduced during rewriting, effectively addressing errors encountered during compilation and test case execution.We implement PR2 and evaluate it using gpt-4o-mini on 28 real-world C projects. It is shown that PR2 successfully eliminates 18.57% of local raw pointers across these projects, significantly enhancing the safety of the translated Rust code. On average, PR2 completes the transformation of a project in 5.02 hours, at a cost of $1.13. Our code is available at https://github.com/bhcsayx/PR2.
论文检索
输入标题、作者或关键词，从 9,460 篇学术成果中精准定位
Recent advances in summarization research focus on improving summary quality across multiple criteria, such as completeness, conciseness, and faithfulness, by jointly optimizing these dimensions. However, these efforts largely overlook the challenge of controlling summary generation with respect to individual criteria, especially in the presence of their inherent trade-offs. For example, enhancing conciseness can compromise completeness, and vice versa. In this work, we address this gap by proposing a loss function that aligns model outputs with fine-grained, model-based evaluation scores (e.g., from FineSurE), enabling both improvement in summary quality and dimension-specific control. Our approach improves the overall quality of summaries while maintaining the ability to selectively prioritize one criterion over others. Experiments on three pretrained models (LLaMA, Qwen, and Mistral) demonstrate that our method achieves performance comparable to state-of-the-art summarizers, while uniquely offering strong controllability over individual quality dimensions.
SafeMERGE: Preserving Safety Alignment in Fine-Tuned Large Language Models via Selective Layer-Wise Model Merging
PDF ↗Fine-tuning large language models (LLMs) is a common practice to adapt generalist models to specialized domains. However, recent studies show that fine-tuning can erode safety alignment, causing LLMs to respond to harmful or unethical prompts. Many methods to realign safety have been proposed, but often introduce custom algorithms that are difficult to implement or compromise task utility. In this work, we propose SafeMERGE, a lightweight, post-fine-tuning framework that restores safety while maintaining downstream performance. SafeMERGE selectively merges fine-tuned with safety-aligned model layers only when they deviate from safe behavior, measured by a cosine similarity criterion. Across four LLMs and several tasks, SafeMERGE consistently reduces harmful outputs compared to other defenses, with negligible or even positive impact on utility. Our results demonstrate that selective, layer-wise merging offers a robust safeguard against the inadvertent loss of safety during fine-tuning, establishing SafeMERGE as a simple yet effective post-fine-tuning defense.
Recent advances in summary evaluation are based on model-based metrics to assess quality dimensions, such as completeness, conciseness, and faithfulness. However, these methods often require large language models, and predicted scores are frequently miscalibrated, limiting their reliability. Moreover, evaluating the average quality across different summaries for a single document typically requires access to multiple reference summaries. Here, we propose a general framework that generates individual and average proxy scores without relying on reference summaries, human annotations, or expensive model-based metrics. We also propose group isotonic regression binning (GIRB), a calibration method that adjusts the raw predictions to better align with ground-truth evaluation metrics. While we focus on continuous-value scenarios, such as summarization, the method is applicable to discrete-value tasks, such as question answering. Experiments on seven datasets demonstrate that our approach consistently outperforms existing baselines.
Modern instruction-following language models are optimized to be helpful and cooperative, often through preference-based alignment such as RLHF and related methods. A growing body of evidence shows that this training can also induce sycophancy: models may agree with a user even when the user is wrong, undermining reliability in decision support and high-stakes advice. We introduce SycoBench-600, a controlled multiple-choice benchmark that measures (i) susceptibility to three social-pressure perturbations (doubt, authority, and an explicit wrong suggestion) and (ii) correction selectivity, the ability to accept correct suggestions while resisting incorrect ones. The released benchmark contains 600 English MCQ instances over 272 normalized question stems, covers 8 domains and 3 difficulty tiers, and evaluates each instance under 3 fixed paraphrase variants of the perturbation prompts. We evaluate seven widely used assistants spanning proprietary and open-weight families. Results show substantial variation in pressure robustness and selective updating, and further show that willingness to update does not by itself imply selectivity. We release raw logs, validation scripts, and code that regenerates every table and figure from the model outputs.
Visual text compression is an emerging paradigm for rendering text as images for processing by vision-language models (VLMs), enabling higher information density per context token. However, the robustness of VLMs under dense, text-based visual inputs remains unevaluated. We introduce Fico, a benchmark designed to assess VLM robustness across seven controlled variants of visual fidelity and information density. Fico spans documents of 8k to 64k tokens and includes three tasks of increasing semantic granularity: optical character recognition (OCR), needle-in-a-haystack (NIAH) retrieval, and visual question answering (VQA). Evaluating 13 general-purpose VLMs and 3 OCR-specialized models reveals three consistent trends: performance drops sharply under increased density or reduced resolution; cross-task transfer between OCR, NIAH, and VQA is limited; and VQA is comparatively robust because low-level details are lost before high-level semantics. By exposing failure modes that remain invisible under conventional VLM evaluations, Fico establishes a rigorous test-bed for visual text compression.
Conversational AI is constrained in many real-world settings where only one side of a dialogue can be recorded. We formalize the one-sided conversation problem (1SC): inferring and learning from only one side of a conversation. We study two tasks: (1) reconstructing the missing speaker’s turns and (2) generating summaries from one-sided transcripts. Evaluating models on MultiWOZ, DailyDialog, SpokenWOZ and Candor with both human A/B testing and LLM-as-a-judge metrics, we find that additional context improves reconstruction, and while large models generate promising reconstructions with prompting, smaller models require finetuning. Further, high-quality summaries can be generated without reconstructing missing turns. We present 1SC as a novel challenge and report promising results that mark a step toward privacy-aware conversational AI.
Uncovering hidden symbolic laws from time series data, as an aspiration dating back to Kepler’s discovery of planetary motion, remains a core challenge in scientific discovery and artificial intelligence. While Large Language Models show promise in structured reasoning tasks, their ability to infer interpretable, context-aligned symbolic structures from time series data is still underexplored. To systematically evaluate this capability, we introduce SymbolBench, a comprehensive benchmark designed to assess symbolic reasoning over real-world time series across three tasks: multivariate symbolic regression, Boolean network inference, and causal discovery. Unlike prior efforts limited to simple algebraic equations, SymbolBench spans a diverse set of symbolic forms with varying complexity. We further propose a unified framework that integrates LLMs with genetic programming to form a closed-loop symbolic reasoning system. Our empirical results reveal key strengths and limitations of current models, highlighting the importance of combining domain knowledge, context alignment, and reasoning structure to improve LLMs in automated scientific discovery.
Multimodal large language models (MLLMs) have recently achieved remarkable progress in radiology by integrating visual perception with natural language understanding. However, they often generate clinically unsupported descriptions, known as medical hallucinations, which pose serious risks in medical applications that demand accuracy and image-grounded outputs. Through empirical analysis, we find that prompt-induced hallucinations remain prevalent in radiology MLLMs, largely due to over-sensitivity to clinical sections. To address this, we introduce **C**linical **C**ontrastive **D**ecoding (**CCD**), a ***training-free*** and ***retrieval-free*** inference framework that integrates structured clinical signals from task-specific radiology expert models. CCD introduces a dual-stage contrastive mechanism to refine token-level logits during generation, thereby enhancing clinical fidelity without modifying the base MLLM. Experiments on three datasets and multiple models demonstrate that CCD consistently improves overall performance on radiology report generation (RRG). On the MIMIC-CXR dataset, it yields up to a 17% improvement in RadGraph-F1 when applied to state-of-the-art RRG models. Our approach provides a lightweight and generalisable solution for mitigating medical hallucinations, effectively bridging expert models and MLLMs in radiology.
While scaling test-time compute can substantially improve model performance, existing approaches either rely on static compute allocation or sample from fixed generation distributions.In this work, we introduce a test-time compute allocation framework that jointly adapts where computation is spent and how generation is performed. Our method begins with a warm-up phase that identifies easy queries and assembles an initial pool of question-response pairs from the test set itself. An adaptive phase then concentrates further computation on unresolved queries while reshaping their generation distributions through evolving in-context demonstrations—conditioning each generation on successful responses from semantically related queries rather than resampling from a fixed distribution.Experiments across math, coding, and reasoning benchmarks demonstrate that our approach consistently outperforms existing baselines while consuming substantially less inference-time compute.
Topic modeling seeks to uncover latent semantic structure in text corpora with minimal supervision. Neural approaches achieve strong performance but require extensive tuning and struggle with lifelong learning due to catastrophic forgetting and fixed capacity, while classical probabilistic models lack flexibility and adaptability to streaming data. We introduce CobwebTM, a low-parameter lifelong hierarchical topic model based on incremental probabilistic concept formation. By adapting the Cobweb algorithm to continuous document embeddings, CobwebTM constructs semantic hierarchies online, enabling unsupervised topic discovery, dynamic topic creation, and hierarchical organization without predefining the number of topics. Across diverse datasets, CobwebTM achieves strong topic coherence, stable topics over time, and high-quality hierarchies, demonstrating that incremental symbolic concept formation combined with pretrained representations is an efficient approach to topic modeling.
Large Language Models (LLMs) have achieved remarkable progress through Reinforcement Learning with Verifiable Rewards (RLVR), yet still rely heavily on external supervision (e.g., curated labels). Adversarial learning, particularly through self-play, offers a promising alternative that enables models to learn from themselves—thus reducing reliance on external supervision. Dual-play extends adversarial learning by assigning specialized roles to two models and training them against each other, fostering sustained competition and mutual evolution. Despite its promise, adapting dual-play training to LLMs remains limited. In this paper, we introduce PasoDoble, a novel LLM dual-play framework. PasoDoble adversarially trains two models initialized from the same base model: a Proposer, which generates challenging questions with ground-truth answers, and a Solver, which attempts to solve them. We enrich the Proposer with knowledge from a pre-training dataset to ensure the questions’ quality and diversity. To avoid reward hacking, the Proposer is rewarded for producing only valid questions that push the Solver’s limit, while the Solver is rewarded for solving them correctly, and both are updated jointly. Experimental results show that PasoDoble can improve the math reasoning performance of LLMs.
EMSDialog: Synthetic Multi-person Emergency Medical Service Dialogue Generation from Electronic Patient Care Reports via Multi-LLM Agents
PDF ↗Conversational diagnosis prediction requires models to track evolving evidence in streaming clinical conversations and decide when to commit to a diagnosis. Existing medical dialogue corpora are largely dyadic or lack the multi-party workflow and annotations needed for this setting. We introduce an ePCR-grounded, topic-flow-based multi-agent generation pipeline that iteratively plans, generates, and self-refines dialogues with rule-based factual and topic flow checks. The pipeline yields EMSDialog, a dataset of 4,414 synthetic multi-speaker EMS conversations based on a real-world ePCR dataset, annotated with 43 diagnoses, speaker roles, and turn-level topics. Human and LLM evaluations confirm high quality and realism of EMSDialog using both utterance- and conversation-level metrics. Results show that EMSDialog-augmented training improves accuracy, timeliness, and stability of EMS conversational diagnosis prediction. Our datasets and code are publicly available at https://uva-dsa.github.io/EMSDialog
Although large language models (LLMs) excel at factual recall, they can still propagate stale or incorrect knowledge, making in-context knowledge editing a gradient-free remedy suitable for black-box APIs. These knowledge editors that use in-context learning typically rely on a single retriever and surface-similarity heuristics to build prompts. However, a key observation in this study is that retrievers can be complementary: semantic rankers may recover paraphrased evidence, while lexical or feature-based retrievers may preserve precise entities and cues. This creates two gaps in single-retriever editors: they (i) miss complementary evidence that different retrievers surface and (ii) cannot adapt when one retriever is clearly more reliable for a query. We introduce a Feature-Weighted Ensemble for In-context Knowledge Editing (FWE-IKE) that calibrates three heterogeneous rankers (LLM-, BERT-, and MLP-based), extracts simple confidence features from each ranker, predicts per-query mixture weights, and applies a conservative margin-based routing gate that selects a single expert when confident; otherwise we mix calibrated distributions with learned per-query weights. On the CounterFact benchmark, FWE-IKE attains 88.33% Edit-Success Rate, a +3.0 point gain over the best single retriever and approaching the oracle upper bound (91%). Case studies, an ablation study, and analyses show the method systematically recovers complementary wins (e.g., BERT-only, LLM-only, MLP-only slices). FWE-IKE improves edit accuracy without touching model weights and provides a practical path to more robust, confidence-aware retrieval for IKE.
Existing work on value alignment typically characterizes value relations statically, ignoring how alignment interventions—such as prompting, fine-tuning, or preference optimization—reshape the broader value system. In practice, aligning a target value can implicitly shift other values, creating value trade-offs that remain largely unmeasured.We introduce the VAT, a framework that quantifies value trade-offs by measuring how alignment-induced changes propagate across interconnected values relative to achieved on-target gain. VAT captures the system-level dynamics of value expression under alignment intervention, enabling evaluation of both intended improvements and unintended side effects.Using a controlled scenario–action dataset grounded in Schwartz value theory, we collect paired pre–post normative judgments and analyze alignment effects across models, values, and interventions. Results show that alignment often produces uneven and structured co-movement among values, revealing systematic trade-offs between target and non-target values. These effects are largely invisible under conventional target-only evaluation, but become evident via VAT, highlighting process-level alignment risks and offering new insights into the dynamic nature of value alignment in LLMs.Dataset and code are open-sourced.
Large language models (LLMs) have demonstrated impressive reasoning capabilities, yet they often struggle when dealing with complex, ill-formed, or noisy inputs that frequently occur in interactions with real users. LLMs typically lack crucial refining capabilities needed to filter out irrelevant details, restructure key points before reasoning over the text and responding, resulting in suboptimal performance and incorrect answers. From an information theory perspective, this behavior is akin to decoding a high-entropy problem without first reducing its entropy. In this work, we first introduce GSM-Noise, a benchmark featuring grade-school math problems systematically perturbed to reflect real-world input variability. We show that the reasoning ability of open-source models (e.g., LLaMA and Qwen series) can be compromised by noise, while closed-source models are more robust. To improve LLM robustness under noisy conditions, we propose that LLMs first refine inputs — thereby reducing their entropy — before engaging in in-depth analysis. We investigate three approaches to instill this refinement capability: prompt engineering (PE), supervised finetuning (SFT), and reinforcement learning (RL). Experimental results show that input refinement leads to consistent performance gains: 2–12% with PE, 4–13% with SFT, and 3–25% with RL. These results highlight the importance of incorporating an explicit refinement phase to enhance the robustness and reliability of LLM reasoning in real-world scenarios.
Text revision is a core process in document creation, capturing how authors iteratively refine, reorganize, and improve written content. With the increasing availability of large-scale revision histories from platforms such as Wikipedia and arXiv, NLP research has begun to move beyond modeling what changes are made to understanding why they are made, i.e., the underlying edit intentions. To our knowledge, this is the first survey that synthesizes text revision research through the lens of edit intentions, providing a unified view of datasets, taxonomies, identification methods, and applications. We review prior work across the full revision workflow, including revision corpus construction, edit intention taxonomy design, and edit intention identification. We further categorize representative datasets and methods, summarize downstream applications such as writing assistance and document edit summarization, and highlight key open research directions.
Automatic evaluators such as reward models play a central role in the alignment and evaluation of large vision–language models (LVLMs). Despite their growing importance, these evaluators are almost exclusively assessed on English-centric benchmarks, leaving open the question of how well these evaluators generalize across languages. To answer this question, we introduce MM-JudgeBench, the first large-scale benchmark for multilingual and multimodal judge model evaluation, which includes over 60K pairwise preference instances spanning 25 typologically diverse languages. MM-JudgeBench integrates two complementary subsets: a general vision–language preference evaluation subset extending VL-RewardBench, and a chart-centric visual–text reasoning subset derived from OpenCQA, enabling systematic analysis of reward models (i.e., LVLM judges) across diverse settings. We additionally release a multilingual training set derived from MM-RewardBench, disjoint from our evaluation data, to support domain adaptation. By evaluating 22 LVLMs (15 open-source, 7 proprietary), we uncover substantial cross-lingual performance variance in our proposed benchmark. Our analysis further shows that model size and architecture are poor predictors of multilingual robustness, and that even state-of-the-art LVLM judges exhibit inconsistent behavior across languages. Together, these findings expose fundamental limitations of current reward modeling and underscore the necessity of multilingual, multimodal benchmarks for developing reliable automated evaluators.
We introduce TableVista, a comprehensive benchmark for evaluating foundation models in multimodal table reasoning under visual and structural complexity. TableVista consists of 3,000 high-quality table reasoning problems, where each instance is expanded into 10 distinct visual variants through our multi-style rendering and transformation pipeline. This process encompasses diverse scenario styles, robustness perturbations, and vision-only configurations, culminating in 30,000 multimodal samples for a multi-dimensional evaluation. We conduct an extensive evaluation of 29 state-of-the-art open-source and proprietary foundation models on TableVista. Through comprehensive quantitative and qualitative analysis, we find that while evaluated models remain largely stable across diverse rendering styles, they exhibit pronounced performance degradation on complex structural layouts and vision-only settings, revealing that current models struggle to maintain reasoning consistency when structural complexity combines with visually integrated presentations. These findings highlight critical gaps in current multimodal capabilities, providing insights for advancing more robust and reliable table understanding models.
Despite demonstrating remarkable performance across a wide range of tasks, large language models (LLMs) have also been found to frequently produce outputs that are incomplete or selectively omit key information. In sensitive domains, such omissions can result in significant harm comparable to that posed by factual inaccuracies, including hallucinations. In this study, we address the challenge of evaluating the comprehensiveness of LLM-generated texts, focusing on the detection of missing information or underrepresented viewpoints. We investigate three automated evaluation metrics: (1) an NLI-based method that decomposes texts into atomic statements and uses natural language inference (NLI) to identify missing facts, (2) a Q A-based metric that extracts question-answer pairs and compares responses across sources, and (3) an end-to-end approach that directly identifies missing content using LLMs. Our experiments demonstrate the surprising effectiveness of the simple end-to-end metric compared to more complex metrics, though at the cost of reduced robustness, interpretability and result granularity. We further assess the comprehensiveness of responses from several popular open-weight LLMs when answering user queries based on multiple sources.