Research map / Video / Video-language understanding

ParaVT: Taming the Tool Prior Paradox for Parallel Tool Use in Agentic Video Reinforcement Learning

NeurIPS 2026 · Video › Video-language understanding · approach: Long video understanding · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes ParaVT, an end-to-end RL-trained LMM that dispatches multiple time-window crop tool calls in one turn for long-video QA, with PARA-GRPO adding format anchoring and frame-budget gating to fix format collapse and tool-skipping under GRPO.

Problem

Native-RL tool-augmented LMMs for long-video understanding dispatch tool calls sequentially, which propagates errors and inflates context; standard GRPO on such models suffers format collapse under temperature sampling and a reward shortcut of skipping tools.

Approach

Single-turn parallel crop_video dispatch to sub-agents returning text summaries, plus PARA-GRPO: a selective format/anchor reward at closing-tag positions and per-prompt random overview-frame budgets that create reward contrast between calling and skipping the tool.

Key results

• ParaVT improves over Qwen3-VL-8B base by +7.9% average relative gain across seven splits, reaching 60.4 on LongVideoBench and 39.8 on LVBench.
• PARA-GRPO lifts training-time format reward at tau=0.7 from 0.13 (vanilla GRPO) to 0.41 mean (peak 0.64).
• ParaVT reaches 50.1 mIoU on Charades-STA versus 49.3 for the Qwen3-VL-8B base.

Surprising findings

• Vanilla GRPO drives tool-call rate to near zero within about 7 steps, with the policy learning to skip tools.
• Qwen2.5-VL, with weaker tool priors, keeps format near-perfect under RL but emits zero tool calls, supporting the Tool Prior Paradox.

Limitations

• Full pipeline validated only on Qwen3-VL-8B; the prior-strength claim rests on a single Qwen2.5-VL vs Qwen3-VL contrast.
• Only the crop_video tool is evaluated.
• RL mainly improves deployment-time format and tool-use stability rather than large greedy-eval gains over the cold-start checkpoint.

Experimental setup

Datasets: LLaVA-Video-178K LongVideo-Reason Charades-STA LongVT rollouts VideoR1 RL pool MUSEG LongVT-selfQA-v2

Benchmarks: VideoMME LongVideoBench LVBench MLVU MMVU Charades-STA

Metrics: accuracy mIoU format reward tool-call rate

Baselines: Qwen3-VL-8B Conan-7B LongVT-RFT-7B SAGE-7B VideoZoomer-7B Video-R1-7B VideoChat-R1-7B VideoRFT-7B Time-R1-7B ReWatch-R1-7B Video-Thinker-7B Qwen2.5-VL-7B

Builds on: GRPO DAPO LongVT Qwen3-VL AReaL

Closest papers by meaning

• ReAgent-V: A Reward-Driven Multi-Agent Framework for Video Understanding (NeurIPS 2025 · similarity 0.88)
• PyVision-RL: Forging Open Agentic Vision Models via RL (ICML 2026 · similarity 0.88)
• Find, Fix, Reason: Context Repair for Video Reasoning (ICML 2026 · similarity 0.88)
• DeepEyesV2: Toward Agentic Multimodal Model (ICLR 2026 · similarity 0.87)
• TempSamp-R1: Effective Temporal Sampling with Reinforcement Fine-Tuning for Video LLMs (NeurIPS 2025 · similarity 0.87)
• SenseSearch: Empowering Vision-Language Models with High-Resolution Agentic Search-Reasoning via Reinforcement Learning (CVPR 2026 · similarity 0.86)