General ML · reinforcement learning

Robust perception and reasoning require consistency across sensory modalities. Yet, current multimodal models often violate this principle, yielding contradictory predictions for visual versus textual representations of the same input. Rather than masking these failures with standard voting mechanisms--which amplify systematic biases--we demonstrate that cross-modal inconsistency provides a rich, natural signal for learning. We introduce R-C2, a reinforcement learning framework that resolves internal conflicts by enforcing cross-modal cycle consistency. By requiring a model to perform backward inference, switches modalities, and reliably reconstruct the answer via forward inference, we establish a dense, label-free reward. This cyclic constraint forces the model to autonomously align its representations. Optimizing for this structure mitigates modality-specific errors and improves reasoning accuracy by up to 7.6 points. Our results suggest that advanced reasoning emerges not just from scaling data, but from enforcing a structurally consistent understanding of the world.

Spatio-Temporal Video Grounding (STVG) requires models to localize objects both spatially and temporally. Despite recent progress, existing methods struggle with complex and fine-grained spatial semantics in language descriptions, leading to error propagation from temporal to spatial grounding stages. We attribute this limitation to the lack of iterative refinement between temporal and spatial predictions. To address these challenges, we propose SARL-STG, the first RL-based framework for STVG. It progressively refines spatio-temporal grounding through multi-stage training with dynamic interaction between temporal and spatial modules. Specifically, SARL-STG contains: (1) a unified architecture that seamlessly integrates a pretrained MLLM for temporal reasoning with an open-vocabulary detector for spatial localization, (2) a hierarchical RL training strategy that progresses from coarse temporal to fine-grained spatio-temporal optimization, and (3) a spatial knowledge-injected reward mechanism that uses spatial grounding quality as discriminative signals for temporal refinement. To facilitate training at scale, we also construct STVG-Wild, a large-scale dataset with diverse spatio-temporal annotations. Experiments demonstrate that our method achieves state-of-the-art performance on multiple benchmarks (HCSTVG, VidSTG, Charades-STA, etc.), significantly reducing error accumulation and enhancing both temporal and spatial grounding accuracy.

The primary obstacle for applying reinforcement learning (RL) to real-world robotics is the design of effective reward functions. While recently learning-based Process Reward Models (PRMs) are a promising direction, they are often hindered by two fundamental limitations: their reward models lack step-aware understanding and rely on single-view perception, leading to unreliable assessments of fine-grained manipulation progress; and their reward shaping procedures are theoretically unsound, often inducing a semantic trap that misguides policy optimization. To address these, we introduce Dopamine-Reward, a novel reward modeling method for learning a general-purpose, step-aware process reward model from multi-view inputs. At its core is our General Reward Model (GRM), trained on a vast 3,400+ hour dataset, which leverages Step-wise Reward Discretization for structural understanding and Multi-Perspective Reward Fusion to overcome perceptual limitations. Building upon Dopamine-Reward, we propose Dopamine-RL, a robust policy learning framework that employs a theoretically-sound Policy-Invariant Reward Shaping method, which enables the agent to leverage dense rewards for efficient self-improvement without altering the optimal policy, thereby fundamentally avoiding the semantic trap. Extensive experiments across diverse simulated and real-world tasks validate our approach. GRM achieves state-of-the-art accuracy in reward assessment, and Dopamine-RL built on GRM significantly improves policy learning efficiency. For instance, after GRM is adapted to a new task in a one-shot manner from a single expert trajectory, the resulting reward model enables Dopamine-RL to improve the policy from near-zero to 95% success with only 150 online rollouts (approximately 1 hour of real robot interaction), while retaining strong generalization across tasks.

Reinforcement fine-tuning (RFT) is increasingly used to strengthen the reasoning abilities of large models, yet its effectiveness is bound by how training data are selected and used. Most data-centric RFT methods rely on static or heuristic sample selection, implicitly assuming a sample's value is fixed over training. This overlooks the non-stationary dynamics of policy learning and can lead to suboptimal updates. We propose Dynamic Important Example Mining (DIEM), a principled and fully automated framework that makes data utilization adaptive throughout RFT. DIEM integrates two components into each optimization step: (i) a gradient-alignment importance estimator that efficiently approximates each sample's marginal contribution to policy improvement; and (ii) a constrained batch reweighting scheme that maximizes aggregate utility while preserving the update's gradient magnitude to stabilize optimization. Across several reasoning benchmarks, DIEM consistently outperforms strong static and dynamic baselines. The code will be released via https://github.com/hrtan/DIEM.

Pedestrian trajectory prediction is crucial for applications such as autonomous driving and social robots. Recently, language model (LM)-based trajectory prediction has offered both prediction accuracy and interpretability. However, the L2 loss commonly used in trajectory prediction cannot be directly applied to LM optimization, resulting in degraded prediction performance. Moreover, current LM-based trajectory prediction methods lack explicit expressions of social interactions, and their scene descriptions are overly simplistic, making it challenging to impose practical scene constraints. To address these issues, we propose Write-to-Walk (W2W). First, we convert observed trajectories and interaction cues (companion/following/obstacle) into parsable textual prompts, so that interaction semantics are expressed more explicitly in the model input. Afterward, a T5-Small backbone is trained in a two-stage manner: (1) Full-parameter supervised fine-tuning with cross-entropy loss for language learning, enabling formatted question answering; (2) Reinforcement Learning (RL) to optimize W2W, where a reward function combining ADE error and off-road penalties strengthens scene constraints, producing future trajectories consistent with the scene context and further improving prediction accuracy. Experiments on benchmark datasets (ETH/UCY and SDD) demonstrate that W2W remains competitive with recent LM-based prediction methods and strong trajectory prediction baselines on ADE/FDE. The interpretability of LMs further supports W2W's deployment in safety-critical domains such as autonomous driving.

Reading text from images or scanned documents via OCR models has been a longstanding focus of researchers. Intuitively, text reading is perceived as a straightforward perceptual task, and existing work primarily focuses on constructing enriched data engineering to enhance SFT capabilities. In this work, we observe that even advanced OCR models exhibit significantly higher entropy in formatted text (e.g., formula, table) compared to plain text, often by an order of magnitude. These statistical patterns reveal that advanced OCR models struggle with high output uncertainty when dealing with format sensitive document, suggesting that reasoning over diverse reading pathways may improve OCR performance. To address this, we propose Format Decoupled Reinforcement Learning (FD-RL) , which leverages high-entropy patterns for targeted optimization. Our approach employs entropy-based data filtration strategy to identify format-intensive instances, and adopt format decoupled rewards tailored to different format types, enabling format-level validation rather than token-level memorization. FD-RL achieves an average score of 90.41 on OmniDocBench, demonstrating highly competitive performance among end-to-end models on this highly popular benchmark. More importantly, we conduct comprehensive ablation studies over data, training, filtering, and rewarding strategies, thoroughly validating their effectiveness.

Standard supervised training for deepfake detection treats all samples with uniform importance, which can be suboptimal for learning robust and generalizable features. In this work, we propose a novel Tutor-Student Reinforcement Learning (TSRL) framework to dynamically optimize the training curriculum. Our method models the training process as a Markov Decision Process where a "Tutor" agent learns to guide a "Student" (the deepfake detector). The Tutor, implemented as a Proximal Policy Optimization (PPO) agent, observes a rich state representation for each training sample, encapsulating not only its visual features but also its historical learning dynamics, such as EMA loss and forgetting counts. Based on this state, the Tutor takes an action by assigning a continuous weight (0-1) to the sample's loss, thereby dynamically re-weighting the training batch. The Tutor is rewarded based on the Student's immediate performance change, specifically rewarding transitions from incorrect to correct predictions. This strategy encourages the Tutor to learn a curriculum that prioritizes high-value samples, such as hard-but-learnable examples, leading to a more efficient and effective training process. We demonstrate that this adaptive curriculum improves the Student's generalization capabilities against unseen manipulation techniques compared to traditional training methods. Code is available at https://github.com/wannac1/TSRL

We present ReMoT, a unified training paradigm to systematically address the fundamental shortcomings of VLMs in spatio-temporal consistency--a critical failure point in navigation, robotics, and autonomous driving. ReMoT integrates two core components: (i) A rule-based automatic framework that generates ReMoT-16K, a large-scale (16.5K triplets) motion-contrast dataset derived from video meta-annotations, surpassing costly manual or model-based generation. (ii) Group Relative Policy Optimization, which we empirically validate, yields optimal performance and data efficiency for learning this contrastive reasoning, far exceeding standard Supervised Fine-Tuning. We also construct the first benchmark for fine-grained motion contrast triplets to measure a VLM's discrimination of subtle motion attributes (e.g., opposing directions). The resulting model achieves SOTA performance on our new benchmark and multiple standard VLM benchmarks, culminating in a remarkable 25.1 performance leap on spatio-temporal reasoning tasks.

Pre-training on large-scale videos to improve reinforcement learning efficiency is promising yet remains challenging. Existing methods typically treat the agent as an indivisible entity, modeling motion patterns globally. Such global modeling is tightly coupled with the morphology, hindering transfer across domains. In contrast, despite the vast disparity in global motions, the local components exhibit similar motion patterns across different agents. Building on this insight, we propose a novel Deconstruct-Recompose Paradigm (DRP) for learning transferable local motion representations. Specifically, in the Deconstruct phase, we identify multiple local points and track their frame-wise motions, defining each as an Atomic Action. We introduce a Dual-Attention Encoder (DAE) to learn local motion representations from these Atomic Actions, capturing their spatiotemporal relationships. In the Recompose phase, we compose local motion representations with a learnable Motion Aggregation Token '[MAT]' via latent dynamics model learning. Additionally, an adapter bridges local motion and downstream action-specific dynamics to accelerate policy learning. Extensive experiments demonstrate that our method effectively transfers to diverse robotic control and manipulation tasks, significantly improving sample efficiency and performance.

The paradigm of agentic AI is shifting from engineered complex workflows to post-training native models. However, existing agents are typically confined to static, predefined action spaces--such as exclusively using APIs, GUI events, or robotic commands. This rigidity limits their adaptability in dynamic environments where the optimal granularity of interaction varies contextually. To bridge this gap, we propose CrossHA, a unified agentic model that masters heterogeneous action spaces and autonomously selects the most effective interface for each step of a trajectory. We introduce a comprehensive training pipeline that integrates cold-start supervised fine-tuning with a Multi-Turn Group Relative Policy Optimization (GRPO) algorithm. This approach enables the agent to learn adaptive action switching--balancing high-level efficiency with low-level precision--without human-specified rules. Extensive experiments on over 800 tasks in the open-world Minecraft environment demonstrate that CrossHA achieves state-of-the-art performance. By dynamically leveraging the strengths of diverse action spaces, our model significantly outperforms fixed-action baselines, exhibiting superior generalization and efficiency in long-horizon reasoning. All code and models are available at https://github.com/CraftJarvis/OpenHA.

High-fidelity generative models are increasingly needed in privacy-sensitive scenarios, where access to data is severely restricted due to regulatory and copyright constraints. This scarcity hampers model development--ironically, in settings where generative models are most needed to compensate for the lack of data. This creates a self-reinforcing challenge: limited data leads to poor generative models, which in turn fail to mitigate data scarcity. To break this cycle, we propose a reinforcement-guided synthetic data generation framework that adapts general-domain generative priors to privacy-sensitive identity recognition tasks. We first perform a cold-start adaptation to align a pretrained generator with the target domain, establishing semantic relevance and initial fidelity. Building on this foundation, we introduce a multi-objective reward that jointly optimizes semantic consistency, coverage diversity, and expression richness, guiding the generator to produce both realistic and task-effective samples. During downstream training, a dynamic sample selection mechanism further prioritizes high-utility synthetic samples, enabling adaptive data scaling and improved domain alignment. Extensive experiments on benchmark datasets demonstrate that our framework significantly improves both generation fidelity and classification accuracy, while also exhibiting strong generalization to novel categories in small-data regimes.

This paper proposes the Continual Dual-Critic with Cross-Attention (CD-CCA) framework for visual reinforcement learning to address the plasticity-stability conflict. Our method introduces continual learning techniques into the visual RL architecture, constructing two complementary critics using Continual Backpropagation (CBP) and Elastic Weight Consolidation (EWC) -- one for maintaining representational plasticity for rapid environmental adaptation, and the other for preserving knowledge stability to prevent catastrophic forgetting. Furthermore, we design a cross-attention based fusion mechanism that balances the value estimates from the dual critics according to observation characteristics. Experimental results on DeepMind Control and CARLA benchmarks show that CD-CCA effective mitigates issues of representation drift and policy degradation. Compared to existing visual RL methods, our approach exhibits enhanced robustness and adaptability in non-stationary environments and long-horizon decision-making tasks, providing a new architectural paradigm for the advancement of continual reinforcement learning.

Video understanding with multimodal large language models (MLLMs) remains challenging due to the long token sequences of videos, which contain extensive temporal dependencies and redundant frames.Existing approaches typically treat MLLMs as passive recognizers, processing entire videos or uniformly sampled frames without adaptive reasoning.Recent agent-based methods introduce external tools, yet still depend on manually designed workflows and perception-first strategies, resulting in inefficiency on long videos. We present EVA, an Efficient Reinforcement Learning framework for End-to-End Video Agent, which enables planning-before-perception through iterative summary-plan-action-reflection reasoning.EVA autonomously decides what to watch, when to watch, and how to watch, achieving query-driven and efficient video understanding.To train such agents, we design a simple yet effective three-stage learning pipeline--comprising supervised fine-tuning (SFT), Kahneman-Tversky Optimization (KTO), and Group Relative Policy Optimization (GRPO)--that bridges supervised imitation and reinforcement learning.We further construct high-quality datasets for each stage, supporting stable and reproducible training.We evaluate EVA on six video understanding benchmarks, demonstrating its comprehensive capabilities. Compared with existing baselines, EVA achieves a substantial improvement of 6--12% over general MLLM baselines and a further 1--3% gain over prior adaptive agent methods.

360deg panoramic images are increasingly used in VR, autonomous driving, and robotics for holistic scene understanding. However, current Vision-Language Models (VLMs) struggle with 3D spatial reasoning on Equirectangular Projection (ERP) images due to geometric distortion and limited 3D supervision. We introduce PanoEnv, a large-scale VQA benchmark built from synthetic 3D environments, containing 14.8K questions across five categories (e.g., relative position, attribute comparison) grounded in accurate 3D annotations--depth, segmentation, and bounding boxes. Benchmarking 14 state-of-the-art VLMs reveals limited 3D understanding, achieving only 49.34% overall and 8.36% on open-ended (OE) questions. To enhance 3D reasoning, we propose a reinforcement learning post-training framework based on Group Relative Policy Optimization (GRPO) with a ground-truth-guided reward combining five geometry-aware strategies (e.g., distance tolerance, spatial consistency). A two-stage curriculum further mitigates catastrophic forgetting: Stage 1 trains on structured tasks (T/F, MCQ), and Stage 2 fine-tunes on mixed OE data for generalization. Our 7B model sets a new SoTA performance, improving total accuracy to 52.93% (+3.59%) and OE accuracy to 14.83% while maintaining structured-task performance. It also achieves top semantic scores (Q-Score 6.24, P-Score 5.95), surpassing 32B models. These results demonstrate that PanoEnv-QA and our curriculum-based RL framework effectively instill 3D spatial intelligence in VLMs for omnidirectional perception.