Skip to content

VideoZoomer: Reinforcement-Learned Temporal Focusing for Long Video Reasoning

🕒 Published (v1): 2026-01-01 · Source: ICLR · Venue: ICLR 2026 · link

Why this paper was selected

RL temporal focusing on relevant frames enables efficient long-video VLM reasoning

Ask a follow-up

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

💬 Ask ChatGPT✦ Ask Claude

TL;DR

VideoZoomer is an agentic framework that enables MLLMs to dynamically allocate their visual attention during long video reasoning by iteratively invoking a temporal zoom tool to retrieve high-frame-rate clips at model-chosen timestamps. It is trained via a two-stage strategy: cold-start SFT on distilled exemplar and reflection trajectories, followed by GRPO-based multi-turn RL. A 7B model surpasses all open-source baselines and even proprietary models on long video reasoning benchmarks.

Problem

Existing long video understanding approaches rely on uniform frame sampling or static pre-selection modules that (1) treat all moments as equally important, missing short critical events, and (2) are decoupled from the reasoning process, with no mechanism to correct suboptimal initial selections. Context window constraints further prevent processing videos at sufficient temporal resolution throughout.

Method

VideoZoomer frames long video understanding as a sequential tool interaction task. The model receives a coarse low-frame-rate overview \(V_\text{low}\) (64 uniformly sampled frames) and can invoke a <video_zoom> tool with arguments \([t_\text{start}, t_\text{end}]\) to retrieve a high-frame-rate clip \(V_\text{clip} = T(V, t_\text{start}, t_\text{end}, f_\text{high})\). Each tool call is frame-budget constrained: \(f_\text{high} \times (t_\text{end} - t_\text{start}) \leq B\), with up to 4 turns of 16 frames each, yielding a maximum of 128 total frames.

Training (two stages): 1. Cold-start SFT: ~11K multi-turn trajectories curated from GPT-4o/Gemini-2.5-Pro expert demonstrations (exemplar trajectories) plus reflection data — failed rollouts corrected by the expert model to teach error recovery and diverse reasoning patterns. 2. Multi-turn RL with GRPO: A token-level loss mask ignores tool-response tokens. The reward is \(R = R_\text{acc} + R_\text{format} + R_\text{tool}\), where \(R_\text{tool}\) is a conditional exploration bonus (awarded only when the tool is used and the final answer is correct) to prevent redundant calls.

The base model is Qwen2.5-VL-7B-Instruct, trained on LongVideoReason (52K QA pairs).

Key Contributions

  • An agentic "first-glance, then zoom" framework enabling iterative, on-demand high-frame-rate evidence gathering during reasoning.
  • A two-stage training pipeline combining cold-start SFT with curated exemplar+reflection trajectories and multi-turn GRPO-based RL.
  • Extension of GRPO to multi-turn tool-calling with a token-level loss mask that correctly attributes credit to model-generated tokens only.
  • A conditional tool-use exploration reward \(R_\text{tool}\) that incentivizes exploration without rewarding spurious tool calls.
  • Demonstration of emergent diverse reasoning patterns: direct-hit, progressive, and self-refine strategies.

Results

  • MLVU: 68.8 (dev) / 55.8 (test) vs. Qwen2.5-VL base 58.3/45.5 (+10.5/+10.3 pts); best among open-source 7B models.
  • LongVideoBench: 57.7 overall, outperforming all listed open-source baselines.
  • VideoMME: 65.2 overall / 55.8 long-set vs. base 63.5/53.9.
  • LVBench: 41.5 vs. base 36.9.
  • VideoMMLU / VideoMMMU: 67.9 / 52.2 — highest among open-source models.
  • LongVideoReason-eval: 80.3, surpassing GPT-4o (60.7) and Gemini-1.5-Pro (67.3).
  • Action Count (MLVU multi-detail): 13.6 → 50.5 (+36.9 pts), the largest single-task gain.
  • Ablations: removing RL causes −17.0 on LongVideoReason; removing cold-start causes −20.7; removing reflection data causes −5.2; removing \(R_\text{tool}\) causes degraded tool-call frequency during training.
  • Achieves better accuracy-vs-frames efficiency than Qwen2.5-VL across frame budgets 32–256.

Limitations

  • Maximum of 4 tool calls per sample; performance gains plateau if the required evidence exceeds this interaction budget.
  • The cold-start dataset relies on expensive proprietary models (GPT-4o, Gemini-2.5-Pro) as expert demonstrators.
  • Evaluation is confined to multiple-choice QA formats; open-ended generation scenarios are not assessed.
  • The frame-budget constraint \(B \times N\) may still be insufficient for extremely dense temporal events in very long videos.
  • Training dataset (LongVideoReason) is fixed; domain generalization to specialized video types (medical, industrial) is not examined.

Relevance to Vision-Language Models

VideoZoomer directly advances the VLM capability frontier on long video, a domain where context-window constraints have been a persistent bottleneck. It demonstrates that RL with tool use can teach a 7B VLM to develop human-like attentional strategies — selectively escalating temporal resolution only where needed — rather than relying on passive frame compression. The two-stage SFT→RL recipe with reflection data is a transferable training paradigm applicable to other VLM agentic tasks (e.g., document retrieval, spatial zoom for high-resolution images). The result that a 7B open-source model with adaptive sampling outperforms proprietary systems on reasoning-heavy benchmarks is a key data point for the ongoing scaling-vs-efficiency debate in the VLM community.