Over the years, scalar MT metrics have advanced rapidly on benchmarks. Yet they remain black boxes, offering little insight into their decisions and sometimes degrading under out-of-distribution inputs. We introduce Remedy-R, a reasoning-driven generative MT metric trained with reinforcement learning from pairwise translation preferences, without requiring error-span annotations or distillation from closed LLMs. Unlike scalar MT metrics that only outputs translation quality scores, Remedy-R produces step-by-step analyses of accuracy, fluency, and completeness, enabling more interpretable assessments. With only 60K pairwise training samples across two language pairs, Remedy-R remains competitive with top scalar metrics and GPT-4-based judges on WMT22–24 metric benchmarks, generalizes to other languages, and shows strong robustness on OOD stress tests. Moreover, Remedy-R generates self-reflective feedback that can be reused for translation refinement. We validate the faithfulness of such feedback with GPT-4 and show that a simple evaluate–revise pipeline leveraging Remedy-R’s analyses consistently improves translation quality across diverse models without any task-specific tuning.
论文检索
输入标题、作者或关键词，从 7,537 篇学术成果中精准定位
Recent advances in AI and wearable devices, such as augmented-reality glasses, have made it possible to augment human memory by retrieving personal experiences in response to natural language queries. However, existing egocentric video datasets fall short in supporting the personalization and long-context reasoning required for episodic memory retrieval. To address these limitations, we introduce EgoMemory, a benchmark derived from Ego4D, enriched with 165,795 user-specific object annotations over 245 videos from 45 participants, yielding 639 distinct, human-curated, and evaluated queries for rich and individualized episodic memory retrieval. Leveraging this resource, we present EgoRetriever, a novel, training-free retrieval framework that combines Multimodal Large Language Models with reflective Chain-of-Thought prompting. Our approach enables interpretive inference of user intent and generates detailed target video descriptions by leveraging contextualized personal memory for video retrieval. Extensive experiments on three benchmarks, including EgoMemory, EgoCVR, and EgoLife, demonstrate that EgoRetriever consistently and substantially outperforms state-of-the-art baselines, highlighting its strong generalizability and practical potential for personalized, long-context egocentric video retrieval.
Understanding New-Knowledge-Induced Factual Hallucinations in LLMs: Analysis and Interpretation
PDF ↗Prior works have shown that fine-tuning on new knowledge can induce factual hallucinations in large language models (LLMs), leading to incorrect outputs when evaluated on previously known information. However, the specific manifestations of such hallucination and its underlying mechanisms remain insufficiently understood. Our work addresses this gap by designing a controlled dataset Biography-Reasoning, and conducting a fine-grained analysis across multiple knowledge types and two task types, including knowledge question answering (QA) and knowledge reasoning tasks. We find that hallucinations not only severely affect tasks involving newly introduced knowledge, but also propagate to other evaluation tasks. Moreover, when fine-tuning on a dataset in which a specific knowledge type consists entirely of new knowledge, LLMs exhibit elevated hallucination tendencies. This suggests that the degree of unfamiliarity within a particular knowledge type, rather than the overall proportion of new knowledge, is a stronger driver of hallucinations. Through interpretability analysis, we show that learning new knowledge weakens the model’s attention to key entities in the input question, leading to an over-reliance on surrounding context and a higher risk of hallucination. Conversely, reintroducing a small amount of known knowledge during the later stages of training restores attention to key entities and substantially mitigates hallucination behavior. Finally, we demonstrate that disrupted attention patterns can propagate across lexically similar contexts, facilitating the spread of hallucinations beyond the original task.
Curriculum Learning based Hierarchical Scoring and Analysis Framework for Question Answering Task Evaluation
PDF ↗The rapid progress of large language models (LLMs) has increased the demand for efficient and reliable evaluation of question answering (QA) systems. Existing evaluation methods either rely on rule-based matching with shallow semantic understanding or adopt LLM-as-a-Judge approaches that incur high cost and latency while offering limited error interpretability. Accordingly, we propose HiEval, a curriculum learning based hierarchical framework for QA task evaluation that supports both quick scoring and fine-grained error analysis. HiEval contains a quick scoring model (HiEval-QS) that predicts three-level correctness labels, and an error analysis model (HiEval-EA) that identifies incorrect responses into five error types. HiEval incorporates a class-balanced focal loss to handle label imbalance, experience replay to prevent forgetting, and contrastive unlikelihood optimization to improve error discrimination. We also construct two large-scale human-annotated evaluation datasets collected from 50 QA-related datasets, covering 8 task types and release two challenging benchmarks. Extensive experiments show that HiEval achieves state-of-the-art performance on both quick scoring and error analysis tasks, outperforming all baseline methods, including GPT-5, while being approximately 25\times faster.
Multimodal Large Language Models (MLLMs) excel at structural reasoning yet suffer from a sharp logical brittleness in structural consistency. We term this phenomenon Structural Cognitive Overload (SCO), a byproduct of the contention between deep reasoning and safety alignment. However, prior work has predominantly targeted typographic and pixel-level perturbations, leaving the study of SCO largely unexplored. To this end, we propose StructBreak, an automated end-to-end framework designed to quantify SCO. By leveraging StructBreak, we uncover a novel higher-order cognitive overload attack paradigm; notably, this attack operates under a practical black-box setting, requiring no internal model access. Consequently, we utilize this framework to establish a comprehensive benchmark spanning ten diverse threat scenarios. Empirical evaluations on six leading MLLMs reveal that SCO readily triggers toxic generation, yielding a 92% average ASR (up to 97% on Gemini 2.5). To elucidate the mechanism of SCO, we further conduct model-level interpretations spanning attention dynamics, latent space topology, and geometric analysis. Our findings reveal that StructBreak acts as a novel structural channel to circumvent safety filters. Furthermore, the limited efficacy of inherent safety mechanisms underscores that current alignment paradigms are insufficient for the era of complex multimodal reasoning.
Large Language Models (LLMs) are increasingly deployed worldwide, yet they exhibit strong Western-centric biases, and the internal mechanisms governing their cultural behaviors remain poorly understood. Prior work has identified so-called cultural neurons, but individual neurons are often polysemous, conflating abstract cultural knowledge with surface-level lexical cues due to superposition. We apply Sparse Autoencoders (SAEs) to decompose intermediate LLM activations into sparse, interpretable feature representations that disentangle these factors. This analysis reveals culturally selective features that remain invariant across paraphrasing and task formats, indicating abstraction beyond lexical correlations. Through targeted feature ablation, we provide causal evidence that these features are necessary for cultural reasoning: their removal selectively degrades performance on culturally conditioned tasks. Furthermore, we show that steering model activations along these feature directions is sufficient to systematically modulate cultural-related knowledge generation, without retraining. Together, our results offer the first causal evidence that LLMs encode cultural knowledge as decoupled semantic structures rather than surface patterns, enabling a scalable pathway toward cultural alignment through mechanistic intervention. Code is available at https://github.com/IAN-YE/Cultural-features-SAE.
K-MetBench: A Multi-Dimensional Benchmark for Fine-Grained Evaluation of Expert Reasoning, Locality, and Multimodality in Meteorology
PDF ↗The development of practical (multimodal) large language model assistants for Korean weather forecasters is hindered by the absence of a multidimensional, expert-level evaluation framework grounded in authoritative sources. To address this, we introduce K-MetBench, a diagnostic benchmark grounded in national qualification exams. It exposes critical gaps across four dimensions: expert visual reasoning of charts, logical validity via expert-verified rationales, Korean-specific geo-cultural comprehension, and fine-grained domain analysis. Our evaluation of 55 models reveals a profound *modality gap* in interpreting specialized diagrams and a *reasoning gap* where models hallucinate logic despite correct predictions. Crucially, Korean models outperform significantly larger global models in local contexts, demonstrating that parameter scaling alone cannot resolve cultural dependencies. K-MetBench serves as a roadmap for developing reliable, culturally aware expert AI agents.
Large language models (LLMs) demonstrate strong performance, but they often lack transparency. We introduce GeoLAN, a training framework that treats token representations as geometric trajectories and applies stickiness conditions inspired by recent developments related to the Kakeya Conjecture. We have developed two differentiable regularizers, Katz-Tao Convex Wolff (KT-CW) and Katz-Tao Attention (KT-Attn), that promote isotropy and encourage diverse attention. Our experiments with Gemma-3 (1B, 4B, 12B) and Llama-3-8B show that GeoLAN frequently maintains task accuracy while improving geometric metrics and reducing certain fairness biases. These benefits are most significant in mid-sized models. Our findings reveal scale-dependent trade-offs between geometric precision and performance, suggesting that geometry-aware training is a promising approach to enhance mechanistic interpretability.
Next location prediction aims to infer the next location users are likely to visit based on their historical check-in data. However, existing methods assume that check-in data is complete, overlooking the subjective nature of users’ check-in behavior, leading to inaccurate capture of user preferences. Recently, Large Language Models (LLMs) have offered a promising approach to location completion due to their extensive world knowledge. Nevertheless, our experiments reveal that LLMs struggle to interpret raw geographic coordinate information. To address these challenges, we propose LaMDA, an LLM-driven Multi-perspective Data Augmentation framework that employs dual completion agents to complement user mobility behaviors. Driven by our empirical findings that natural language descriptions align more closely with real-world geographic logic, LaMDA translates geographic coordinates into text to enhance spatial reasoning. Leveraging these semantic descriptions, LaMDA constructs dual agents to complement user mobility: "Micro-Level Completion" fills short-term omissions, while "Macro-Level Completion" infers unrecorded locations based on periodic preferences. Reliability is ensured through tailored real-world point-of-interest (POI) pools and a self-verification mechanism. Finally, a collaborative dual-graph module leverages this augmented data for fine-grained preference modeling. Extensive experiments on three real-world datasets demonstrate that LaMDA significantly outperforms state-of-the-art methods.
MCLE-Mol: Empowering LLM with Molecular Comprehension and Low-Cost Continual Evolution for Interpretable Property Prediction
PDF ↗Large language models (LLMs) offer a new paradigm for molecular property prediction (MPP), yet a semantic gap between natural language and molecular representations limits LLMs’ ability to capture structure–activity relationships (SAR). Recent approaches have explored injecting structure-level information into LLMs, primarily modeling associations based on statistical regularities. However, these methods are prone to misinterpreting coincidental associations as general principles, imposing a bottleneck on predictive performance. To tackle the challenges above, we propose MCLE-Mol, an ML–LLM–Rule collaborative framework for MPP. It bridges the semantic gap by injecting ML-derived substructure attribution values into LLMs, utilizing Context-Calibrated Substructure Attribution Rules (CCSAR) to calibrate these attributions under specific chemical contexts to mitigate such misinterpretation. In addition, MCLE-Mol introduces a low-cost continual evolution strategy that updates CCSAR with frozen model parameters to adapt to dynamic chemical spaces. Experiments on multiple benchmark datasets demonstrate that MCLE-Mol outperforms all baselines, successfully resolving the trade-off between predictive performance and interpretability.
RAG-KT: Cross-platform Explainable Knowledge Tracing with Multi-view Fusion Retrieval Generation
PDF ↗Knowledge Tracing (KT) infers a student’s knowledge state from past interactions to predict future performance. Conventional Deep Learning (DL)-based KT models are typically tied to platform-specific identifiers and latent representations, making them hard to transfer and interpret. Large Language Model (LLM)-based methods can be either ungrounded under prompting or overly domain-dependent under fine-tuning. In addition, most existing KT methods are developed and evaluated under a same-distribution assumption. In real deployments, educational data often arise from heterogeneous platforms with substantial distribution shift, which often degrades generalization. To this end, we propose RAG-KT, a retrieval-augmented paradigm that frames cross-platform KT as reliable context constrained inference with LLMs. It builds a unified multi-source structured context with cross-source alignment via Question Group abstractions and retrieves complementary rich and reliable context for each prediction, enabling grounded prediction and interpretable diagnosis. Experiments on three public KT benchmarks demonstrate consistent gains in accuracy and robustness, including strong performance under cross-platform conditions.
How Chain-of-Thought Works? Tracing Information Flow from Decoding, Projection, and Activation
PDF ↗Chain-of-Thought (CoT) prompting significantly enhances model reasoning, yet its internal mechanisms remain poorly understood. We analyze CoT’s operational principles by reversely tracing information flow across decoding, projection, and activation phases. Our quantitative analysis suggests that CoT may serve as a decoding space pruner, leveraging answer templates to guide output generation, with higher template adherence strongly correlating with improved performance. Furthermore, we surprisingly find that CoT modulates neuron engagement in a task-dependent manner: reducing neuron activation in open-domain tasks, yet increasing it in closed-domain scenarios. These findings offer a novel mechanistic interpretability framework and critical insights for the NLP community, enabling targeted CoT interventions to design more efficient and robust prompts. We released our code and data at https://github.com/How-Young-X/cot
Large Language Models (LLMs) have shown promise in simulating human behavior, yet existing agents often exhibit behavioral rigidity, a flaw frequently masked by the self-referential bias of current "LLM-as-a-judge" evaluations. By evaluating against empirical ground truth, we reveal a counter-intuitive phenomenon: increasing the intensity of prompt-driven reasoning does not enhance fidelity but rather exacerbates value polarization, collapsing population diversity. To address this, we propose the Context-Value-Action (CVA) architecture, grounded in the Stimulus-Organism-Response (S-O-R) model and Schwartz’s Theory of Basic Human Values. Unlike methods relying on self-verification, CVA decouples action generation from cognitive reasoning via a novel Value Verifier trained on authentic human data to explicitly model dynamic value activation. Experiments on CVABench, which comprises over 1.1 million real-world interaction traces, demonstrate that CVA significantly outperforms baselines. Our approach effectively mitigates polarization while offering superior behavioral fidelity and interpretability.
Once Correct, Still Wrong: Counterfactual Hallucination in Multilingual Vision-Language Models
PDF ↗Vision–language models (VLMs) can achieve high accuracy while still accepting **culturally plausible but visually incorrect** interpretations. Existing hallucination benchmarks rarely test this failure mode, particularly outside Western contexts and English. We introduce **M^2CQA**, a culturally grounded multimodal benchmark built from images spanning 17 MENA countries, paired with contrastive true and counterfactual statements in English, Arabic, and its dialects. To isolate hallucination beyond raw accuracy, we propose the **CounterFactual Hallucination Rate (CFHR)**, which measures counterfactual acceptance conditioned on correctly answering the true statement. Evaluating state-of-the-art VLMs under multiple prompting strategies, we find that CFHR rises sharply in Arabic, especially in dialects, even when true-statement accuracy remains high.Moreover, reasoning-first prompting consistently increases counterfactual hallucination, while answering before justifying improves robustness. We make the dataset publicly available for the community (https://huggingface.co/datasets/QCRI/M2CQA)).
Recent work on domain-specific reasoning with large language models (LLMs) has largely relied on training-intensive approaches that require updating model parameters. Although activation steering has emerged as a parameter-efficient alternative, existing methods typically rely on static and manually designed interventions, limiting their ability to adapt to the dynamic nature of complex reasoning. To address this limitation, we propose RISER (Router-based Intervention for Steerable Enhancement of Reasoning), a plug-and-play intervention framework that adaptively steers LLM reasoning in activation space. RISER builds a library of reusable reasoning vectors and employs a lightweight Router to dynamically compose these vectors for each input. The Router is optimized via reinforcement learning under task-level rewards, enabling the emergent and compositional activation of latent cognitive primitives. Across seven diverse benchmarks, RISER achieves average zero-shot accuracy improvements of 3.4–6.5% over the base model, while outperforming chain-of-thought-style reasoning with 2–3× higher token efficiency and robust accuracy gains. Further analysis demonstrates that RISER autonomously combines multiple vectors into interpretable and precise control strategies, pointing toward more controllable and efficient LLM reasoning.
We introduce AVACraft — the first multimodal benchmark environment for complex decision-making in StarCraft II, supporting both traditional Multi-Agent Reinforcement Learning (MARL) and modern Vision-Language Model (VLM) paradigms. Existing StarCraft II environments like SMAC rely on abstract state representations that deviate from human perception and lack support for emerging VLM-based decision-making. AVACraft mitigates these limitations via a unified framework, which provides RGB visual inputs, natural language observations and structured state information, enabling systematic comparisons between training-based and zero-shot decision-making methods. Our benchmark features 21 carefully designed scenarios covering micromanagement, coordination and strategic planning, with standardized evaluation protocols for both paradigms. We establish comprehensive baselines using four MARL algorithms (IQL, QMIX, QTRAN, VDN) and multiple state-of-the-art VLMs (GPT-4o, Qwen-VL, etc.). Experimental results reveal their complementary strengths: MARL methods achieve up to 27.1% win rate after 1M training steps in complex scenarios, while VLMs deliver superior zero-shot performance (75–81% win rate) and human-aligned decision processes without any training. Systematic analysis (including expert human evaluation) also identifies key trade-offs between training efficiency, performance ceilings and interpretability across the two paradigms. Our implementation is available at https://anonymous.4open.science/r/VLM-Play-StarCraft2-70C4 .
Reinforcement learning (RL) has emerged as an effective paradigm for enhancing model reasoning. However, existing RL methods like GRPO often rely on unstructured self-sampling to fit scalar rewards, often producing inefficient rollouts that fail to capture transferable problem-solving strategies. To address these limitations, we propose **TemplateRL**, a structured template-guided RL framework that augments policy optimization with explicit template guidance. Our approach first constructs a problem-solving template library via MCTS on a small seed set, then seamlessly integrates this high-level structured guidance into RL training. By guiding rollout generation to align with proven template structures, TemplateRL significantly improves high-quality trajectory hit rates while reducing ineffective exploration. This structure-guided design steers the policy toward validated strategic patterns, stabilizing training dynamics, and enhancing RL sampling efficiency. Notably, the explicit template library is interpretable, editable, and supports online updates-enabling continuous updates during both training and inference. Extensive experiments demonstrate that TemplateRL outperforms GRPO by 99% on AIME and 41% on AMC, with superior stability on weak models and remarkable cross-domain generalization, highlighting its potential for broader tasks.
RiT: Rubrics-in-Thinking Reinforcement Learning for Improved Reasoning in Large Language Models
PDF ↗Large Reasoning Models (LRMs) benefit from generating intermediate reasoning steps, enabling more reliable and interpretable decision-making. While outcome-based supervision has proven effective for LRMs across diverse tasks, it focuses solely on final answers and cannot guarantee high-quality intermediate reasoning. In contrast, existing process supervision is largely limited to verifiable domains such as mathematics or code, where intermediate steps can be explicitly checked, restricting its applicability to open-ended reasoning tasks. To address these limitations, we propose Rubrics-in-Thinking Reinforcement Learning (RiT), the first framework to introduce thinking-rubric supervision into intermediate reasoning. RiT automatically generates fine-grained rubrics and integrates them into a reward function via gated fusion with outcome-based rewards, guiding models to reason in a coherent and task-aligned manner, improving both intermediate steps and the final response. Experiments on reasoning-intensive and open-ended benchmarks demonstrate that RiT consistently outperforms outcome-only RL baselines.
Reliable interpretation of multimodal dental data is essential for automated oral healthcare, yet current multimodal large language models (MLLMs) show limited understanding of dental images. Although complex reasoning improves performance, its gains in dentistry are substantially smaller than in other medical domains, suggesting that complex reasoning is not yet sufficiently incentivized for dental diagnosis, likely due to insufficient domain knowledge and limited reinforcement learning on dental questions. We present DentalGPT, a dentistry-specialized MLLM trained via staged multimodal alignment and reinforcement learning. By constructing the largest annotated multimodal dental dataset to date with over 120k images, multimodal alignment provides the necessary domain knowledge foundation to support and incentivize complex reasoning, which is further strengthened through reinforcement learning. Experiments on expert-annotated benchmarks and dental subsets of medical VQA benchmarks show that DentalGPT achieves superior performance on disease classification and dental VQA tasks, outperforming many state-of-the-art MLLMs despite its compact 7B parameter scale.
While LLMs demonstrate impressive reasoning capabilities, their internal decision dynamics remain opaque. To render these process interpretable and intervenable, we propose Dynamic Entropy Tracing, a mechanism-aware framework that interprets the evolving "choice state" of attention heads during CoT generation through stepwise head-wise option-logit and entropy tracing. Our analysis reveals distinct functional behaviors at attention heads: Steadfast Heads, characterized by consistently low entropy and producing a sharp, option-selective logit pattern with a stable top choice, and Wavering Heads, characterized by consistently high entropy and producing flat or oscillatory option logits without a persistent winner. Leveraging these traces, we identify a set of intervention targets and perform Selective Head Fine-Tuning, updating solely these selected heads against a frozen backbone. Experiments across the LLaMA and Qwen families reveal a striking plasticity hierarchy: fine-tuning just 30 Wavering Heads recovers over 98% of the performance achieved by full-parameter tuning, and in some settings modestly exceeds it. In contrast, intervening on Steadfast Heads yields much less gains. Our findings translate process-level mechanistic observables into a principled criterion for selective fine-tuning, offering a fundamental insight: the most effective tuning knobs are not the components that signal the final decision, but those that retain uncertainty, and thus plasticity, during its formation.