Skip to content

VideoZoomer: Reinforcement-Learned Temporal Focusing for Long Video Reasoning

🕒 Published (v1): 2025-12-26 11:43 UTC · Source: Arxiv · link

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

💬 Ask ChatGPT✦ Ask Claude

TL;DR

VideoZoomer trains a 7B MLLM to act as an active agent for long video understanding by iteratively invoking a <video_zoom> tool to retrieve high-frame-rate clips of autonomously selected temporal segments. A two-stage pipeline—cold-start SFT on distilled exemplar and reflection trajectories, followed by GRPO-based RL with a conditional tool-use bonus reward—teaches the model when, where, and how many times to zoom. The result is a dynamic, multi-turn evidence-gathering system that outperforms much larger frame-budget baselines.

Problem

MLLMs are bottlenecked on long video understanding by fixed context windows. Uniform frame sampling treats all moments as equally important; static pre-selection (frame selectors) is non-interactive and cannot correct initial errors. Neither approach adapts to per-question complexity or recovers from mis-selected frames.

Method

Framework. The agent begins with a coarse low-FPS overview \(V_\text{low}\) (64 uniformly sampled frames). It can invoke <video_zoom>(t_start, t_end, f_high) up to 4 times per query, each retrieving a high-FPS clip of ≤16 frames from the original video. A per-call frame budget \(B\) enforces \(f_\text{high} \times (t_\text{end} - t_\text{start}) \le B\); illegal calls return error messages.

Cold-start SFT. A dataset of ~11,000 multi-turn trajectories is constructed in two steps: (1) GPT-4o/Gemini-2.5-Pro generate exemplar "golden" trajectories; (2) reflection data is created by having the initial SFT model roll out on hard examples, then having the expert model correct failures into richer, multi-call reasoning paths. Both streams are filtered by a verifier before inclusion.

RL phase. GRPO is extended to multi-turn tool-call contexts with a token-level loss mask (only model-generated tokens receive gradient). The reward is \(R = R_\text{acc} + R_\text{format} + R_\text{tool}\), where \(R_\text{tool}\) is a conditional bonus: awarded only when the model calls the tool and the final answer is correct, preventing degenerate tool-spam.

Key Contributions

  • <video_zoom> agentic tool that enables iterative, budget-constrained temporal zoom within a single multi-turn reasoning trajectory
  • Cold-start dataset combining distilled exemplar trajectories and expert-corrected reflection trajectories to seed diverse reasoning patterns
  • Multi-turn GRPO extension with conditional tool-use bonus to prevent policy collapse and encourage non-trivial exploration
  • Demonstrated emergent reasoning patterns: direct-hit, progressive, and self-refine reasoning
  • Compatibility with external frame selectors (TSPO) as drop-in replacements for the initial overview

Results

  • MLVU: 68.8 dev / 55.8 test vs. Qwen2.5-VL-7B base 58.3/45.5 (+10.5/+10.3); beats all listed open-source 7B models
  • LongVideoBench: 57.7 val vs. base 51.0 (+6.7)
  • LongVideoReason-eval: 80.3 vs. GPT-4o 60.7, Gemini-1.5-Pro 67.3—surpasses both proprietary models
  • VideoMMLU / VideoMMMU: 67.9 / 52.2, highest among open-source
  • Efficiency: on MLVU, achieves 0.64 accuracy at 48 avg frames vs. base 0.581 at 128 frames; on LVBench, 77 frames beats base at 256
  • Ablations: removing RL drops LongVideoReason by −17.0; removing cold-start causes non-convergence; removing reflection data collapses avg tool calls to ~1.0; removing \(R_\text{tool}\) drives tool usage to zero

Limitations

  • Maximum of 4 tool calls and 16 frames per call are fixed hyperparameters; optimal values per task are not explored
  • Initial overview is still uniform sampling by default, which may miss rare events entirely before any zoom opportunity
  • Cold-start relies on proprietary teacher models (GPT-4o, Gemini-2.5-Pro) for trajectory generation
  • Evaluation is multiple-choice QA; open-ended generation quality is not assessed
  • Frame budget \(B \times N\) grows linearly with turns; very long videos requiring many zooms may still overflow context

Relevance to Harnesses / Meta-Harnesses

VideoZoomer is a concrete instance of a multi-turn agentic harness built on top of a base VLM: the framework defines the tool contract, observation loop, trajectory format, and reward structure that orchestrate the model's behavior—classic harness concerns. Its two-stage training pipeline (SFT cold-start with curated exemplar+reflection data, then RL policy refinement) is itself a meta-harness over the base model training loop, analogous to how meta-harnesses in language settings coordinate multiple reasoning or retrieval tools. The design of the conditional \(R_\text{tool}\) reward and the reflection data generation loop—where a failing agent's trajectories are corrected and fed back—demonstrates a self-improving harness pattern directly applicable to any tool-augmented agentic system. The modular swap of the initial overview (uniform sampling → TSPO frame selector with +2–3 point gains) further illustrates how harness components can be hot-swapped without retraining the core agent.