Flow-matching policies hold great promise for reinforcement learning (RL) by capturing complex, multi-modal action distributions. However, their practical application is often hindered by prohibitive inference latency and ineffective online exploration. Although recent works have employed one-step distillation for fast inference, the structure of the initial noise distribution remains an overlooked factor that presents significant untapped potential. This overlooked factor, along with the challenge of controlling policy stochasticity, constitutes two critical areas for advancing distilled flow-matching policies. To overcome these limitations, we propose GoldenStart (GS-flow), a policy distillation method with Q-guided priors and explicit entropy control. Instead of initializing generation from uninformed noise, we introduce a Q-guided prior modeled by a conditional VAE. This state-conditioned prior repositions the starting points of the one-step generation process into high-Q regions, effectively providing a ``golden start'' that shortcuts the policy to promising actions. Furthermore, for effective online exploration, we enable our distilled actor to output a stochastic distribution instead of a deterministic point. This is governed by entropy regularization, allowing the policy to shift from pure exploitation to principled exploration. Our integrated framework demonstrates that by designing the generative startpoint and explicitly controlling policy entropy, it is possible to achieve efficient and exploratory policies, bridging the generative models and the practical actor-critic methods. We conduct extensive experiments on offline and online continuous control benchmarks, where our method significantly outperforms prior state-of-the-art approaches.

Fine-tuning LLMs to inject new knowledge faces a critical challenge: LLMs can quickly memorize new facts, yet fail to use them for downstream reasoning tasks. We formalize this failure as the Knowing–Using Gap, characterized by an accuracy gap and a temporal lag between memorization and generalization. To understand this phenomenon, we fine-tune LLMs with unseen knowledge and monitor the spatial permeation dynamics of the knowledge internally using a novel intervention technique called self-patching. Self-patching identifies activation locations where relocating representations substantially improves failed generalization cases. These results are consistent with a knowledge-circuit misalignment hypothesis: memorized representations can exist internally but may not be routed to computation-effective layers. To demonstrate the practicality of this diagnostic finding, we design a simple heuristic strategy which recovers 58–75% of the oracle headroom in generalization failure. Experiments are done cross-domain for the robustness of this finding.

The integration of Large Language Models (LLMs) into the natural sciences marks a shift from specialized discriminative models to generalist generative agents. However, adapting LLMs to chemistry presents a core challenge: the mismatch between the linear, ambiguous nature of natural language and the discrete, topological, and physical nature of molecules. This review provides a systematic framework for bridging this gap, organizing recent advances through the dual lenses of representation alignment and cognitive mechanism. First, we examine how molecular modalities are adapted to the transformer architecture, progressing from Linguistic Linearization, which optimizes symbolic grammars such as SMILES and SELFIES, to Topological Perception, which injects graph inductive biases through adapters and projectors, and finally to Geometric Grounding, which encodes 3D physical reality and equivariance. Second, we analyze how LLMs acquire and deploy chemical knowledge through three complementary pathways: Internalization, which cultivates parametric intuition through pre-training and alignment; Externalization, which anchors generation to physical reality via retrieval and tool orchestration; and Reasoning, which we further decompose into four strategies spanning contextual analogy, sequential deduction, strategic planning, and introspective refinement. Building on this dual-lens framework, we map downstream capabilities onto a four-level task hierarchy spanning Semantic Translation, Predictive Inference, Constrained Generation, and Autonomous Discovery, and we identify three coupled challenges at the representation-cognition interface, outlining a research roadmap toward chemically grounded AI systems.

Large language models (LLMs) excel at complex tasks with advances in reasoning capabilities. However, existing reward mechanisms remain tightly coupled to final correctness and pay little attention to the underlying reasoning process: trajectories with sound reasoning but wrong answers receive low credit, while lucky guesses with flawed logic may be highly rewarded, affecting reasoning generalization. From a causal perspective, we interpret multi-candidate reasoning for a fixed question as a family of counterfactual experiments with theoretical supports. Building on this, we propose Group Causal Counterfactual Policy Optimization to explicitly train LLMs to learn generalizable reasoning patterns. It proposes an episodic causal counterfactual reward that jointly captures (i) robustness, encouraging the answer distribution induced by a reasoning step to remain stable under counterfactual perturbations; and (ii) effectiveness, enforcing sufficient variability so that the learned reasoning strategy can transfer across questions. We then construct token-level advantages from this reward and optimize the policy, encouraging LLMs to favor reasoning patterns that are process-valid and counterfactually robust. Extensive experiments on diverse benchmarks demonstrate its advantages.

Recent years have witnessed the widespread application of artificial intelligence technologies in major hospitals for accurate diagnoses and effective treatments. Existing methods for personalizing treatment mainly focus on recommending drug combinations for patients based on their profiles and symptoms. However, the generation of prescriptions, including drugs and the corresponding doses, based on context information during hospitalization has been largely ignored. Therefore, in this paper, we propose a multisource, context-aware prescription generation model, namely hierarchical transformer-based prescription generation (HTPG), to solve this problem. Specifically, we first formulate prescription generation as a sequence generation task, where each prescription is regarded as a sequential composition of drug-dose pairs. Then, we propose to model the multisource context information of patients and generate prescriptions based on a hierarchical transformer structure. Extensive experiments on a public real-world dataset demonstrate the effectiveness of our HTPG model compared with several competitive baseline methods.

Recent advances in pretrained language models (PLMs) have significantly improved conversational recommender systems (CRS), enabling more fluent and context-aware interactions. To further enhance accuracy and mitigate hallucination, many methods integrate PLMs with knowledge graphs (KGs), but face key challenges: failing to fully exploit PLM reasoning over graph relationships, indiscriminately incorporating retrieved knowledge without context filtering, and neglecting collaborative preferences in multi-turn dialogues. To this end, we propose PCRS-TKA, a prompt-based framework employing retrieval-augmented generation to integrate PLMs with KGs. PCRS-TKA constructs dialogue-specific knowledge trees from KGs and serializes them into texts, enabling structure-aware reasoning while capturing rich entity semantics. Our approach selectively filters context-relevant knowledge and explicitly models collaborative preferences using specialized supervision signals. A semantic alignment module harmonizes heterogeneous inputs, reducing noise and enhancing accuracy. Extensive experiments demonstrate that PCRS-TKA consistently outperforms all baselines in both recommendation and conversational quality.

On-policy distillation (OPD) provides dense token-level supervision by asking a teacher to score student-generated rollouts. However, when the student drifts into an unrecoverable prefix, the teacher may locally agree with the degraded state, producing low reverse KL but little corrective training signal. We identify this persistent regime as a low-KL agreement trap. Further analyses show that tokens during and after such traps produce less useful supervision signals. We propose KAT (KL Agreement Trap Termination), an online OPD termination rule that detects persistent low-KL agreement with a dynamic training-adaptive threshold. By filtering weak supervision from degenerate agreement, KAT improves avg@k accuracy by 2.66

Large Language Models (LLMs) have shown great potential for enhancing recommender systems through their extensive world knowledge and reasoning capabilities. However, effectively translating these semantic signals into traditional collaborative embeddings remains an open challenge. Existing approaches typically fall into two extremes: direct inference methods are computationally prohibitive for large-scale retrieval, while embedding-based methods primarily focus on unilateral feature augmentation rather than holistic collaborative signal enhancement. To bridge this gap, we propose Topology-Augmented Graph Collaborative Filtering (TAGCF), a novel framework that transforms semantic knowledge into topological connectivity. Unlike existing approaches that depend on textual features or direct interaction synthesis, TAGCF employs LLMs to infer interaction intents and underlying causal relationships from user-item pairs, representing these insights as intermediate attribute nodes within an enriched User-Attribute-Item (U-A-I) graph. Furthermore, to effectively model the heterogeneous relations in this augmented structure, we propose Adaptive Relation-weighted Graph Convolution (ARGC), which employs relation-specific prediction networks to dynamically estimate the importance of each relation type. Extensive experiments across multiple benchmark datasets and CF backbones demonstrate consistent improvements, with comprehensive evaluations including cold-start scenarios validating the effectiveness and robustness of our framework. All code will be made publicly available. For anonymous review, our code is available at the following anonymous link: https://anonymous.4open.science/r/AGCF-2441353190/.

Multimodal large language models (MLLMs) have become a key interface for visual reasoning and grounded question answering, yet they remain vulnerable to visual hallucinations, where generated responses contradict image content or mention nonexistent objects. A central challenge is that hallucination is not always caused by a simple lack of visual attention: the model may still assign substantial attention mass to image tokens while internally drifting toward an incorrect answer. In this paper, we show that the high-frequency structure of visual attention, measured by layer-wise Laplacian energy, reveals both the layer where hallucinated preferences emerge and the layer where the ground-truth answer transiently recovers. Building on this finding, we propose LaSCD (Laplacian-Spectral Contrastive Decoding), a training-free decoding strategy that selects informative layers via Laplacian energy and remaps next-token logits in closed form. Experiments on hallucination and general multimodal benchmarks show that LaSCD consistently reduces hallucination while preserving general capabilities, highlighting its potential as a faithful decoding paradigm. The code is available at https://github.com/macovaseas/LaSCD.

Large language models (LLMs) often suffer from catastrophic forgetting in continual learning: after learning new tasks sequentially, they perform worse on earlier tasks. Existing methods mitigate catastrophic forgetting by data replay, parameter freezing, or regularization. However, these methods lack semantic awareness of internal knowledge distribution in LLMs. As a result, they cannot distinguish parameters that should be preserved or updated. We propose an attribution-guided continual fine-tuning framework for LLMs. Our method estimates task-specific, element-wise parameter importance in each Transformer layer and uses these scores to modulate gradients. Parameters important to previous tasks receive smaller updates, while less relevant ones remain plastic for learning new tasks. Experiments on continual learning benchmarks show that our method consistently outperforms baselines, achieving better retention of old tasks while maintaining competitive performance on new tasks.

Large pretrained vision models have substantially improved appearance-based 3D shape retrieval, but they still confuse shapes that look similar while differing in geometry. Although geometry-aware features can reduce these errors, naive fusion of geometry and appearance may hurt retrieval when the two modalities are already well aligned. We propose GATE-3D, a lightweight query-adaptive reranking method that incorporates geometry without retraining the retrieval backbone. For each query, GATE-3D predicts how much a geometry-aware score should adjust the appearance-based ranking using features that capture disagreement between the two modalities. This selective design lets geometry contribute where it helps and stay silent where it would hurt. Experiments on three open-set 3D retrieval benchmarks show that GATE-3D improves over appearance-only retrieval and is more robust than always-on fusion. On the primary benchmark, it improves mAP@10 by 2.00 points over appearance-only retrieval (p=0.041); it also improves leave-one-category-out generalization and reduces geometric false positives by 10.8

Multi-modal methods establish comprehensive superiority over uni-modal methods. However, the imbalanced contributions of different modalities to task-dependent predictions constantly degrade the discriminative performance of canonical multi-modal methods. Based on the contribution to task-dependent predictions, modalities can be identified as predominant and auxiliary modalities. Benchmark methods raise a tractable solution: augmenting the auxiliary modality with a minor contribution during training. However, our empirical explorations challenge the fundamental idea behind such behavior, and we further conclude that benchmark approaches suffer from certain defects: insufficient theoretical interpretability and limited exploration capability of discriminative knowledge. To this end, we revisit multi-modal representation learning from a causal perspective and build the Structural Causal Model. Following the empirical explorations, we determine to capture the true causality between the discriminative knowledge of predominant modality and predictive label while considering the auxiliary modality. Thus, we introduce the $\beta$-generalization front-door criterion. Furthermore, we propose a novel network for sufficiently exploring multi-modal discriminative knowledge. Rigorous theoretical analyses and various empirical evaluations are provided to support the effectiveness of the innate mechanism behind our proposed method.

Embodied navigation is moving toward long-horizon missions, yet existing long-horizon benchmarks are largely acoustically silent, and audio-visual navigation tasks typically focus on a single goal. We introduce LH-AVLN, a benchmark for Long-Horizon Audio-Visual-Language Navigation that combines multi-goal mission execution, heterogeneous goal specifications, and persistent spatialized acoustic cues. In LH-AVLN, an agent receives a global mission of two to four goals specified by category, language description, or reference image, and navigates with RGB-D observations, pose, and binaural audio in indoor 3D environments. The benchmark supports both ordered and unordered missions, where alternating goal-associated sounds can guide non-line-of-sight search but may also become distractors as mission progress changes. We further develop PAG-Nav, a training-free reference agent that maintains a temporal uniform semantic map and performs progressive goal-state planning, using sound for search while reserving completion for visual-semantic verification. Experiments show that existing vision-language, memory-based, and audio-visual agents struggle to complete full LH-AVLN missions, and that PAG-Nav provides a stronger diagnostic baseline while leaving substantial room for future progress.

Training stability remains a critical bottleneck for Group Relative Policy Optimization (GRPO), often manifesting as a trade-off between reasoning plasticity and general capability retention. We identify a root cause as the geometric conflict between plasticity and stability gradients, which leads to destructive interference. Crucially, we argue that deterministic projection methods are suboptimal for GRPO as they overlook the intrinsic stochasticity of group-based gradient estimates. To address this, we propose Probabilistic Conflict Resolution (PCR), a Bayesian framework that models gradients as random variables. PCR dynamically arbitrates conflicts via an uncertainty-aware “soft projection” mechanism, optimizing the signal-to-noise ratio. Extensive experiments demonstrate that PCR significantly smooths the training trajectory and achieves superior performance in various reasoning tasks.

Foundation-model post-training usually relies on human labels, preference data, stronger teachers, or executable verifiers. We study Unsupervised Post-Training (UPT): update-bearing adaptation on unlabeled inputs whose learning signal is derived from same-lineage model artifacts rather than an external oracle. We catalog 80 strict UPT methods and organize them by the object that supplies the update signal: a prediction statistic, a sample relation, a self-generated target, or an internal evaluator. Beyond inventory, we show how the choice of internal signal and task structure determines whether post-training improves the model or recursively amplifies error. An orthogonal Input Visibility × Update Persistence view maps deployment regimes and defines a unified framework for UPT selection and evaluation.

Video contrastive learning (V-CL) has emerged as a popular framework for unsupervised video representation learning, demonstrating strong results in tasks such as action classification and detection. Yet, to harness these benefits, it is critical for the learned representations to fully capture both static and dynamic semantics. However, our experiments show that existing V-CL methods fail to effectively learn either type of feature. Through a rigorous theoretical analysis based on the Structural Causal Model and gradient update, we find that in a given dataset, certain static semantics consistently co-occur with specific dynamic semantics. This phenomenon creates spurious correlations between static and dynamic semantics in the dataset. However, existing V-CL methods do not differentiate static and dynamic similarities when computing sample similarity. As a result, learning only one type of semantics is sufficient for the model to minimize the contrastive loss. Ultimately, this causes the V-CL pre-training process to prioritize learning the easier-to-learn semantics. To address this limitation, we propose Bi-level Optimization with Decoupling for Video Contrastive Learning. (BOD-VCL). In BOD-VCL, we model videos as linear dynamical systems based on Koopman theory. In this system, all frame-to-frame transitions are represented by a linear Koopman operator. By performing eigen-decomposition on this operator, we can separate time-variant and time-invariant components of semantics, which allows us to explicitly separate the static and dynamic semantics in the video. By modeling static and dynamic similarity separately, both types of semantics can be fully exploited during the V-CL training process. BOD-VCL can be seamlessly integrated into existing V-CL frameworks, and experimental results highlight the significant improvements achieved by our method. The source code is released at https://github.com/ZeenSong/Video_contrastive .

Federated Learning (FL) has emerged as a promising privacy-preserving collaborative model training paradigm without sharing raw data. However, recent studies have revealed that private information can still be leaked through shared gradient information and attacked by Gradient Inversion Attacks (GIA). While many GIA methods have been proposed, a detailed analysis, evaluation, and summary of these methods are still lacking. Although various survey papers summarize existing privacy attacks in FL, few studies have conducted extensive experiments to unveil the effectiveness of GIA and their associated limiting factors in this context. To fill this gap, we first undertake a systematic review of GIA and categorize existing methods into three types, i.e., optimization-based GIA (OP-GIA), generation-based GIA (GEN-GIA), and analytics-based GIA (ANA-GIA). Then, we comprehensively analyze and evaluate the three types of GIA in FL, providing insights into the factors that influence their performance, practicality, and potential threats. Our findings indicate that OP-GIA is the most practical attack setting despite its unsatisfactory performance, while GEN-GIA has many dependencies and ANA-GIA is easily detectable, making them both impractical. Finally, we offer a three-stage defense pipeline to users when designing FL frameworks and protocols for better privacy protection and share some future research directions from the perspectives of attackers and defenders that we believe should be pursued. We hope that our study can help researchers design more robust FL frameworks to defend against these attacks.

Generalized Category Discovery (GCD) aims to identify both known and novel categories from partially labeled data, reflecting more realistic open-world learning scenarios. However, most existing methods rely solely on one-hot discriminative supervision, leading to overfitting on seen classes and poor generalization to unseen ones. Recent advances introduce large language models (LLMs) to incorporate external semantics, yet they often suffer from semantic–label misalignment and weak semantic integration during training. We propose GenDis, a Generative–Discriminative Dual-View Co-Training framework that unifies discriminative classification and semantic label generation within an LLM. Discriminative pseudo-labels guide the formation of a separable generative latent space, enabling semantically meaningful supervision for novel classes. To ensure consistency between the two views, we employ Canonical Correlation Analysis (CCA)-based alignment and a curriculum-guided, dispersion-aware pseudo-labeling strategy for iterative refinement. Extensive experiments on five GCD benchmarks demonstrate that GenDis substantially outperforms prior methods, validating the effectiveness of dual-view co-training with semantically enriched supervision. The anonymized repository is available at https://anonymous.4open.science/r/GenDis.