Demystifying On-Policy Distillation: Roles, Pathologies, and Regulations

On-policy distillation (OPD) has become a key paradigm in LLM post-training, yet its training dynamics remain poorly understood. We present a systematic study examining the role, pathologies, and regulations of OPD. We first clarify the role of OPD as an exploration catalyst: it steers the student toward correct reasoning paths via dense token-level guidance, without expanding capability ceiling. We confirm this by showing that prompt diversity matters more than per-problem sampling numbers, and critically, that the effectiveness of OPD hinges entirely on the quality of its guiding signal. This dependency exposes two pathologies that derail exploration. The Student-Teacher Mismatch occurs when a large teacher-student distributional gap causes the guiding signal to misalign with task correctness, steering exploration in counterproductive directions. Length Exploitation arises when the aggregated token-level objective creates length-dependent shortcuts, allowing the student to game the reward landscape through response truncation or redundant padding, exploring degenerate length modes rather than reasoning strategies. To tame these pathologies, we investigate lightweight signal regulations: advantage clipping and log-scale compression, ensuring exploration is guided by faithful signals. Experiments across seven benchmarks demonstrate that these regulations alleviate length exploitation and enable effective distillation, stably surpassing OPD variants and RLVR baselines, thereby confirming that well-regulated signal quality, rather than mere teacher scale, governs successful exploration in OPD.

Just-In-Time Scene Graph Growth: Combating Perceptual Saturation in Long-Horizon Robotics

While 3D Scene Graphs (3DSGs) provide crucial structured representations for embodied agents, conventional Ahead-of-Time, build-everything-then-filter pipelines conflict with the real-time, low-latency demands of edge platforms, inducing a perceptual saturation effect via severe observation redundancy. To resolve this, we present JITOMA (Just-In-Time On-demand Memory Activation), a closed-loop framework that unifies task reasoning, perception, and memory into a just-in-time growth process. Instead of exhaustively mapping the entire environment, JITOMA leverages a top-down task heatmap at the frontend to filter continuous observations, routing minimal streams to maintain a global foundation of low-cost, dormant anchors. Upon a cognitive query, the backend Large Language Model (LLM) parses the robotic intent to dynamically awaken task-relevant anchors, triggering resource-intensive operations -- such as dense node captioning and functional inference -- exclusively within the activated local subgraph. To evaluate these dynamic capabilities and study perceptual saturation trade-offs, we introduce JITOMA-Bench, a comprehensive suite for long-horizon multi-tasking and complex multi-step reasoning. Extensive experiments demonstrate that JITOMA substantially reduces active graph size and captioning latency, while maintaining stable processing time under long-horizon task switching.

VINE: Taming Generative Control Policies for Reinforcement Learning

Flow-matching policies have emerged as an effective policy parameterization for robot learning. They iteratively generate actions from noise, enabling highly expressive modeling of complex and multimodal action distributions. However, prior works observed that scaling these policies with value-gradient reinforcement learning (RL) often leads to training instability. Existing methods attribute this instability to iterative generation and therefore avoid end-to-end value-gradient optimization by sacrificing iterative generation, high expressiveness, or value-gradient optimization. Contrary to prior belief, we show the instability does not stem from iterative generation itself, but from the vanilla sampling strategy originally designed for behavior cloning, which becomes brittle under value-gradient RL. Motivated by this insight, we propose VINE, an RL-oriented sampling method that enables stable end-to-end value-gradient optimization for flow-matching policies. Instead of following a single flow trajectory, VINE reconstructs a new interpolation state at every denoising step, creating a stable differentiable path for value-gradient propagation while remaining compatible with the original flow-matching denoising process. As a result, VINE preserves the expressiveness and iterative generation of flow-matching without sacrificing end-to-end value-gradient optimization. Despite performing end-to-end backpropagation through all ten denoising steps, VINE achieves stable policy improvement and consistently outperforms state-of-the-art RL methods on the OGBench offline RL benchmark and real-world robotic manipulation task. Videos are available on our website: https://agibottech.github.io/vine.

Decision-Weighted Flow Matching for Contextual Stochastic Optimization

Conditional generative models are increasingly used as scenario generators for stochastic optimization, but standard training objectives emphasize uniform distributional fit rather than the downstream decisions induced by generated scenarios. This creates an objective mismatch: errors in statistically common regions may have little effect on decision regret, whereas errors in decision-sensitive regions can substantially change the optimal action. We propose Decision-Weighted Flow Matching (DW-FM), a regret-aligned training framework that preserves the simplicity of standard flow matching while reweighting its velocity-regression objective using decision-sensitive endpoint information. Theoretically, we connect downstream regret to pathwise velocity mismatch through a loss-induced decision discrepancy and an adjoint transport argument, yielding an ideal regret-aligned surrogate and practical endpoint-weighted objectives with regret guarantees. Empirically, we demonstrate the effectiveness of DW-FM on three CVaR-based contextual stochastic optimization benchmarks spanning synthetic portfolio, semi-real financial, and traffic-CVaR tasks, where DW-FM improves downstream regret over standard baselines.

Making Foresight Actionable: Repurposing Representation Alignment in World Action Models

World Action Models (WAMs) offer a promising route for robot manipulation by using video generation models to model future scene evolution before producing control actions. However, our empirical observations reveal a phenomenon: generating plausible visual futures does not always guarantee the extraction of accurate actions. To diagnose this failure, we conduct action-head attention analysis and causal interventions. We find that the action decoder fails to focus on task-relevant interaction regions and remains sensitive to perturbations in task-irrelevant areas. This reveals a representation mismatch: hidden states optimized for visual reconstruction are not inherently organized in a form useful for low-level action control. In this paper, we propose AGRA, an Action-Grounded Representation Alignment objective that regularizes the world-action interface by aligning intermediate video diffusion features with spatially coherent semantic representations from a foundation visual encoder. We evaluate AGRA on real-world manipulation tasks. Experiments show that AGRA makes world model representations more action-grounded: by focusing the action decoder on the correct interaction regions, it improves object localization accuracy and affordance understanding, and makes the policy more robust to perturbations in task-irrelevant regions. As a result, AGRA consistently improves both in-distribution performance and out-of-distribution generalization over the baseline world action model.

Bellman-Taylor Score Decoding for Markov Decision Processes with State-Dependent Feasible Action Sets

Many Markov decision processes (MDPs) in operations research have feasible actions that are state dependent and defined implicitly by various operational constraints. These features make it difficult to use standard deep reinforcement learning (DRL) algorithms, whose action interfaces typically assume either a fixed finite action catalog or a simple Euclidean space. Motivated by a Taylor expansion of the optimal action-value function, we propose Bellman--Taylor score decoding, a framework that moves policy learning to a Euclidean score space while enforcing feasibility through an action decoder. The induced latent-score MDP then can be optimized by standard DRL algorithms without differentiating through the decoder. We provide a performance guarantee showing that the optimality gap of this approach decomposes into a structural approximation error and an algorithmic learning error. Lastly, we apply this framework to a queueing network control problem, where the policy essentially learns a state-dependent index-based dispatching rule. Numerical experiments show near-optimal performance in small instances and considerable improvements over benchmarks in larger systems.

ConMem: Structured Memory-Guided Adaptation in Training-Free Multi-Agent Systems

Recent advances have improved the adaptive capabilities of LLM-based multi-agent systems (MAS) through memory-, skill-, and learning-based approaches, yet these approaches remain challenged by noisy trajectories, insufficient modeling of memory-skill relations, and reliance on additional training or high-quality supervision. To address these limitations, we propose ConMem, a relation-aware and training-free framework that enables efficient multi-agent adaptation through cross-experience coordination. Specifically, ConMem distills historical interaction trajectories into structured memory cards to capture reusable strategies and cues, organizing them into a relation-aware memory graph. At runtime, ConMem retrieves cards according to task needs and coordinates them through the card graph to resolve strategy conflicts and recover their dependencies. Combined, these modules yield structured and relation-aware guidance, enabling robust, lightweight adaptation in multi-agent systems without additional training. Extensive experiments across multiple benchmarks and mainstream MAS architectures show consistent gains over existing memory architectures, with improved inference-time efficiency through pruning more than 50% of expanded candidates and reducing planning overhead by over 80%. Our codes are available at https://anonymous.4open.science/r/ConMemCode

Unveiling the Unknown: Open Vocabulary Object Detection with Scene Graphs

Open-vocabulary object detection seeks to identify novel object categories that were not part of the training data. Many knowledge distillation-based approaches have shown promising performance by transferring knowledge from pre-trained vision-language models to object detection. However, these methods often overlook structured, image-specific relationships between objects, such as interactions and spatial arrangements. This oversight can significantly restrict the effectiveness of detecting novel categories. To address this issue, we propose a Scene-guided Relational Modeling detection framework. This framework utilizes scene graphs to capture structured semantic and spatial relationships between candidate regions and their contextual objects. It explicitly models interactions among neighboring regions and incorporates a Relation Attention Module to implicitly amplify the key relational cues extracted from the scene graph. Furthermore, we present a scene-based textual alignment branch that distills category knowledge from captions to guide relational alignment. This approach facilitates a seamless integration of visual relations with semantic information for enhanced detection performance. Comprehensive experiments show that our model achieves superior performance compared to other OVOD methods, improving the AP for novel categories on COCO and LVIS datasets.

Detect by Yourself: Self-Designing Agentic Workflows for Few-Shot Graph Anomaly Detection

Graph anomaly detection aims to identify anomaly nodes in attributed graphs and plays an important role in real-world applications. However, existing graph anomaly detection methods still face two key challenges: 1) fixed pipelines, which restrict their adaptability across different graph tasks under limited supervision; 2) weak evidence, which prevents them from explicitly incorporating contextual and structural anomaly signals into the detection process. In this paper, we propose a novel framework, self-designing agentic workflows for few-shot graph anomaly detection (SignGAD). Specifically, we propose a novel paradigm that reformulates graph anomaly detection task from training a fixed anomaly detector to designing task-conditioned detection workflows. By constructing detection workflows, SignGAD selects suitable graph encodings and detector designs to exploit task-specific anomaly evidence. Meanwhile, we introduce a guarded final refit strategy to refine the selected workflow by calibrating refit acceptance, enhancing reliability under limited supervision. Extensive experiments conducted on several real-world datasets demonstrate that SignGAD achieves strong performance against state-of-the-art methods, highlighting its effectiveness on graph anomaly detection tasks.

DenseScout: Algorithm-System Co-design for Budgeted Tiny Object Selection on Edge Platforms

Deploying tiny object perception on edge platforms is challenging because practical systems must satisfy both strict compute budgets and end-to-end latency constraints. A common strategy is to first select a small number of candidate patches from a high-resolution image and then apply downstream processing only to the selected regions. However, existing detector-based frontends are not well aligned with this setting: strong offline detection accuracy does not necessarily yield effective low-budget patch prioritization, nor does it guarantee usable performance once transport and inference delays are considered. In this work, we study budgeted tiny object selection on edge platforms from a joint algorithm--system perspective. We present DenseScout, a lightweight dense-response selector with only 1.01M parameters, which directly ranks candidate patch locations from a high-resolution scene via a lightweight proxy input and is better aligned with low-budget tiny-object prioritization than detector-style frontends. To bridge offline selector quality and deployable utility, we further develop a transport-aware runtime realization on heterogeneous edge devices and adopt QoS-constrained recall, which counts a target as successfully perceived only if it is covered by the selected regions and the end-to-end processing finishes before the deadline. Experiments show that DenseScout consistently outperforms detector-based baselines in offline budgeted patch-selection evaluation, especially in low-budget regimes, while cross-platform results on RK3588 and Jetson Orin NX show that deployable performance depends jointly on selector quality and runtime realization efficiency. These results suggest that edge tiny object perception should be optimized as an algorithm--system co-design problem rather than as isolated model selection.