Video-OPD: Efficient Post-Training of Multimodal Large Language Models for Temporal Video Grounding via On-Policy Distillation

Reinforcement learning has emerged as a principled post-training paradigm for Temporal Video Grounding (TVG) due to its on-policy optimization, yet existing GRPO-based methods remain fundamentally constrained by sparse reward signals and substantial computational overhead. We propose Video-OPD, an efficient post-training framework for TVG inspired by recent advances in on-policy distillation. Video-OPD optimizes trajectories sampled directly from the current policy, thereby preserving alignment between training and inference distributions, while a frontier teacher supplies dense, token-level supervision via a reverse KL divergence objective. This formulation preserves the on-policy property critical for mitigating distributional shift, while converting sparse, episode-level feedback into fine-grained, step-wise learning signals. Building on Video-OPD, we introduce Teacher-Validated Disagreement Focusing (TVDF), a lightweight training curriculum that iteratively prioritizes trajectories that are both teacher-reliable and maximally informative for the student, thereby improving training efficiency. Empirical results demonstrate that Video-OPD consistently outperforms GRPO while achieving substantially faster convergence and lower computational cost, establishing on-policy distillation as an effective alternative to conventional reinforcement learning for TVG.

Restoring Exploration after Post-Training: Latent Exploration Decoding for Large Reasoning Models

Large Reasoning Models (LRMs) have recently achieved strong mathematical and code reasoning performance through Reinforcement Learning (RL) post-training. However, we show that modern reasoning post-training induces an unintended exploration collapse: temperature-based sampling no longer increases pass@$n$ accuracy. Empirically, the final-layer posterior of post-trained LRMs exhibit sharply reduced entropy, while the entropy of intermediate layers remains relatively high. Motivated by this entropy asymmetry, we propose Latent Exploration Decoding (LED), a depth-conditioned decoding strategy. LED aggregates intermediate posteriors via cumulative sum and selects depth configurations with maximal entropy as exploration candidates. Without additional training or parameters, LED consistently improves pass@1 and pass@16 accuracy by 0.61 and 1.03 percentage points across multiple reasoning benchmarks and models. Project page: https://GitHub.com/Xiaomi-Research/LED.

FutureMind: Equipping Small Language Models with Strategic Thinking-Pattern Priors via Adaptive Knowledge Distillation

Small Language Models (SLMs) are attractive for cost-sensitive and resource-limited settings due to their efficient, low-latency inference. However, they often struggle with complex, knowledge-intensive tasks that require structured reasoning and effective retrieval. To address these limitations, we propose FutureMind, a modular reasoning framework that equips SLMs with strategic thinking-pattern priors via adaptive knowledge distillation from large language models (LLMs). FutureMind introduces a dynamic reasoning pipeline composed of four key modules: Problem Analysis, Logical Reasoning, Strategy Planning, and Retrieval Guidance. This pipeline is augmented by three distinct retrieval paradigms that decompose complex queries into tractable subproblems, ensuring efficient and accurate retrieval execution. Extensive experiments on multi-hop QA benchmarks, including 2WikiMultihopQA, MuSiQue, Bamboogle, and Frames, demonstrate the superiority of FutureMind. It consistently outperforms strong baselines such as Search-o1, achieving state-of-the-art results under free training conditions across diverse SLM architectures and scales. Beyond empirical gains, our analysis reveals that the process of thinking-pattern distillation is restricted by the cognitive bias bottleneck between the teacher (LLMs) and student (SLMs) models. This provides new perspectives on the transferability of reasoning skills, paving the way for the development of SLMs that combine efficiency with genuine cognitive capability.

MobileBench-OL: A Comprehensive Chinese Benchmark for Evaluating Mobile GUI Agents in Real-World Environment

Recent advances in mobile Graphical User Interface (GUI) agents highlight the growing need for comprehensive evaluation benchmarks. While new online benchmarks offer more realistic testing than offline ones, they tend to focus on the agents' task instruction-following ability while neglecting their reasoning and exploration ability. Moreover, these benchmarks do not consider the random noise in real-world mobile environments. This leads to a gap between benchmarks and real-world environments. To addressing these limitations, we propose MobileBench-OL, an online benchmark with 1080 tasks from 80 Chinese apps. It measures task execution, complex reasoning, and noise robustness of agents by including 5 subsets, which set multiple evaluation dimensions. We also provide an auto-eval framework with a reset mechanism, enabling stable and repeatable real-world benchmarking. Evaluating 12 leading GUI agents on MobileBench-OL shows significant room for improvement to meet real-world requirements. Human evaluation further confirms that MobileBench-OL can reliably measure the performance of leading GUI agents in real environments. Our data and code will be released upon acceptance.

R^3: Replay, Reflection, and Ranking Rewards for LLM Reinforcement Learning

Large reasoning models (LRMs) aim to solve diverse and complex problems through structured reasoning. Recent advances in group-based policy optimization methods have shown promise in enabling stable advantage estimation without reliance on process-level annotations. However, these methods rely on advantage gaps induced by high-quality samples within the same batch, which makes the training process fragile and inefficient when intra-group advantages collapse under challenging tasks. To address these problems, we propose a reinforcement learning mechanism named \emph{\textbf{R^3}} that along three directions: (1) a \emph{cross-context \underline{\textbf{R}}eplay} strategy that maintains the intra-group advantage by recalling valuable examples from historical trajectories of the same query, (2) an \emph{in-context self-\underline{\textbf{R}}eflection} mechanism enabling models to refine outputs by leveraging past failures, and (3) a \emph{structural entropy \underline{\textbf{R}}anking reward}, which assigns relative rewards to truncated or failed samples by ranking responses based on token-level entropy patterns, capturing both local exploration and global stability. We implement our method on Deepseek-R1-Distill-Qwen-1.5B and train it on the DeepscaleR-40k in the math domain. Experiments demonstrate our method achieves SoTA performance on several math benchmarks, representing significant improvements and fewer reasoning tokens over the base models. Code and model will be released.

GAIA: A Data Flywheel System for Training GUI Test-Time Scaling Critic Models

While Large Vision-Language Models (LVLMs) have significantly advanced GUI agents' capabilities in parsing textual instructions, interpreting screen content, and executing tasks, a critical challenge persists: the irreversibility of agent operations, where a single erroneous action can trigger catastrophic deviations. To address this, we propose the GUI Action Critic's Data Flywheel System (GAIA), a training framework that enables the models to have iterative critic capabilities, which are used to improve the Test-Time Scaling (TTS) of basic GUI agents' performance. Specifically, we train an Intuitive Critic Model (ICM) using positive and negative action examples from a base agent first. This critic evaluates the immediate correctness of the agent's intended actions, thereby selecting operations with higher success probability. Then, the initial critic guides agent actions to collect refined positive/negative samples, initiating the self-improving cycle. The augmented data then trains a second-round critic with enhanced discernment capability. We conduct experiments on various datasets and demonstrate that the proposed ICM can improve the test-time performance of various closed-source and open-source models, and the performance can be gradually improved as the data is recycled. The code and dataset will be publicly released.

Federated Balanced Learning

Federated learning is a paradigm of joint learning in which clients collaborate by sharing model parameters instead of data. However, in the non-iid setting, the global model experiences client drift, which can seriously affect the final performance of the model. Previous methods tend to correct the global model that has already deviated based on the loss function or gradient, overlooking the impact of the client samples. In this paper, we rethink the role of the client side and propose Federated Balanced Learning, i.e., FBL, to prevent this issue from the beginning through sample balance on the client side. Technically, FBL allows unbalanced data on the client side to achieve sample balance through knowledge filling and knowledge sampling using edge-side generation models, under the limitation of a fixed number of data samples on clients. Furthermore, we design a Knowledge Alignment Strategy to bridge the gap between synthetic and real data, and a Knowledge Drop Strategy to regularize our method. Meanwhile, we scale our method to real and complex scenarios, allowing different clients to adopt various methods, and extend our framework to further improve performance. Numerous experiments show that our method outperforms state-of-the-art baselines. The code is released upon acceptance.

Xiaomi MiMo-VL-Miloco Technical Report

We open-source MiMo-VL-Miloco-7B and its quantized variant MiMo-VL-Miloco-7B-GGUF, a pair of home-centric vision-language models that achieve strong performance on both home-scenario understanding and general multimodal reasoning. Built on the MiMo-VL-7B backbone, MiMo-VL-Miloco-7B is specialized for smart-home environments, attaining leading F1 scores on gesture recognition and common home-scenario understanding, while also delivering consistent gains across video benchmarks such as Video-MME, Video-MMMU, and Charades-STA, as well as language understanding benchmarks including MMMU-Pro and MMLU-Pro. In our experiments, MiMo-VL-Miloco-7B outperforms strong closed-source and open-source baselines on home-scenario understanding and several multimodal reasoning benchmarks. To balance specialization and generality, we design a two-stage training pipeline that combines supervised fine-tuning with reinforcement learning based on Group Relative Policy Optimization, leveraging efficient multi-domain data. We further incorporate chain-of-thought supervision and token-budget-aware reasoning, enabling the model to learn knowledge in a data-efficient manner while also performing reasoning efficiently. Our analysis shows that targeted home-scenario training not only enhances activity and gesture understanding, but also improves text-only reasoning with only modest trade-offs on document-centric tasks. Model checkpoints, quantized GGUF weights, and our home-scenario evaluation toolkit are publicly available at https://github.com/XiaoMi/xiaomi-mimo-vl-miloco to support research and deployment in real-world smart-home applications.

STEP: Success-Rate-Aware Trajectory-Efficient Policy Optimization

Multi-turn interaction remains challenging for online reinforcement learning. A common solution is trajectory-level optimization, which treats each trajectory as a single training sample. However, this approach can be inefficient and yield misleading learning signals: it applies uniform sampling across tasks regardless of difficulty, penalizes correct intermediate actions in failed trajectories, and incurs high sample-collection costs. To address these issues, we propose STEP (Success-rate-aware Trajectory-Efficient Policy optimization), a framework that dynamically allocates sampling based on per-task success rates and performs step-level optimization. STEP maintains a smoothed success-rate record to guide adaptive trajectory resampling, allocating more effort to harder tasks. It then computes success-rate-weighted advantages and decomposes trajectories into step-level samples. Finally, it applies a step-level GRPO augmentation to refine updates for low-success tasks. Experiments on OSWorld and AndroidWorld show that STEP substantially improves sample efficiency and training stability over trajectory-level GRPO, converging faster and generalizing better under the same sampling budget.

REVISOR: Beyond Textual Reflection, Towards Multimodal Introspective Reasoning in Long-Form Video Understanding

Self-reflection mechanisms that rely on purely text-based rethinking processes perform well in most multimodal tasks. However, when directly applied to long-form video understanding scenarios, they exhibit clear limitations. The fundamental reasons for this lie in two points: (1)long-form video understanding involves richer and more dynamic visual input, meaning rethinking only the text information is insufficient and necessitates a further rethinking process specifically targeting visual information; (2) purely text-based reflection mechanisms lack cross-modal interaction capabilities, preventing them from fully integrating visual information during reflection. Motivated by these insights, we propose REVISOR (REflective VIsual Segment Oriented Reasoning), a novel framework for tool-augmented multimodal reflection. REVISOR enables MLLMs to collaboratively construct introspective reflection processes across textual and visual modalities, significantly enhancing their reasoning capability for long-form video understanding. To ensure that REVISOR can learn to accurately review video segments highly relevant to the question during reinforcement learning, we designed the Dual Attribution Decoupled Reward (DADR) mechanism. Integrated into the GRPO training strategy, this mechanism enforces causal alignment between the model's reasoning and the selected video evidence. Notably, the REVISOR framework significantly enhances long-form video understanding capability of MLLMs without requiring supplementary supervised fine-tuning or external models, achieving impressive results on four benchmarks including VideoMME, LongVideoBench, MLVU, and LVBench.