Large Language Models (LLMs) powerful generative capabilities also pose significant risks, underscoring the need for effective detoxification methods to ensure safer deployment. Due to the polysemantic nature of LLM neurons, recent neuron intervention methods inevitably entangle unrelated concepts, compromising generation quality and interpretability. Sparse Autoencoders (SAEs) have opened new horizons for decomposing model activations into monosemantic features, offering interpretability and targeted feature-level steering. Empirical findings reveal that, despite capturing interpretable features, indiscriminate interventions on toxicity-related features expose the fragility of LLMs, achieving toxicity mitigation at the cost of degraded fluency. Building upon this finding, we propose DeLFE, a lightweight controlled detoxification approach that identifies specific toxic features across model layers and performs targeted interventions on them. DeLFE learns toxicity subspaces from label-guided SAE feature subsets to characterize toxic v.s. non-toxic activation patterns. When auto-completing a response token-by-token, DeLFE tracks the toxicity-triggering risks and steers toxic features away from the subspace via a flow-matching feature transformation. We further design three feature-level strategies that adjust intervention timing and strength to reconstruct the target model’s original activations. Extensive experiments demonstrate that our method achieves strong detoxification effectiveness while maintaining high generation quality across models of varying sizes and diverse base LLMs.
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
The Dynamic Flexible Job Shop Scheduling Problem (DFJSP) necessitates a trade-off between instant reaction to stochastic disturbances and global optimization of production goals. Conventional priority rules are insufficiently flexible to handle complex disruptions, whereas learning-based approaches often compromise interpretability or fail to generalize across problem scales. Although Large Language Models (LLMs) offer advanced reasoning capabilities to bridge this gap, their substantial inference latency is incompatible with the millisecond-level decision cycles of industrial control systems. To resolve this conflict, we introduce RACE-Sched, an asynchronous agent-based framework that decouples policy execution from logical reasoning via a dual-stream architecture. The Reactive Stream executes low-latency symbolic heuristics to enable real-time dispatching, while the parallel Deliberative Stream leverages an LLM to synthesize, validate, and evolve these rules. Candidate rules undergo rigorous testing in a sandbox and are deployed via atomic updates, ensuring safety without blocking the control loop. Additionally, a semantic rule repository indexes validated heuristics for retrieval-based initialization which enhances transferability across problem scales. Extensive evaluations on GEN-Bench, MK-Bench, and JMS-Bench demonstrate that RACE-Sched outperforms leading Deep Reinforcement Learning and other LLM-based baselines. This approach harmonizes real-time constraints with long-horizon reasoning to achieve superior solution quality and robust adaptation to dynamic events.
Adaptive Learning Systems (ALS) play a key role in promoting the equity of education, which can provide personalized teaching on a large scale. However, the existing question recommendation methods face the problem of data sparsity at the scenario of cold-start, and lack the ability of long-term teaching reasoning. To address these challenges, we propose QuesRecAgent, a multi-agent framework which involves Large Language Models (LLMs) and the Relexion mechanism for adaptive question recommendation. Different from the traditional black-box methods, the QuesRecAgent applies a framework Dual-Loop State Decoupling architecture, which achieves a rapid updating of beliefs through simulated interaction in the inner loop and do the authoritative diagnosis and meta-strategy reflection to optimize the teaching strategy in the outer loop. Furthermore, with the support of a Topological Knowledge Graph (TKG) and the theory of Zone of Proximal Development (ZPD), the QuesRecAgent ensures a high-quality learning guidance even facing students with limited interaction records. Our experiments on three real-world datasets show that the QuesRecAgent outperforms state-of-the-art baselines in total knowledge gain and other metrics, providing a robust and interpretable solution for promoting high-quality personalized education.
Artificial intelligence models typically perform well on large-scale datasets, yet their effectiveness tends to degrade in real-world scenarios with scarce data, such as medical diagnostics. In contrast, humans can learn and reason effectively from few examples. Even when novel objects differ significantly from prior observations, humans can quickly infer category distributions accurately. Inspired by this, we explore how to leverage human-inspired cognitive mechanisms to improve the distribution calibration in few-shot learning models. Based on the ``fast-slow thinking'' dual-process theory, we propose a novel cognition-inspired few-shot learning framework. It mimics the human cognition when handling novel information: it first rapidly screens relevant knowledge through ``fast thinking'' and then infers inter-class relationships and calibrates distributions via ``slow thinking''. Specifically, the fast-thinking stage employs a gating mechanism to quickly match input samples with known base-class prototypes, activating relevant candidate knowledge. In the slow-thinking stage, the model aggregates inter-class edge embeddings into a summary relation graph and then applies divergent Gaussian sampling to generate multiple relation graphs representing different association strengths, thereby achieving distribution calibration for few-shot classes. Experiments on public few-shot benchmarks and medical image datasets show competitive performance. Visualization analyses further reveal that our framework exhibits meaningful interpretability.
3D Gaussian Splatting enables efficient, high-fidelity novel view synthesis with explicit Gaussians and differentiable rendering. However, adverse weather introduces rain streaks and droplets as well as volumetric scattering, producing view-dependent, spatially varying degradations that break the clean multi-view consistency assumption and lead to geometric drift and unstable appearance. Existing approaches are largely confined to the 2D image domain and seldom model the 3D degradation formation process, limiting controllable weather synthesis and consistent restoration in 3DGS. To our knowledge, we are the first to incorporate real meteorological observations as an external prior and propose a 3D-level unified rendering–restoration framework for joint deraining and dehazing within a single 3DGS pipeline. Precipitation is mapped to interpretable parameters controlling the density and morphology of rain Gaussians, while an atmospheric-scattering extinction strength drives 3D fog generation for reproducible weather modeling. We adopt a closed-loop dual-branch optimization where the rendering branch fits degraded observations to capture weather degradation patterns, and the restoration branch regularizes scene Gaussians toward a clean domain with multi-scale perceptual consistency, suppressing artifacts and improving cross-view detail fidelity. Experiments across diverse scenes and weather conditions consistently outperform strong baselines on all three metrics, with a particularly notable PSNR improvement of 1.78 dB on average.
With the rising demand for trustworthy AI in clinical practice, strong interpretability is now a critical requirement as well as accuracy. However, the modality gap for medical visual question answering is quite severe when continuous visual signals are forcibly projected into discrete text space for reasoning, and the loss of necessary diagnostic information leads to low precision and black-box opacity. To address this problem, we propose MedVCoT, which incorporates latent visual reasoning into the medical visual question answering(VQA) domain. Rather than merely integrating modules, MedVCoT utilizes the specialized expertise of MedSAM to train a large vision-language model so that it can autonomously generate consistent and continuous latent visual tokens within Visual Chain-of-Thought. This mechanism forces the model to explicitly "see" the lesion in the latent space before formulating a textual diagnosis, ensuring answers are causally rooted in verifiable visual evidence rather than statistical hallucination. We achieve this through a progressive 3-stage training procedure: medical feature alignment, visual reasoning learning by utilizing latent tokens generated, and instruction tuning for complex clinical scenarios. Extensive experiments show that MedVCoT can achieve state-of-the-art performance on multiple benchmarks, outperforming other methods by large margins. Meanwhile, it provides pixel-level segmentation masks to validate its diagnostic reasoning. Our demo is available at https://zhuqh19.github.io/MedVCoT.
We present two new algorithms for the problem of lexicographic inference from conditional knowledge bases. These algorithms are based on SAT and MaxSAT encodings of the underlying problem of lexicographic comparisons of classical interpretations and require only a polynomial number of SAT solver calls. In our experimental evaluation we show that our new algorithms signficantly outperform the state of the art.
Current Parameter-Efficient Fine-Tuning (PEFT) methods typically operate under an implicit assumption: once a target module is selected, every token passing through it contributes equally to the downstream task and requires a parameter update. In this paper, we challenge this convention by revealing a pervasive token-level redundancy in the fine-tuning of large models (LMs). We propose TS-PEFT, a theoretical framework utilizing proximal optimization that acts as a dynamic probe to identify token-level redundancy during the fine-tuning process. Extensive experiments demonstrate that indiscriminately updating all tokens is not only computationally superfluous but often introduces optimization noise. Surprisingly, by discarding 30%-70% of token updates, TS-PEFT consistently matches or exceeds the performance of dense baselines such as LoRA, DoRA. Our in-depth analysis shows that the learned token-level sparsity is a superior indicator of module importance compared to traditional weight criteria, providing a novel data-driven perspective on the intrinsic adaptation mechanism of LMs.
Designing effective reward functions remains a major challenge in reinforcement learning (RL), particularly in open-ended environments where task goals are abstract and difficult to quantify. In this work, we present VLM-AR3L, a framework that leverages Vision-Language Models (VLMs) to provide both absolute and relative rewards for RL. VLM-AR3L interprets an agent’s visual observations in the context of a natural language task goal, and learns both absolute and relative rewards from VLM-generated preference labels. The absolute reward model predicts scalar evaluations for individual states, while the relative reward model compares consecutive observations to infer progress or regression toward the task goal. Their integration combines the stability of state-based evaluation with the robustness of comparative supervision. We evaluate VLM-AR3L across benchmarks spanning classic control, manipulation, and open-world embodied tasks, with a particular focus on Minecraft given its visual complexity and long-horizon decision-making requirements. Experimental results show that VLM-AR3L consistently outperforms prior VLM-based reward learning methods. Videos and code are available on the project website: https://vlm-ar3l.github.io/.
Graph Neural Networks (GNNs) are widely used in high-stakes decision-making, where explanation robustness is crucial. Yet, even minor structural perturbations can drastically change explanations without changing predictions. Existing methods primarily characterize structural perturbations by their edit distance or budget, and enforce robustness through uniform constraints or averaging mechanisms over this perturbation set, which fail to capture the inherently direction-dependent effects of structural changes. In this paper, we propose GeoXGNN (Local Geometry Improves Explanation Robustness for Graph Neural Networks), a robust explanation framework that explicitly models the directional sensitivity of structural perturbations from a geometric perspective. GeoXGNN introduces local, direction-aware metrics in the node representation space to characterize how perturbations along different directions affect node embeddings and explanation outcomes. Based on the learned local metrics, GeoXGNN further enforces directional constraints, encouraging explanations to remain robust along high-sensitivity directions while preserving explanatory fidelity. Extensive experiments on multiple node-level and graph-level benchmark datasets demonstrate that GeoXGNN significantly improves explanation robustness over existing methods.
As Large Language Models (LLMs) scale to handle massive context windows, achieving surgical feature-level interpretation is essential for high-stakes tasks like legal auditing and code debugging. However, existing local model-agnostic explanation methods face a critical dilemma in these scenarios: feature-based methods suffer from attribution dilution due to high feature dimensionality, which prevents them from providing faithful explanations. In this paper, we propose Focus-LIME, a coarse-to-fine framework designed to restore the tractability of surgical interpretation. Focus-LIME utilizes a proxy model to curate the perturbation neighborhood, allowing the target model to perform fine-grained attribution exclusively within the optimized context. Empirical evaluations on long-context benchmarks demonstrate that our method makes surgical explanations practical and provides faithful explanations to users.
Inductive Logic Programming (ILP) learns interpretable logical rules from data. Existing methods are transductive: their learned parameters are bound to specific predicates and require retraining for each new task. We introduce Neural Rule Inducer (NRI), a pretrained model for zero-shot rule induction. Rather than encoding literal identities, NRI represents literals using domain-agnostic statistical properties such as class-conditional rates, entropy, and co-occurrence, which generalize across variable identities and counts without retraining. The model consists of a statistical encoder and a parallel slot-based decoder. Parallel decoding preserves the permutation invariance of logical disjunction; an autoregressive decoder would instead impose an arbitrary clause order. Product T-norm relaxation makes rule execution differentiable, allowing end-to-end training on prediction accuracy alone. We evaluate NRI on rule recovery, robustness to label noise and spurious correlations, and zero-shot transfer to real-world benchmarks, and we believe this work opens up the possibility of foundation models for symbolic reasoning. Code and the reference checkpoint are available at https://github.com/phuayj/neural-rule-inducer. An extended version with full appendices is available at https://arxiv.org/abs/2605.04916.
Reinforcement learning (RL) agents often struggle to generalize to new tasks and contexts without updating their parameters, mainly because their learned representations and policies are overfit to the specifics of their training environments. To boost agents' in-context RL (ICRL) ability, this work formulates ICRL as a two-agent emergent communication problem and introduces CORAL (Communicative Representation for Adaptive RL), a framework that learns a transferable communicative context by functionally separating latent representation learning from control. In CORAL, an Information Agent (IA) is pre-trained as a world model on a diverse distribution of tasks. Its objective is not direct return maximization, but world modeling and distilling its understanding into concise messages. The emergent communication protocol is shaped by a novel Causal Influence Loss, which measures the effect that the message has on the next action. During deployment, the previously trained IA serves as a fixed contextualizer for a new Control Agent (CA), which learns to solve tasks by interpreting the provided communicative context. Our experiments demonstrate that this approach enables the CA to achieve significant gains in sample efficiency and successfully perform zero-shot adaptation with the help of pre-trained IA in diverse online and offline environments, validating the efficacy of learning a transferable communicative representation.
Symbolic regression (SR) aims to recover compact and interpretable mathematical expressions from data. Genetic programming (GP) directly searches over symbolic structures, but its population dynamics can make it difficult to reliably preserve and accumulate useful subexpressions. In contrast, reinforcement learning (RL)-based SR has shown strong empirical performance, suggesting that learned sampling dynamics may capture useful regularities in symbolic search. Motivated by this contrast, we analyze expressions sampled during RL training and identify a recurring pattern, termed progressive subexpression reuse, where useful simple subexpressions emerge early, become increasingly frequent, and support the formation of more complex structures. Based on this observation, we propose Reinforcement Genetic Programming (RGP), a purely GP-based and non-RL framework that explicitly realizes a stage-wise retention--reintroduction loop through a dynamically maintained subexpression pool and pool-guided population initialization. Experiments on standard SR benchmarks show that RGP matches or outperforms strong baselines without RL policy training, suggesting that progressive subexpression reuse is an effective mechanism for SR search. We release our code at https://github.com/wuxiangdong586-max/RGP.
Identifying cancer driver genes (CDGs) requires integrating heterogeneous biological networks, each encoding distinct mechanistic insights into tumorigenesis. Existing multi-network methods typically fuse views at the feature level or enforce uniform representations, which obscures network-specific signals and limits interpretability. To address this, we propose LogicFusion, a novel differentiable framework that treats each biological network as an independent probabilistic evidence source and explicitly learns how to combine them using logical reasoning. Specifically, LogicFusion maps the output of each view into probabilistic space as independent evidence, and learns a set of logical rules in a differentiable manner. These rules adaptively select and fuse these view evidences through logical conjunction (AND) and disjunction (OR). The final prediction is then obtained via gene-specific weighting. LogicFusion enables both high-accuracy CDG identification and interpretable attribution of predictions to specific networks. Experiments on pan-cancer and cancer-specific datasets show that LogicFusion consistently outperforms state-of-the-art methods, while providing transparency in multi-view genomic reasoning through learned logical rules.
We investigate the interaction between quantization and pruning in the compression of open-weight small language models (SLMs), as these techniques are frequently combined in practice without a clear understanding of whether their effects are antagonistic, synergistic or additive. Prior work has reported ordering effects and studied pruning or quantization in isolation, but rarely quantifies their interaction across models and evaluation metrics. We introduce an interaction coefficient that isolates non-additive effects and apply it to Falcon3-1B-Base and LLaMA-3.2-1B, with a limited study on Qwen2.5-1.5B. Our experiments cover four hybrid application orders, multiple pruning sparsities and 4/8-bit quantization. We find non-monotonic interactions: memory savings are dominated by 4-bit quantization, while pruning alone provides limited benefits. Hybrids exhibit the greatest effects at global unstructured pruning sparsity s ∈ {0.2, 0.3} while getting close to additivity at s=0.5. At low sparsity, truthfulness improves, while reasoning losses diminish at higher sparsity. Future work will extend interaction analysis to hardware-aware sparse kernels and larger models.
The knowing--doing gap, the mismatch between ideas articulated during model reasoning and the realized creative artifact, remains a fundamental challenge in creative AI and persists in LLM-based artistic creation. This paper responds to this gap by introducing a systematic and interpretable framework for examining how user intent is articulated during model reasoning and selectively realized, or lost, during action in LLMs, using symbolic music composition as an analytical lens. We present a realizational process theory that formalizes creative generation and localizes the knowing--doing gap at the realization stage. We instantiate this theory with Music Atelier (Mutelier), an LLM-as-a-Judge framework that operationalizes idea-level realization analysis and makes the gap observable and analyzable in practice. Across diverse evaluation settings, we show that Mutelier reveals reliable, previously invisible failure modes in which intent-aligned ideas are articulated during reasoning but fail to materialize in the final artifact. By reframing artistic creation as a realizational process, this work provides a principled process-level foundation for understanding how knowing--doing gaps emerge in machine creativity.
Contrastive learning has become a dominant paradigm for learning time series representations from large-scale unlabeled data. However, current methods are often adapted from computer vision and rely on random time-domain augmentations (e.g., jittering and cropping). Such augmentations can unpredictably disrupt the natural frequency structure of signals, leading to representations failing to capture crucial patterns in the data. To address this, we propose a framework of Frequency-Aware Augmentation and Alignment for Time Series Contrastive Learning (FACL), which comprises two key innovations. First, FACL employs a novel frequency-structured augmentation mechanism based on wavelet transforms. The mechanism constructs controlled and interpretable contrastive views by the structured attenuation and recombination of specific wavelet components. Second, FACL introduces a multi-level contrastive objective that incorporates a subspace alignment strategy. This objective explicitly aligns representations within their corresponding frequency subspaces. Experiments across six forecasting and four classification benchmarks show that FACL achieves superior performance compared to recent baselines. Ablation studies and model analysis highlight the contribution of each component in FACL. Furthermore, low-sample semi-supervised learning experiments confirm the robustness and generalization of FACL.
We present Magnol.AI Copilot, an extension of the Magnol.AI digital biomarker platform that integrates multimodal large language models (LLMs) to transform digital health technology (DHT) trial dashboards into conversational systems. Copilot augments the platform with a multi-agent orchestration layer and vision-enabled LLMs that interpret visualizations, tabular summaries, and textual metadata. The system enables natural language queries and automatic generation of contextual insights, allowing researchers to interact with wearable data through dialogue rather than static inspection. A case study with an actigraphy device demonstrates Copilot’s ability to identify nightly compliance gaps and provide contextual explanations, reducing cognitive load compared to manual dashboard review. This work presents a novel integration of IoMT infrastructure with multimodal LLMs, advancing digital biomarker research toward conversational and accessible DHT trial platforms.
Chronic diseases have become the leading cause of death worldwide, a challenge intensified by strained medical resources and an aging population. Individually, patients often struggle to interpret early signs of deterioration or maintain adherence to care plans. In this paper, we introduce VitalDiagnosis, an LLM-driven ecosystem designed to shift chronic disease management from passive monitoring to proactive, interactive engagement. By integrating continuous data from wearable devices with the reasoning capabilities of LLMs, the system addresses both acute health anomalies and routine adherence. It analyzes triggers through context-aware inquiries, produces provisional insights within a collaborative patient–clinician workflow, and offers personalized guidance. This approach aims to promote a more proactive and cooperative care paradigm, with the potential to enhance patient self-management and reduce avoidable clinical workload.