EvoVid: Temporal-Centric Self-Evolution for Video Large Language Models

2026-05-22 · Source: cs.CV updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning · Depth: Expert, extended

Summary

EvoVid introduces a temporal-centric self-evolving framework for Video Large Language Models (Video-LLMs), enabling them to improve video understanding and reasoning directly from 5,778 raw, unannotated videos. This framework bypasses costly human annotations by employing a Questioner-Solver self-play mechanism. It features two novel temporal-centric rewards: a temporal-aware Questioner reward that encourages temporally dependent question generation by assessing sensitivity to frame perturbations, and a temporal-grounded Solver reward providing automatic temporal supervision through video segment localization. Experiments across four base models, including Qwen2.5-VL-3B/7B and Qwen3-VL-4B/8B, and six benchmarks demonstrate consistent performance improvements over both base models and existing self-evolving baselines, achieving competitive results with supervised methods.

Key takeaway

For AI Scientists and Machine Learning Engineers developing Video-LLMs, EvoVid presents a scalable paradigm to enhance temporal reasoning without extensive human annotation. You should consider integrating temporal-centric self-evolution, leveraging frame perturbation sensitivity for question generation and IoU-based segment localization for Solver supervision. This approach offers a path to more robust and autonomously improving video understanding models.

Key insights

EvoVid enables Video-LLMs to self-evolve temporal reasoning from raw video, eliminating human annotation dependency.

Principles

• Temporal dynamics are central to video reasoning.
• Self-play can generate verifiable learning signals.
• Perturbing temporal order reveals question sensitivity.

Method

EvoVid co-optimizes a Questioner and Solver using GRPO. The Questioner generates temporally dependent questions via frame shuffling, while the Solver predicts answers and localizes relevant K-frame video segments.

In practice

• Evaluate Solver confidence on original vs. shuffled frames.
• Reward Solver for accurate IoU-based segment localization.
• Iteratively train Questioner and Solver roles.

Topics

• Video-LLMs
• Self-evolving AI
• Temporal Reasoning
• Reinforcement Learning
• Video Understanding
• Label-free Learning

Best for: Research Scientist, AI Scientist, Machine Learning Engineer

Related on AIssential

• Temporal-Aware Reasoning Optimization for Video Temporal Grounding — TaRO improves MLLM video temporal grounding by optimizing time-aware reasoning through guided exploration and quality-focused rewards.
• Active Zero: Self-Evolving Vision-Language Models through Active Environment Exploration — Active environment exploration with co-evolving agents significantly improves VLM reasoning and general understanding.
• ReFoCUS: Reinforcement-guided Frame Optimization for Contextual Understanding — Optimizing visual input selection via reinforcement learning significantly enhances video-LLM contextual understanding and reasoning.
• CurEvo: Curriculum-Guided Self-Evolution for Video Understanding — CurEvo uses curriculum learning to guide self-evolution, improving autonomous video understanding through structured progression.
• AI News: This Video Model Has Everyone Freaked Out! — New AI models are pushing boundaries in multimodal video generation, accelerated code synthesis, and cost-effective, agentic LLMs.
• EvoTrainer: Co-Evolving LLM Policies and Training Harnesses for Autonomous Agentic Reinforcement Learning — EvoTrainer co-evolves LLM policies and training harnesses, moving beyond static recipe search for agentic RL.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.CV updates on arXiv.org.