Recent research has shown that large language models (LLMs) can enhance translation quality through self-refinement. In this paper, we build on this idea by extending the refinement from sentence-level to document-level translation, specifically focusing on document-to-document (Doc2Doc) translation refinement. Since sentence-to-sentence (Sent2Sent) and Doc2Doc translation address different aspects of the translation process, we propose fine-tuning LLMs for translation refinement using two intermediate translations, combining the strengths of both Sent2Sent and Doc2Doc. Additionally, recognizing that the quality of intermediate translations varies, we introduce an enhanced fine-tuning method with quality awareness that assigns lower weights to easier translations and higher weights to more difficult ones, enabling the model to focus on challenging translation cases. Experimental results across ten translation tasks with LLaMA-3-8B-Instruct and Mistral-Nemo-Instruct demonstrate the effectiveness of our approach.

Extractive reading comprehension systems are designed to locate the correct answer to a question within a given text. However, a persistent challenge lies in ensuring these models maintain high accuracy in answering questions while reliably recognizing unanswerable queries. Despite significant advances in large language models (LLMs) for reading comprehension, this issue remains critical, particularly as the length of supported contexts continues to expand. To address this challenge, we propose an innovative data augmentation methodology grounded in a multi-agent collaborative framework. Unlike traditional methods, such as the costly human annotation process required for datasets like SQuAD 2.0, our method autonomously generates evidence-based question-answer pairs and systematically constructs unanswerable questions. Using this methodology, we developed the FactGuard-Bench dataset, which comprises 25,220 examples of both answerable and unanswerable question scenarios, with context lengths ranging from 8K to 128K. Experimental evaluations conducted on seven popular LLMs reveal that even the most advanced models achieve only 61.79% overall accuracy. Furthermore, we emphasize the importance of a model's ability to reason about unanswerable questions to avoid generating plausible but incorrect answers. By implementing efficient data selection and generation within the multi-agent collaborative framework, our method significantly reduces the traditionally high costs associated with manual annotation and provides valuable insights for the training and optimization of LLMs.

Chain-of-Thought (CoT) enhances an LLM's ability to perform complex reasoning tasks, but it also introduces new security issues. In this work, we present ShadowCoT, a novel backdoor attack framework that targets the internal reasoning mechanism of LLMs. Unlike prior token-level or prompt-based attacks, ShadowCoT directly manipulates the model's cognitive reasoning path, enabling it to hijack multi-step reasoning chains and produce logically coherent but adversarial outcomes. By conditioning on internal reasoning states, ShadowCoT learns to recognize and selectively disrupt key reasoning steps, effectively mounting a self-reflective cognitive attack within the target model. Our approach introduces a lightweight yet effective multi-stage injection pipeline, which selectively rewires attention pathways and perturbs intermediate representations with minimal parameter overhead (only 0.15% updated). ShadowCoT further leverages reinforcement learning and reasoning chain pollution (RCP) to autonomously synthesize stealthy adversarial CoTs that remain undetectable to advanced defenses. Extensive experiments across diverse reasoning benchmarks and LLMs show that ShadowCoT consistently achieves high Attack Success Rate (94.4%) and Hijacking Success Rate (88.4%) while preserving benign performance. These results reveal an emergent class of cognition-level threats and highlight the urgent need for defenses beyond shallow surface-level consistency.

Medical Visual Question Answering (VQA) enhances clinical decision-making by enabling systems to interpret medical images and answer clinical queries. However, developing efficient, high-performance VQA models is challenging due to the complexity of medical imagery and diverse modalities. In this paper, we introduce a lightweight, multimodal VQA model integrating BiomedCLIP for image feature extraction and LLaMA-3 for text processing. Designed for medical VQA tasks, our model achieves state-of-the-art performance on the OmniMedVQA dataset. With approximately 8 billion parameters, it requires only two NVIDIA 40 GB A100 GPUs, demonstrating superior efficiency over larger models. Our results show 73.4% accuracy for open-end questions, surpassing existing models and validating its potential for real-world medical applications. Key contributions include a specialized multimodal VQA model, a resource-efficient architecture, and strong performance in answering open-ended clinical questions.

While Large Language Models (LLMs) acquire vast knowledge during pre-training, they often lack domain-specific, new, or niche information. Continual pre-training (CPT) attempts to address this gap but suffers from catastrophic forgetting and inefficiencies in low-data regimes. We introduce Knowledge-Instruct, a novel approach to efficiently inject knowledge from limited corpora through pure instruction-tuning. By generating information-dense synthetic instruction data, it effectively integrates new knowledge while preserving general reasoning and instruction-following abilities. Knowledge-Instruct demonstrates superior factual memorization, minimizes catastrophic forgetting, and remains scalable by leveraging synthetic data from relatively small language models. Additionally, it enhances contextual understanding, including complex multi-hop reasoning, facilitating integration with retrieval systems. We validate its effectiveness across diverse benchmarks, including Companies, a new dataset that we release to measure knowledge injection capabilities.

Large Language Models (LLMs) hold promise as dynamic instructional aids. Yet, it remains unclear whether LLMs can replicate the adaptivity of intelligent tutoring systems (ITS)--where student knowledge and pedagogical strategies are explicitly modeled. We propose a prompt variation framework to assess LLM-generated instructional moves' adaptivity and pedagogical soundness across 75 real-world tutoring scenarios from an ITS. We systematically remove key context components (e.g., student errors and knowledge components) from prompts to create variations of each scenario. Three representative LLMs (Llama3-8B, Llama3-70B, and GPT-4o) generate 1,350 instructional moves. We use text embeddings and randomization tests to measure how the omission of each context feature impacts the LLMs' outputs (adaptivity) and a validated tutor-training classifier to evaluate response quality (pedagogical soundness). Surprisingly, even the best-performing model only marginally mimics the adaptivity of ITS. Specifically, Llama3-70B demonstrates statistically significant adaptivity to student errors. Although Llama3-8B's recommendations receive higher pedagogical soundness scores than the other models, it struggles with instruction-following behaviors, including output formatting. By contrast, GPT-4o reliably adheres to instructions but tends to provide overly direct feedback that diverges from effective tutoring, prompting learners with open-ended questions to gauge knowledge. Given these results, we discuss how current LLM-based tutoring is unlikely to produce learning benefits rivaling known-to-be-effective ITS tutoring. Through our open-source benchmarking code, we contribute a reproducible method for evaluating LLMs' instructional adaptivity and fidelity.

Aligning large language models (LLMs) with human preferences has achieved remarkable success. However, existing Chinese preference datasets are limited by small scale, narrow domain coverage, and lack of rigorous data validation. Additionally, the reliance on human annotators for instruction and response labeling significantly constrains the scalability of human preference datasets. To address these challenges, we design an LLM-based Chinese preference dataset annotation pipeline with no human intervention. Specifically, we crawled and carefully filtered 92k high-quality Chinese queries and employed 15 mainstream LLMs to generate and score chosen-rejected response pairs. Based on it, we introduce COIG-P (Chinese Open Instruction Generalist - Preference), a high-quality, large-scale Chinese preference dataset, comprises 1,009k Chinese preference pairs spanning 6 diverse domains: Chat, Code, Math, Logic, Novel, and Role. Building upon COIG-P, to reduce the overhead of using LLMs for scoring, we trained a 8B-sized Chinese Reward Model (CRM) and meticulously constructed a Chinese Reward Benchmark (CRBench). Evaluation results based on AlignBench \citep{liu2024alignbenchbenchmarkingchinesealignment} show that that COIG-P significantly outperforms other Chinese preference datasets, and it brings significant performance improvements ranging from 2% to 12% for the Qwen2/2.5 and Infinity-Instruct-3M-0625 model series, respectively. The results on CRBench demonstrate that our CRM has a strong and robust scoring ability. We apply it to filter chosen-rejected response pairs in a test split of COIG-P, and our experiments show that it is comparable to GPT-4o in identifying low-quality samples while maintaining efficiency and cost-effectiveness. Our codes and data are released in https://github.com/multimodal-art-projection/COIG-P.

We assess how the code reasoning abilities of large language models (LLMs) generalize to different kinds of programs. We present techniques for obtaining in- and out-of-distribution programs with different characteristics: code sampled from a domain-specific language, code automatically generated by an LLM, code collected from competitive programming contests, and mutated versions of these programs. We also present an experimental methodology for evaluating LLM generalization by comparing their performance on these programs. We perform an extensive evaluation across 10 state-of-the-art models from the past year, obtaining insights into their generalization capabilities over time and across different classes of programs. Our results highlight that while earlier models exhibit behavior consistent with pattern matching, the latest models exhibit strong generalization abilities on code reasoning.

Hypothesis generation is a fundamental step in scientific discovery, yet it is increasingly challenged by information overload and disciplinary fragmentation. Recent advances in Large Language Models (LLMs) have sparked growing interest in their potential to enhance and automate this process. This paper presents a comprehensive survey of hypothesis generation with LLMs by (i) reviewing existing methods, from simple prompting techniques to more complex frameworks, and proposing a taxonomy that categorizes these approaches; (ii) analyzing techniques for improving hypothesis quality, such as novelty boosting and structured reasoning; (iii) providing an overview of evaluation strategies; and (iv) discussing key challenges and future directions, including multimodal integration and human-AI collaboration. Our survey aims to serve as a reference for researchers exploring LLMs for hypothesis generation.

Service and assistive robots are increasingly being deployed in dynamic social environments; however, ensuring transparent and explainable interactions remains a significant challenge. This paper presents a multimodal explainability module that integrates vision language models and heat maps to improve transparency during navigation. The proposed system enables robots to perceive, analyze, and articulate their observations through natural language summaries. User studies (n=30) showed a preference of majority for real-time explanations, indicating improved trust and understanding. Our experiments were validated through confusion matrix analysis to assess the level of agreement with human expectations. Our experimental and simulation results emphasize the effectiveness of explainability in autonomous navigation, enhancing trust and interpretability.