Retrieval‑augmented generation (RAG) improves knowledge‑intensive QA tasks by incorporating external evidence, yet retrieval remains imperfect and often returns irrelevant or misleading passages. We refer to this as Noisy RAG, where LLMs suffer from inconsistent relevance signals due to the fact that relevance identification is carried out by only a small subset of internal modules (e.g., retrieval heads) and can easily be confounded by other modules. Recent work attempts to enhance the denoising ability of the LLM through external emphasis, prompting the LLM to highlight helpful passages as the emphasis before answering. However, these textual identifications themselves depend on the same inconsistent signals and thus become less reliable. Our key insight is that, because relevance signals reside in only a few modules, emphasis should be applied directly within these modules rather than only at the text level. We therefore propose \textit{selective internal emphasis} to amplify the relevance signals, implemented via a lightweight, plug‑and‑play signal amplifier that operates inside the LLM. The amplifier performs token‑ and channel‑level selective emphasis during a standard RAG fine‑tuning pipeline, with the base LLM frozen. Across four QA benchmarks and three LLM scales, our method consistently improves accuracy and robustness. Furthermore, generalization and interpretability analyses show that the amplifier captures retrieval‑related patterns rather than only dataset‑specific patterns, enabling more reliable passage-denoising.
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
Effective multimodal learning in real-world scenarios depends on a nuanced treatment of uncertainty, which arises at three levels: (1) Intrinsic Uncertainty from modality-specific noise or ambiguity; (2) Relational Uncertainty due to cross-modal conflicts or redundancy; and (3) Aggregated Uncertainty when fusing potentially inconsistent signals. Most existing methods overlook this hierarchy, applying a single uncertainty model. We propose Adaptive Evidential Multimodal Representation Learning (AEMRL), a framework aligned with this multi-level view. To address intrinsic uncertainty, Disentangled Evidential Uncertainty Encoding (D-EUE) provides interpretable, class-aware reliability scores per modality. For relational uncertainty, Uncertainty-Conditioned Dynamic Factorization (UDF) uses a hypernetwork to dynamically extract complementary cues and suppress conflict. To resolve aggregated uncertainty, Adaptive Fusion & Conflict-Aware Calibration (AFCAC) adaptively weights evidence streams and calibrates final predictions based on detected conflicts. Extensive experiments on five diverse benchmark datasets show that AEMRL consistently enhances task accuracy, reduces calibration error, and improves robustness to noise, semantic conflict, and missing modalities.
Null-space-based methods have garnered considerable attention in model editing by constraining updates to the null space of the pre-existing knowledge representation, thereby preserving the model's original behavior. However, in practice these methods rely on an approximate null space—leading to knowledge leakage—and further suffer from severe performance degradation during sequential editing. Recent work shows that history-aware editing strategies can empirically mitigate this decline, yet the underlying reason remains unclear. In this paper, we first expose the knowledge leakage inherent in existing null-space approaches and then analyze why history-aware updates effectively preserve both editing performance and general capabilities during long-horizon editing. Building on these insights, we propose BetaEdit, a refined framework that effectively controls the knowledge leakage and integrates history-aware updates into the null-space paradigm. Extensive experiments on three large language models across two standard benchmarks show that BetaEdit consistently outperforms prior methods in the challenging regime of massive-scale sequential editing. Code is available at: https://github.com/lbq8942/BetaEdit.
Skeleton-based human activity recognition (HAR) has achieved strong empirical performance, yet most existing models remain black boxes and difficult to interpret. In this work, we introduce a neurosymbolic formulation of skeleton-based HAR that reframes action recognition as concept-driven first-order logical reasoning over motion primitives. Our framework bridges representation learning and symbolic inference by grounding first-order logic predicates in learnable spatial and temporal motion concepts. Specifically, we employ a standard spatio-temporal skeleton encoder to extract latent motion representations, which are then mapped to interpretable concept predicates via a spatio-temporal concept decoder that explicitly separates pose-centric and dynamics-centric abstractions. These concept predicates are composed through differentiable first-order logic layers, enabling the model to learn human-readable logical rules that govern action semantics. To impose semantic structure on the learned concepts, we align skeleton representations with LLM-derived descriptions of atomic motion primitives, establishing a shared conceptual space for perception and reasoning. Extensive experiments on NTU RGB+D 60/120 and NW-UCLA demonstrate that our approach achieves competitive recognition performance while providing explicit, interpretable explanations grounded in logical structure. Our results highlight neurosymbolic reasoning as an effective paradigm for interpretable spatio-temporal action understanding.
While Vision-Language Models (VLMs) have shown promise in textual understanding, they face significant challenges when handling long context and complex reasoning tasks. In this paper, we dissect the internal mechanisms governing long-context processing in VLMs to understand their performance bottlenecks. Through the lens of attention analysis, we identify specific Visual Evidence Retrieval (VER) Heads —a sparse, dynamic set of attention heads critical for locating visual cues during reasoning, distinct from static OCR heads. We demonstrate that these heads are causal to model performance; masking them leads to significant degradation. Leveraging this discovery, we propose VERA (Visual Evidence Retrieval Augmentation), a training-free framework that detects model uncertainty (i.e., entropy) to trigger the explicit verbalization of visual evidence attended by VER heads. Comprehensive experiments demonstrate that VERA significantly improves long-context understanding of open-source VLMs: it yields an average relative improvement of 21.3% on Qwen3-VL-8B-Instruct and 20.1% on GLM-4.1V-Thinking across five benchmarks.
Structured claim decomposition is often proposed as a solution for verifying complex, multi-faceted claims, yet empirical results have been inconsistent. We argue that these inconsistencies stem from two overlooked bottlenecks: evidence alignment and sub-claim error profiles. To better understand these factors, we introduce a new dataset of real-world complex claims, featuring temporally bounded evidence and human-annotated sub-claim evidence spans. We evaluate decomposition under two evidence alignment setups: Sub-claim Aligned Evidence (SAE) and Repeated Claim-level Evidence (SRE). Our results reveal that decomposition brings significant performance improvement only when evidence is granular and strictly aligned. By contrast, standard setups that rely on repeated claim-level evidence (SRE) fail to improve and often degrade performance as shown across different datasets and domains (PHEMEPlus, MMM-Fact, COVID-Fact). Furthermore, we demonstrate that in the presence of noisy sub-claim labels, the nature of the error ends up determining downstream robustness. We find that conservative "abstention" significantly reduces error propagation compared to aggressive but incorrect predictions. These findings suggest that future claim decomposition frameworks must prioritize precise evidence synthesis and calibrate the label bias of sub-claim verification models.
Offline-to-online reinforcement learning harnesses the stability of offline pretraining and the flexibility of online fine-tuning. A key challenge lies in the non-stationary distribution shift between offline datasets and the evolving online policy. Common approaches often rely on static mixing ratios or heuristic-based replay strategies, which lack adaptability to different environments and varying training dynamics, resulting in suboptimal tradeoff between stability and asymptotic performance. In this work, we propose Reinforcement Learning with Optimized Adaptive Data-mixing (ROAD), a dynamic plug-and-play framework that automates the data replay process. We identify a fundamental objective misalignment in existing approaches. To tackle this, we formulate the data selection problem as a bi-level optimization process, interpreting the data mixing strategy as a meta-decision governing the policy performance (outer-level) during online fine-tuning, while the conventional Q-learning updates operate at the inner level. To make it tractable, we propose a practical algorithm using a multi-armed bandit mechanism. This is guided by a surrogate objective which simultaneously maintains offline priors and prevents value overestimation. Our empirical results demonstrate that this approach consistently outperforms existing data replay methods across various datasets, eliminating the need for manual, context-specific adjustments while achieving superior stability and asymptotic performance.
Evaluation of tool-augmented Large Language Models (LLMs) has not advanced far beyond final answer accuracy, and neglects in-depth evaluation of reasoning ability despite it being a central claim of recent models. We aim to address this gap by developing a dependency graph-based evaluation to give insight into meta-level reasoning ability. We create a two-stage algorithm to construct dependency graphs from a series of LLM-generated tool calls: in the first stage creating nodes for tool calls, then in the second creating edges where tool call results are identified as being used as arguments in subsequent calls. The resulting graph facilitates the development of graph-theoretic metrics which allows deeper evaluation of the reasoning ability of LLMs over the task in question. We also use the dependency graph to assist with the generation of natural language explanations of the question-answering process. We create a baseline explanation generated from the list of tool calls, and a graph-augmented explanation which incorporates information about information provenance. A human evaluation study shows that our graph-augmented explanations are preferred across three key explainability criteria, although the extra detail contained impacts ease of understanding.
Neural Network k-Nearest Neighbor (NN-kNN) was proposed as an interpretable network model that learns feature weights and similarity to retrieve relevant cases for classification. This paper extends it to regression with the goal of generating accurate predictions based on neighboring cases with similar labels. Specifically, we introduce three modular components: an attention mechanism that weights the contribution of retrieved cases, a locality-aware regularizer that favors label-similar neighbors, and an optional case adaptation module that refines the retrieved estimate. Across synthetic and standard tabular regression benchmarks, NN-kNN achieves competitive predictive error against strong baselines (kNN-R, MLKR, and MLPs) while providing cases with similar labels as explanation (later referred as label-similar case). Moreover, NN-kNN supports manual knowledge injection through tuning weights for human-comprehensible features. The result is a simple, general, and interpretable approach to continuous-valued prediction that unifies retrieval, attention, and optional case adaptation within a single neural framework.
Understanding the training dynamics of deep neural networks remains a major open problem, with physics-inspired approaches offering promising insights. Building on this perspective, we develop a thermodynamic framework to describe the stationary distributions of stochastic gradient descent (SGD) with weight decay for scale-invariant neural networks, a setting that both reflects practical architectures with normalization layers and permits theoretical analysis. We establish analogies between training hyperparameters (e.g., learning rate, weight decay) and thermodynamic variables such as temperature, pressure, and volume. Starting with a simplified isotropic noise model, we uncover a close correspondence between SGD dynamics and ideal gas behavior, validated through theory and simulation. Extending to training of neural networks, we show that key predictions of the framework, including the behavior of stationary entropy, align closely with experimental observations. This framework provides a principled foundation for interpreting training dynamics and may guide future work on hyperparameter tuning and the design of learning rate schedulers.
Understanding overthinking in large reasoning models (LRMs) is crucial for interpretability as well as reasoning efficiency and effectiveness. However, existing approaches primarily adopt coarse-grained reasoning strategies, such as truncating Chains-of-Thought or switching reasoning modes, which reduce verbosity but are insufficient to actively guide reasoning toward more effective trajectories. To address these issues, we propose a training-free, interpretable framework that selects thinking words via attention heads to guide LRMs toward more effective reasoning. Specifically, we categorize thinking steps into effective and redundant states, identify the attention head that best discriminates between them as the Thinking Partition Head to construct an Effective Thinking Representation Space, and compute the Information Gain Ratio (IGR) between candidate thinking words and this space to select the word that steers reasoning toward a more effective direction. Extensive experiments on mathematical and scientific reasoning benchmarks, including AIME24, AMC23, MATH-500, GSM8K, and GPQA-D, show that our method consistently outperforms the strong baseline DEER, achieving average improvements of 1.2–1.3% in accuracy and 2.5–4.5% in compression rate. Compared to vanilla baselines, our approach yields larger gains of 2.6–6.3% in accuracy while reducing token usage by 22–43%.
We present a unified framework for modeling agents’ epistemic states and counterfactual conditionals. The novelty of our approach lies at the semantic level. We introduce a computationally grounded semantics in which an agent’s doxastic accessibility relation, needed to model the standard notion of deductively closed belief, is derived from the agent’s belief base, and in which the notion of comparative similarity between states, needed to interpret counterfactual conditionals, is computed from both the atomic formulas describing the agents’ belief bases and those describing the environment. We use this semantics to interpret a language of beliefs and counterfactuals. We demonstrate the expressiveness of our language and the flexibility of the semantics by formalizing a wide range of notions: counterfactual dependence, counterfactual notions of reason and knowledge, conditional belief. Furthermore, we show that our semantics illuminates the subtle interplay between belief change and counterfactuals. On the computational side, we provide a succinct formulation of model checking for our language and establish a PSPACE complexity result.
Chain-of-Thought (CoT) reasoning empowers Large Language Models (LLMs) with remarkable capabilities, but typically requires enormous parameter scales. CoT distillation has emerged as a promising paradigm to transfer reasoning prowess into compact student models (SLMs), yet existing approaches often rely on a solitary teacher, capping the student’s potential since individual LLMs often exhibit distinct capability biases and may suffer from catastrophic forgetting. While leveraging diverse teachers seems appealing, effectively fusing their supervisions remains challenging: teacher-student incompatibility risks amplifying hallucinations, and passive supervision fails to ensure genuine logic internalization. To address this, we introduce COMPACT, a framework that adaptively fuses supervisions from different teachers by dynamically weighting teacher gradients based on the student’s real-time compatibility evaluated by a multi-dimensional metric: (1) Graph-based Consensus to filter misleading rationales by identifying mainstream reasoning paths; (2) Mutual-Information-based Adaptability to detect "epiphany moments" for genuinely understanding the reasoning process rather than merely imitating; and (3) Loss-based Difficulty to assess student receptivity to the teacher's guidance and prevent negative transfer. Our experiments and latent space analysis demonstrate that COMPACT effectively integrates diverse reasoning capabilities without damaging the model's original knowledge structure, achieving state-of-the-art performance on various benchmarks while effectively mitigating catastrophic forgetting. Code available: https://github.com/CAG-Research/COMPACT.git.
Preference alignment is a critical step in making Large Language Models (LLMs) useful and aligned with (human) preferences. Existing approaches such as Reinforcement Learning from Human Feedback or Direct Preference Optimization typically require curated data and expensive optimization over billions of parameters, and eventually lead to persistent task-specific models. In this work, we introduce Preference alignment of Large Language Models via Residual Steering (PaLRS), a training-free method that exploits preference signals encoded in the residual streams of LLMs. From as few as one hundred preference pairs, PaLRS extracts lightweight, plug-and-play steering vectors that can be applied at inference time to push models toward preferred behaviors. We evaluate PaLRS on various small-to-medium-scale open-source LLMs, showing that PaLRS-aligned models achieve consistent gains on mathematical reasoning and code generation benchmarks while preserving baseline general-purpose performance. Moreover, when compared to models aligned with DPO and SimPO, they perform better with great time savings. Our findings highlight that PaLRS offers an effective, much more efficient and flexible alternative to standard preference optimization pipelines, offering a training-free, plug-and-play mechanism for alignment with minimal data. Extended version with Suppl. Mat. is available at https://doi.org/10.48550/arXiv.2509.23982.
Cross-view Referring Expression Comprehension (REC) requires a model to accurately localize objects in a spatial representation of one perspective based on descriptions generated from a different perspective. Essentially, it demands maintaining referential consistency under perspective shifts. This problem is prevalent in engineering application scenarios (e.g., Geographic Information Systems (GIS), urban planning, etc), where descriptions are typically generated from a top-down or planning view, while targets must be identified in a spatial representation distinct from the source. However, existing methods often assume shared perspectives or rely on an explicit observer to interpret observer-centric spatial relations (e.g., "left/right" of the observer). Consequently, they struggle to handle cross-view grounding cases commonly arising in engineering application scenarios, where descriptions are defined by the source-view representation itself and do not involve any observing subject. To this end, we propose a novel Subject-Agnostic Cross-View REC task and construct a controlled simulation environment to systematically study referential consistency under perspective changes. Methodologically, we propose a Reward-Driven Consistency Learning framework. By imposing mutually constraining reward signals on the grounding and localization results between the source and target views, we guide the model to maintain referential consistency across different view representations. Experimental results show that our method significantly outperforms baselines in multiple cross-view settings, exhibiting superior generalization and robustness under perspective shifts.
Unifying Perspectives: Plausible Counterfactual Explanations on Global, Group-wise, and Local Levels
The growing complexity of AI systems has intensified the need for transparency through Explainable AI (XAI). Counterfactual explanations (CFs) offer actionable "what-if" scenarios on three levels: Local CFs providing instance-specific insights, Global CFs addressing broader trends, and Group-wise CFs (GWCFs) striking a balance and revealing patterns within cohesive groups. Despite the availability of methods for each granularity level, the field lacks a unified method that integrates these complementary approaches. We address this limitation by proposing a gradient-based optimization method for differentiable models that generates Local, Global, and Group-wise Counterfactual Explanations in a unified manner. We especially enhance GWCF generation by combining instance grouping and counterfactual generation into a single efficient process, replacing traditional two-step methods. Moreover, to ensure trustworthiness, we innovatively introduce the integration of plausibility criteria into the GWCF domain, making explanations both valid and realistic. Our results demonstrate the method's effectiveness in balancing validity, proximity, and plausibility while optimizing group granularity, with practical utility validated through practical use cases.
Single-cell perturbation screens enable systematic discovery of gene regulatory mechanisms, yet the exponential expansion of perturbation space makes comprehensive experimentation impractical. Although in silico predictors have been increasingly proposed to address this challenge, most existing methods either assume static regulatory priors or learn unconstrained data-driven representations. As a result, they often suffer from poor adaptation to state-dependent shifts and weak out-of-distribution generalization. In response, we propose URFPert, an unrolled regulatory flow network tailored for single-cell perturbation prediction. Specifically, URFPert is composed of three fundamental modules: (i) Regulatory Structure Unrolling module, which employs an iterative unrolling optimization algorithm to infer state-dependent gene regulatory networks (GRNs); (ii) Heterogeneous Graph Flow Network, which incorporates the inferred GRNs into a conditional velocity field to yield state-consistent and perturbation-specific trajectories; (iii)Discriminative Optimization Strategy, which combines Mean Squared Error~(MSE) with Maximal Coding Rate Reduction~(MCR²) to promote discriminative perturbation responses and mitigate training confounders. Comprehensive evaluations on five single-cell perturbation datasets demonstrate that URFPert outperforms state-of-the-art methods in unseen perturbation regimes, providing a powerful tool for interpreting regulatory mechanisms. The project code is public via https://github.com/Jiniretttt/URFPert.
Attribution techniques explain the outcome of an AI model by assigning a numerical score to its inputs. So far, these techniques have mainly focused on attributing importance to static input features at a single point in time, and thus fail to generalize to sequential decision-making settings. This paper fills this gap by introducing techniques to generate attribution-based explanations for Markov Decision Processes (MDPs). We give a formal characterization of what attributions should represent in MDPs, focusing on explanations that assign importance scores to both individual states and execution paths. We show how importance scores can be computed by leveraging techniques for strategy synthesis, enabling the efficient computation of these scores despite the non-determinism inherent in an MDP. We evaluate our approach on five case-studies, demonstrating its utility in providing interpretable insights into the logic of sequential decision-making agents.
Motor imagery (MI) electroencephalography (EEG) decoding has benefited from deep learning, yet methods operate in Euclidean space and behave as opaque black boxes, neglecting the intrinsic geometry of functional brain connectivity. EEG connectivity descriptors, such as phase synchrony and covariance matrices, naturally reside on the manifold of symmetric positive definite (SPD) matrices, where Euclidean operations are geometrically inconsistent and hinder interpretability. This work proposes the Multi-frequency and Multi-scale Riemannian Network (MFMSRNet), an interpretable end-to-end geometry-aware framework for MI EEG decoding on the SPD manifold. The method constructs kernelized phase-locking value (KPLV) functional connectivity (FC) matrices to capture nonlinear phase synchrony while ensuring positive definiteness. An attention-based Riemannian fusion mechanism adaptively integrates information across multiple frequency bands in the tangent space. Furthermore, a multi-scale Riemannian network extracts global, hemispheric, and local connectivity patterns via manifold-preserving bilinear mappings and smooth eigenvalue rectification. Extensive experiments indicate that MFMSRNet yields more expressive and interpretable representations for robust MI decoding, offering a promising solution for reliable brain–computer interface applications. The code is available at https://github.com/Raeno-Rao/MFMSRNet.
Empowering Precise Embodied Agents with Executable Analytic Concepts as Semantic-Physical Blueprints
A core challenge for embodied agents is the ``semantic-to-physical gap"—the difficulty of mapping symbolic reasoning to precise execution. While Vision-Language Models (VLMs) enhance agent task planning, they often fail in problem classes requiring accurate alignment between functional geometry and physical constraints, such as articulated object manipulation or precision assembly. To address these challenges, we propose GRACE, an agent framework that adopts Executable Analytic Concepts (EAC) as a core knowledge annotation paradigm for object understanding. Under this paradigm, the agent does not merely perceive objects as unstructured visual data; instead, it interprets and annotates them as ``Semantic-Physical Blueprints", which provide a structured cognitive representation that encodes geometric primitives, mechanical affordances, and manipulation constraints. Central to our framework is a policy scaffolding mechanism, which allows the agent to dynamically ground VLM-based insights into these instantiated blueprints. This enables the agent to model and resolve complex decision-making tasks involving parameterized object descriptions and constrained motion planning. Experimental results demonstrate that agents utilizing this paradigm achieve high robustness in complex manipulation tasks. Furthermore, we demonstrate that GRACE supports automated concept discovery, offering a scalable and modular solution for high-precision embodied intelligence.