Test-time compute has led to remarkable success in the large language model (LLM) community, particularly for complex tasks, where longer chains of thought (CoTs) are generated to enhance reasoning capabilities. However, growing evidence reveals that such reasoning models often produce CoTs plagued by excessive redundancy, including repetitive verification steps and unnecessary reasoning shifts. The root cause lies in post-training of them that overly rely on outcome reward paradigms, as the data of process reward paradigms, which regulate intermediate reasoning steps, is difficult to construct at scale. To address this, we propose PI, a novel framework for Test-time Prompt Intervention. PI provides an interface to dynamically guide and regulate reasoning paths during inference through timely (When module) and proper (How module) interventions and post-intervention sampling (Which module). This allows human problem-solving expertise and cognitive science principles to be seamlessly integrated into LLMs’ reasoning processes, enhancing controllability and interpretability. Extensive experiments across multiple models and datasets demonstrate that PI significantly shortens CoTs while reducing hallucination, yielding more concise and reliable reasoning.
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
StepFun-Formalizer: Unlocking the Autoformalization Potential of LLMs Through Knowledge-Reasoning Fusion
PDF ↗Autoformalization aims to translate natural-language mathematical statements into a formal language. While LLMs have accelerated progress in this area, existing methods still suffer from low accuracy. We identify two key abilities for effective autoformalization: comprehensive mastery of formal-language domain knowledge, and reasoning capability of natural language problem understanding and informal-formal alignment. Without the former, a model cannot identify the correct formal objects; without the latter, it struggles to interpret real-world contexts and map them precisely into formal expressions. To address these gaps, we introduce ThinkingF, a data synthesis and training pipeline that improves both abilities. First, we construct two datasets: one by distilling and selecting large-scale examples rich in formal knowledge, and another by generating informal-to-formal reasoning trajectories guided by expert-designed templates. We then apply SFT and RLVR with these datasets to further fuse and refine the two abilities. The resulting 7B and 32B models exhibit both comprehensive formal knowledge and strong informal-to-formal reasoning. Notably, StepFun-Formalizer-32B achieves SOTA BEq@1 scores of 40.5% on FormalMATH-Lite and 26.7% on ProverBench, surpassing all prior general-purpose and specialized models.
Introducing Visual Scenes and Reasoning: A More Realistic Benchmark for Spoken Language Understanding
PDF ↗Spoken Language Understanding (SLU) consists of two sub-tasks: intent detection (ID) and slot filling (SF). Given its broad range of real-world applications, enhancing SLU for practical deployment is increasingly critical. Profile-based SLU addresses ambiguous user utterances by incorporating context awareness (CA), user profiles (UP), and knowledge graphs (KG) to support disambiguation, thereby advancing SLU research toward real-world applicability. However, existing SLU datasets still fall short in representing real-world scenarios. Specifically, (1) CA uses one-hot vectors for representation, which is overly idealized, and (2) models typically focuses solely on predicting intents and slot labels, neglecting the reasoning process that could enhance performance and interpretability. To overcome these limitations, we introduce VRSLU, a novel SLU dataset that integrates both Visual images and explicit Reasoning. For over-idealized CA, we use GPT-4o and FLUX.1-dev to generate images reflecting users’ environments and statuses, followed by human verification to ensure quality. For reasoning, GPT-4o is employed to generate explanations for predicted labels, which are then refined by human annotators to ensure accuracy and coherence. Additionally, we propose an instructional template, LR-Instruct, which first predicts labels and then generates corresponding reasoning. This two-step approach helps mitigate the influence of reasoning bias on label prediction. Experimental results confirm the effectiveness of incorporating visual information and highlight the promise of explicit reasoning in advancing SLU.
Recent advances in vision-language models have opened up new possibilities for reasoning-driven image geolocalization. However, existing approaches often rely on synthetic reasoning annotations or external image retrieval, which can limit interpretability and generalizability. In this paper, we present Geo-R, a retrieval-free framework that uncovers structured reasoning paths from existing ground-truth coordinates and optimizes geolocation accuracy via reinforcement learning. We propose the Chain of Region, a rule-based hierarchical reasoning paradigm that generates precise, interpretable supervision by mapping GPS coordinates to geographic entities (e.g., country, province, city) without relying on model-generated or synthetic labels. Building on this, we introduce a lightweight reinforcement learning strategy with coordinate-aligned rewards based on Haversine distance, enabling the model to refine predictions through spatially meaningful feedback. Our approach bridges structured geographic reasoning with direct spatial supervision, yielding improved localization accuracy, stronger generalization, and more transparent inference. Experimental results across multiple benchmarks confirm the effectiveness of Geo-R, establishing a new retrieval-free paradigm for scalable and interpretable image geolocalization. To facilitate further research and ensure reproducibility, both the model and code will be made publicly available.
Temporal knowledge graph question answering (TKGQA) involves multi-hop reasoning over temporally constrained entity relationships in the knowledge graph to answer a given question. However, at each hop, large language models (LLMs) retrieve subgraphs with numerous temporally similar and semantically complex relations, increasing the risk of suboptimal decisions and error propagation. To address these challenges, we propose the multi-hop reasoning enhanced (MRE) framework, which enhances both forward and backward reasoning to improve the identification of globally optimal reasoning trajectories. Specifically, MRE begins with prompt engineering to guide LLM in generating diverse reasoning trajectories for the given question. Valid reasoning trajectories are then selected for supervised fine-tuning, serving as a cold-start strategy. Finally, we introduce Tree-Group Relative Policy Optimization (T-GRPO)—a recursive, tree-structured learning-by-exploration approach. At each hop, exploration establishes strong causal dependencies on the previous hop, while evaluation is informed by multi-path exploration feedback from subsequent hops. Experimental results on two TKGQA benchmarks indicate that the proposed MRE-based model consistently surpasses state-of-the-art (SOTA) approaches in handling complex multi-hop queries. Further analysis highlights improved interpretability and robustness to noisy temporal annotations.
Knowledge editing (KE) provides a scalable approach for updating factual knowledge in large language models without full retraining. While previous studies have demonstrated effectiveness in general domains and medical QA tasks, little attention has been paid to KE in multimodal medical scenarios. Unlike text-only settings, medical KE demands integrating updated knowledge with visual reasoning to support safe and interpretable clinical decisions. To address this gap, we propose MultiMedBench, the first benchmark tailored to evaluating KE in clinical multimodal tasks. Our framework spans both understanding and reasoning task types, defines a three-dimensional metric suite (reliability, generality, and locality), and supports cross-paradigm comparisons across general and domain-specific models. We conduct extensive experiments under single-editing and lifelong-editing settings. Results suggest that current methods struggle with generalization and long-tail reasoning, particularly in complex clinical workflows. We further present an efficiency analysis (e.g., edit latency, memory footprint), revealing practical trade-offs in real-world deployment across KE paradigms. Overall, MultiMedBench not only reveals the limitations of current approaches but also provides a solid foundation for developing clinically robust knowledge editing techniques in the future.
Large reasoning models (LRMs) have shown significant progress in test-time scaling through chain-of-thought prompting. Current approaches like search-o1 integrate retrieval augmented generation (RAG) into multi-step reasoning processes but rely on a single, linear reasoning path while incorporating unstructured textual information in a flat, context-agnostic manner. As a result, these approaches can lead to error accumulation throughout the reasoning chain, which significantly limits its effectiveness in medical question-answering (QA) tasks where both accuracy and traceability are critical requirements. To address these challenges, we propose MIRAGE (Multi-path Inference with Retrieval-Augmented Graph Exploration), a novel test-time scalable reasoning framework that performs dynamic multi-path inference over structured medical knowledge graphs. Specifically, MIRAGE 1) decomposes complex queries into entity-grounded sub-questions, 2) executes parallel inference paths, 3) retrieves evidence adaptively via neighbor expansion and multi-hop traversal, and 4) integrates answers using cross-path verification to resolve contradictions. Experiments on three medical QA benchmarks (GenMedGPT-5k, CMCQA, and ExplainCPE) show that MIRAGE consistently outperforms GPT-4o, Tree-of-Thought variants, and other retrieval-augmented baselines in both automatic and human evaluations. Additionally, MIRAGE improves interpretability by generating explicit reasoning chains that trace each factual claim to concrete paths within the knowledge graph, making it especially suitable for complex medical reasoning scenarios.
Understanding and reasoning over complex spreadsheets remain fundamental challenges for large language models (LLMs), which often struggle with intricate structures and rely solely on neural computation. In this work, we propose SheetBrain, a neuro-symbolic dual-workflow agent framework for precise and interpretable reasoning over tabular data. SheetBrain consists of an understanding module that produces a comprehensive overview of the spreadsheet, including structural summaries and query-specific analyses to guide execution; an execution module that integrates a Python sandbox with preloaded table-processing libraries and an Excel helper toolkit for effective data manipulation; and a validation module that verifies the correctness of reasoning and answers, triggering re-execution if necessary. We evaluate SheetBrain on multiple public QA and manipulation benchmarks, and introduce SheetBench, a new benchmark targeting large, multi-table, and structurally complex spreadsheets. Experimental results show that SheetBrain significantly improves reasoning performance on both existing benchmarks and the more challenging scenarios presented in SheetBench.
Sound separation (SS) and target sound extraction (TSE) are fundamental techniques for addressing complex acoustic scenarios. While existing SS methods struggle with determining the unknown number of sound sources, TSE approaches require precisely specified clues to achieve optimal performance. This paper proposes a unified framework that synergistically combines SS and TSE to overcome their individual limitations. Our architecture employs two complementary components: 1) An Encoder-Decoder Attractor (EDA) network that automatically infers both the source count and corresponding acoustic clues for SS, and 2) A multi-modal fusion network that precisely interprets diverse user-provided clues (acoustic, semantic, or visual) for TSE. Through joint training with cross-task consistency constraints, we establish a unified latent space that bridges both paradigms. During inference, the system adaptively operates in either fully autonomous SS mode or clue-driven TSE mode. Experiments demonstrate remarkable performance in both tasks, with notable improvements of 1.4 dB SDR improvement in SS compared to baseline and 86% TSE accuracy.
Improving Implicit Discourse Relation Recognition with Natural Language Explanations from LLMs
PDF ↗Implicit Discourse Relation Recognition (IDRR) remains a challenging task due to the requirement for deep semantic understanding in the absence of explicit discourse markers. A further limitation is that existing methods only predict relations without providing any supporting explanations. Recent advances in large language models (LLMs) have shown strong reasoning capabilities in both deep language understanding and natural language explanation generation. In this work, we propose a simple yet effective approach to distill the reasoning capabilities of LLMs into lightweight IDRR models to improve both performance and interpretability. Specifically, we first prompt an LLM to generate explanations for each training instance conditioned on its gold label. Then, we introduce a novel classification-generation framework that jointly performs relation prediction and explanation generation, and train it with the additional supervision of LLM-generated explanations. Our framework is plug-and-play, enabling easy integration with most existing IDRR models. Experimental results on PDTB demonstrate that our approach significantly improves IDRR performance, while human evaluation further confirms that the generated explanations enhance model interpretability. Furthermore, we validate the generality of our approach on sentiment classification and natural language inference.
Flow matching and diffusion bridge models have emerged as leading paradigms in generative speech enhancement, modeling stochastic processes between paired noisy and clean speech signals based on principles such as flow matching, score matching, and Schrödinger bridge. In this paper, we present a framework that unifies existing flow and diffusion bridge models by interpreting them as constructions of Gaussian probability paths with varying means and variances between paired data. Furthermore, we investigate the underlying consistency between the training/inference procedures of these generative models and conventional predictive models. Our analysis reveals that each sampling step of a well-trained flow or diffusion bridge model optimized with a data prediction loss is theoretically analogous to executing predictive speech enhancement. Motivated by this insight, we introduce an enhanced bridge model that integrates an effective probability path design with key elements from predictive paradigms, including improved network architecture, tailored loss functions, and optimized training strategies. Experiments on denoising and dereverberation tasks demonstrate that the proposed method outperforms existing flow and diffusion baselines with fewer parameters and reduced computational complexity. The results also highlight that the inherently predictive nature of this generative framework imposes limitations on its achievable upper-bound performance.
Probing Preference Representations: A Multi-Dimensional Evaluation and Analysis Method for Reward Models
PDF ↗Previous methods evaluate reward models by testing them on a fixed pairwise ranking test set, but they typically do not provide performance information on each preference dimension. In this work, we address the evaluation challenge of reward models by probing preference representations. To confirm the effectiveness of this evaluation method, we construct a Multi-dimensional Reward Model Benchmark (MRMBench), a collection of six probing tasks for different preference dimensions. We design it to favor and encourage reward models that better capture preferences across different dimensions. Furthermore, we introduce an analysis method, inference-time probing, which identifies the dimensions used during the reward prediction and enhances its interpretability. Through extensive experiments, we find that MRMBench strongly correlates with LLM alignment performance, supporting it as a reliable reference for developing advanced reward models. By analyzing the evaluation results on MRMBench, we reveal that reward models struggle to simultaneously capture preferences across multiple dimensions, highlighting the potential of multi-objective optimization in reward modeling. Furthermore, our results demonstrate that the proposed inference-time probing method provides a reliable metric for assessing the confidence of reward predictions, leading to improved alignment of large language models.
Mixture-of-Experts (MoE) architectures decouple model capacity from per-token computation, enabling scaling beyond the computational limits imposed by dense scaling laws. Yet how MoE architectures shape knowledge acquisition during pre-training—and how this process differs from dense architectures—remains unknown. To address this issue, we introduce Gated-LPI (Log-Probability Increase), a neuron-level attribution metric that decomposes log-probability increase across neurons. We present a time-resolved comparison of knowledge acquisition dynamics in MoE and dense architectures, tracking checkpoints over 1.2M (~ 5.0T tokens) and 600K (~ 2.5T tokens) training steps, respectively. Our experiments uncover three patterns: (1) Low-entropy backbone. The top approximately 1% of MoE neurons capture over 45% of positive updates, forming a high-utility core, which is absent in the dense baseline. (2) Early consolidation. The MoE model locks into a stable importance profile within < 100K steps, whereas the dense model remains volatile throughout training. (3) Functional robustness. Masking the ten most important MoE attention heads reduces relational HIT@10 by < 10%, compared with > 50% for the dense model, showing that sparsity fosters distributed—rather than brittle—knowledge storage. These patterns collectively demonstrate that sparsity fosters an intrinsically stable and distributed computational backbone from early in training, helping bridge the gap between sparse architectures and training-time interpretability.
PRAGWORLD: A Benchmark Evaluating LLMs’ Local World Model Under Minimal Linguistic Alterations and Conversational Dynamics
PDF ↗Real-world conversations are rich with pragmatic elements, such as entity mentions, references, and implicatures. Understanding such nuances is a requirement for successful natural communication, and often requires building a local _world model_ which encodes such elements and captures the dynamics of their evolving states. However, it is not well-understood whether language models (LMs) construct or maintain a robust implicit representation of conversations. In this work, we evaluate the ability of LMs to encode and update their internal world model in dyadic conversations and test their _malleability_ under linguistic alterations. To facilitate this, we apply seven minimal linguistic alterations to conversations sourced from popular conversational QA datasets and construct a benchmark with two variants (i.e., Manual and Synthetic) comprising yes-no questions. We evaluate nine open and one closed source LMs and observe that they struggle to maintain robust accuracy. Our analysis unveils that LMs struggle to memorize crucial details, such as tracking entities under linguistic alterations to conversations. We then propose a dual-perspective interpretability framework which identifies transformer layers that are _useful_ or _harmful_ and highlights linguistic alterations most influenced by harmful layers, typically due to encoding spurious signals or relying on shortcuts. Inspired by these insights, we propose two layer-regularization based fine-tuning strategies that suppress the effect of the harmful layers.
Efficient Transcoder Adaptation for Fine-Tuned Models: Revealing Medical Reasoning Mechanisms in Large Language Models
PDF ↗Large language models (LLMs) suffer from a lack of decision-making transparency, limiting their deployment in high-stakes domains such as healthcare. We propose a mechanistic interpretability framework that introduces two novel paradigms: Medical Fine-Tuning with Frozen Attention Layers (FTFA) and Posterior Adaptation Transcoders (PAT). FTFA freezes attention layers while fine-tuning only feed-forward network (FFN) parameters, enabling PAT to efficiently adapt pre-trained transcoders on the same data. This approach achieves over 1000× efficiency improvement compared to training transcoders from scratch. We theoretically justify this methodology and demonstrate its cost-effectiveness for cross-domain transfer. Transcoders are sparse autoencoders that replace MLP layers to provide interpretable feature representations. By substituting MLP layers of both base Gemma2-2b and its medical fine-tuned variant with per-layer transcoders, we enable feature-level attribution analysis. Through systematic pruning and node merging of resulting attribution graphs, we construct human-interpretable decision pathways. Our analysis reveals that LLMs employ two parallel mechanisms for medical diagnosis: pattern matching and multi-hop reasoning, with fine-tuned models demonstrating enhanced correct reasoning patterns. This work provides a practical framework for training transcoders on fine-tuned models at minimal cost, enabling broader application of mechanistic interpretability across domains and potentially guiding model training through transcoder-based analysis.
Scope Delineation Before Localization: A Two-Stage Framework for Enhancing Failure Attribution in Multi-Agent Systems
PDF ↗Large language models (LLMs) are seeing growing adoption in multi-agent systems. In these systems, efficient failure attribution is critical for ensuring robustness and interpretability. Current LLM-based attribution methods often face challenges with lengthy logs and lacking expert knowledge. Drawing inspiration from human debugging strategies, we propose an automated failure attribution framework, Scope Delineation Before Localization, which operates in two key stages: (1) identifying the failure scope and (2) pinpointing the failure step. By decoupling failure attribution into the two stages, our approach alleviates the reasoning workload of LLMs, enabling more precise failure attribution. To support scope delineation, we further introduce two strategies: Stepwise Scope Delineation and Expertise-Assisted Scope Delineation. Experiments on the Who&When dataset validate the efficacy of our two-stage framework, demonstrating substantial improvements over prior methods (up to 24.27% on step-level accuracy).
Bridging the Language Gap: Uncovering and Aligning Shared Circuits for Multi-Hop Reasoning in Multilingual LLMs
PDF ↗Large language models (LLMs) present a paradox: they can correctly answer a multi-hop factual query in a high-resource language like English, yet fail on the identical query in another language. This raises a fundamental question about the nature of multilingual knowledge: are facts missing, or merely inaccessible? The underlying mechanisms for this knowledge gap have remained largely unexplored. In this work, we resolve this question by introducing a mechanistic interpretability framework that traces the causal pathways of multi-hop knowledge reasoning. Our analysis reveals a core, non-obvious finding: cross-lingual inconsistencies do not stem from a knowledge deficit. Instead, factual knowledge is robustly stored in a set of **shared, language-agnostic semantic neurons**. The failure originates from **misaligned attention pathways**, where a common set of critical attention heads fails to correctly route information along the reasoning chain to the appropriate knowledge neurons in lower-resource languages. This mechanistic diagnosis motivates a targeted alignment strategy: a surgical fine-tuning of only these critical heads. Experiments demonstrate that our method achieves significant improvements in multilingual multi-hop factuality—with positive cross-lingual transfer—while uniquely preserving general model capabilities, offering a scalable and mechanistically-grounded approach to building more reliable multilingual models.
Source attribution aims to enhance the reliability of AI-generated answers by including references for each statement, helping users validate the provided answers. However, existing work has primarily focused on text-only scenario and largely overlooked the role of multimodality. We introduce MAVIS, the first benchmark designed to evaluate multimodal source attribution systems that understand user intent behind visual questions, retrieve multimodal evidence, and generate long-form answers with citations. Our dataset comprises 157K visual QA instances, where each answer is annotated with fact-level citations referring to multimodal documents. We develop fine-grained automatic metrics along three dimensions of informativeness, groundedness, and fluency, and demonstrate their strong correlation with human judgments. Our key findings are threefold: (1) LVLMs with multimodal RAG generate more informative and fluent answers than unimodal RAG, but they exhibit weaker groundedness for image documents than for text documents, a gap amplified in multimodal settings. (2) Given the same multimodal documents, there is a trade-off between informativeness and groundedness across different prompting methods. (3) Our proposed method highlights mitigating contextual bias in interpreting image documents as a crucial direction for future research.
Causal Reward Adjustment: Mitigating Reward Hacking in External Reasoning via Backdoor Correction
PDF ↗External reasoning systems combine language models with process reward models (PRMs) to select high-quality reasoning paths for complex tasks such as mathematical problem solving. However, these systems are prone to reward hacking, where high-scoring but logically incorrect paths are assigned high scores by the PRMs, leading to incorrect answers. From a causal inference perspective, we attribute this phenomenon primarily to the presence of confounding semantic features. To address it, we propose Causal Reward Adjustment (CRA), a method that mitigates reward hacking by estimating the true reward of a reasoning path. CRA trains sparse autoencoders on the PRM’s internal activations to recover interpretable features, then corrects confounding by using backdoor adjustment. Experiments on math solving datasets demonstrate that CRA mitigates reward hacking and improves final accuracy, without modifying the policy model or retraining PRM.
The deployment of large language models for specialized tasks often requires domain-specific parameter-efficient finetuning through Low-Rank Adaptation (LoRA) modules. However, effectively fusing these adapters to handle complex, multi-domain composite queries remains a critical challenge. Existing LoRA fusion approaches either use static weights, which assign equal relevance to each participating LoRA, or require data-intensive supervised training for every possible LoRA combination to obtain respective optimal fusion weights. We propose qa-FLoRA, a novel query-adaptive data-and-training-free method for LoRA fusion that dynamically computes layer-level fusion weights by measuring distributional divergence between the base model and respective adapters. Our approach eliminates the need for composite training data or domain-representative samples, making it readily applicable to existing adapter collections. Extensive experiments across nine multilingual composite tasks spanning mathematics, coding, and medical domains, show that qa-FLoRA outperforms static fusion by ~5% with LLaMA-2 and ~6% with LLaMA-3, and the training-free baselines by ~7% with LLaMA-2 and ~10% with LLaMA-3, while significantly closing the gap with supervised baselines. Further, layer-level analysis of our fusion weights reveals interpretable fusion patterns, demonstrating the effectiveness of our approach for robust multi-domain adaptation.