Where to Look Matters: On-Policy Self-Distillation for Long-Video Understanding

• URL: http://arxiv.org/abs/2608.25356v1
• Date: Wed, 26 Aug 2026 04:23:21 GMT
• Title: Where to Look Matters: On-Policy Self-Distillation for Long-Video Understanding
• Abstract summary: As videos become longer, the full-video context inevitably contains more question-irrelevant temporal content.<n>We empirically find that focusing the visual input on short annotated clue intervals containing question-relevant evidence consistently improves prediction accuracy.<n>We introduce Clue-OPSD, a clue-privileged on-policy self-distillation framework for long-video understanding.
• Score: 43.28973351947568
• License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
• Abstract: Vision-language models (VLMs) have made substantial progress in long-video understanding, with standard backbone models typically answering questions from frames sampled across the full video. However, as videos become longer, the full-video context inevitably contains more question-irrelevant temporal content, which can distract the model from the evidence needed to answer a specific question. We empirically find that focusing the visual input on short annotated clue intervals containing question-relevant evidence consistently improves prediction accuracy across model scales compared with using the corresponding full videos, while requiring fewer input frames. Based on this finding, we introduce Clue-OPSD, a clue-privileged on-policy self-distillation framework for long-video understanding. During training, a full-video student learns from a self-teacher conditioned on the corresponding clue interval by aligning their next-token distributions along student-generated trajectories. Clue-OPSD thus uses clue intervals as privileged supervision without relying on ground-truth answer labels, while requiring no clue annotations or additional modules at inference time. Extensive experiments across multiple long-video understanding benchmarks and Qwen3.5 model scales demonstrate consistent improvements over the corresponding backbone models and strong performance against supervised post-training baselines.

Related papers

• VideoScout: Learning Agentic Active Exploration with Adaptive Reasoning Pacing for Long Video Understanding [61.00809248898572]
  Long video understanding is a problem in which an agent reads the video turn by turn along the temporal axis, deciding at each turn how fast to watch, what evidence to retain, when to revisit uncertain segments, and when to stop and answer.<n>We propose textbfVideoScout, a multi-turn reasoning agent that instantiates the SEA paradigm through adaptive reasoning pacing.<n>Experiments on long video understanding and reasoning benchmarks demonstrate that our 7B model achieves strong performance compared with existing trained 7B agentic models.
  arXiv  Detail & Related papers  (2026-09-14T14:04:03Z)
• TTA-Vid: Generalized Test-Time Adaptation for Video Reasoning [54.70019148172847]
  We leverage the paradigm of Test-Time Reinforcement Learning on video-language data to allow for adapting a pretrained model to incoming video samples at test-time without explicit labels.<n>The proposed test-time adaptation for video approach (TTA-Vid) combines two components that work simultaneously.<n>TTA-Vid yields consistent improvements across various video reasoning tasks and is able to outperform current state-of-the-art methods trained on large-scale data.
  arXiv  Detail & Related papers  (2026-04-01T09:52:57Z)
• Think with Grounding: Curriculum Reinforced Reasoning with Video Grounding for Long Video Understanding [38.87967229483403]
  Video-TwG is a curriculum reinforced framework that employs a novel Think-with-Grounding paradigm.<n>Video-TwG can be trained end-to-end in a straightforward manner, without relying on complex auxiliary modules or heavily annotated reasoning traces.<n>Our algorithm features the fine-grained grounding reward, self-confirmed pseudo reward and accuracy-gated mechanism.
  arXiv  Detail & Related papers  (2026-02-21T03:16:23Z)
• LongVT: Incentivizing "Thinking with Long Videos" via Native Tool Calling [87.98096428508181]
  LongVT is an end-to-end agentic framework that enables "Thinking with Long Videos" via interleaved Multimodal Chain-of-Tool-Thought.<n>We exploit LMMs' inherent temporal grounding ability as a native video cropping tool to zoom in on a specific video clip and resample finer-grained video frames.<n>Our training dataset consists of 247.9K samples for tool-integrated cold-start supervised fine-tuning, 1.6K samples for agentic reinforcement learning, and 15.4K samples for agentic reinforcement fine-tuning.
  arXiv  Detail & Related papers  (2025-11-25T19:22:48Z)
• SeViCES: Unifying Semantic-Visual Evidence Consensus for Long Video Understanding [36.30263540665245]
  We propose a framework for effective and reliable long video understanding.<n>SeViCES is training-free and model-agnostic, and introduces two key components.<n> experiments on long video understanding benchmarks show that SeViCES consistently outperforms state-of-the-art methods in both accuracy and robustness.
  arXiv  Detail & Related papers  (2025-10-23T14:55:28Z)
• On the Consistency of Video Large Language Models in Temporal Comprehension [57.985769348320616]
  Video large language models (Video-LLMs) can temporally ground language queries and retrieve video moments.<n>We conduct a study on prediction consistency -- a key indicator for robustness and trustworthiness of temporal grounding.
  arXiv  Detail & Related papers  (2024-11-20T00:47:17Z)
• TemporalBench: Benchmarking Fine-grained Temporal Understanding for Multimodal Video Models [75.42002690128486]
  TemporalBench is a new benchmark dedicated to evaluating fine-grained temporal understanding in videos. It consists of 10K video question-answer pairs, derived from 2K high-quality human annotations detailing the temporal dynamics in video clips. Results show that state-of-the-art models like GPT-4o achieve only 38.5% question answering accuracy on TemporalBench.
  arXiv  Detail & Related papers  (2024-10-14T17:59:58Z)
• Weakly Supervised Gaussian Contrastive Grounding with Large Multimodal Models for Video Question Answering [11.244643114253773]
  Video Question (VideoQA) aims to answer natural language questions based on the information observed in videos. We propose a novel weakly supervised framework to enforce the LMMs to reason out the answers with question-critical moments as visual inputs.
  arXiv  Detail & Related papers  (2024-01-19T14:21:46Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.