# Don't Read the Log: Execution Traces Contaminate Verifiers in Video-Generation Agents

arXiv: https://arxiv.org/abs/2609.28564 · HTML: https://mlsanity.com/paper/2609.28564/

## Key facts

- No citing papers in the map yet (data through 7 Oct 2026).
- Placed in Software engineering coding agents, one of 16 directions in LLM agents and tool use.
- Builds on 23 of its 28 references that are in the map.
- First posted on arXiv 23 Sep 2026.

## Placed in

- LLM agents and tool use / [Software engineering coding agents](https://mlsanity.com/agents/software-engineering-coding-agents/)

## Builds on (most cited first, 23 of 23)

- [Qwen2.5-VL Technical Report](https://mlsanity.com/paper/2502.13923/) (2025)
- [Wan: Open and Advanced Large-Scale Video Generative Models](https://mlsanity.com/paper/2503.20314/) (2025)
- [Qwen3-VL Technical Report](https://mlsanity.com/paper/2511.21631/) (2025)
- [DeepSeek-V4: Towards Highly Efficient Million-Token Context Intelligence](https://mlsanity.com/paper/2606.19348/) (2026)
- [GenArtist: Multimodal LLM as an Agent for Unified Image Generation and Editing](https://mlsanity.com/paper/2407.05600/) (2024)
- [The Curse of Multi-Modalities: Evaluating Hallucinations of Large Multimodal Models across Language, Visual, and Audio](https://mlsanity.com/paper/2410.12787/) (2024)
- [T2I-Copilot: A Training-Free Multi-Agent Text-to-Image System for Enhanced Prompt Interpretation and Interactive Generation](https://mlsanity.com/paper/2507.20536/) (2025)
- [VISTA: A Test-Time Self-Improving Video Generation Agent](https://mlsanity.com/paper/2510.15831/) (2025)
- [GenMAC: Compositional Text-to-Video Generation with Multi-Agent Collaboration](https://mlsanity.com/paper/2412.04440/) (2024)
- [Let's Think in Two Steps: Mitigating Agreement Bias in MLLMs with Self-Grounded Verification](https://mlsanity.com/paper/2507.11662/) (2025)
- [ViMax: Agentic Video Generation](https://mlsanity.com/paper/2606.07649/) (2026)
- [SPIRAL: Self-Evolving Action-Conditioned Video Generation via Reflective Planning Agents](https://mlsanity.com/paper/2603.08403/) (2026)
- [VideoWeaver: Evaluating and Evolving Skills for Agentic Long Video Generation](https://mlsanity.com/paper/2606.08091/) (2026)
- [VideoAgent: All-in-One Framework for Video Understanding and Editing](https://mlsanity.com/paper/2606.23327/) (2026)
- [SoliReward: Mitigating Susceptibility to Reward Hacking and Annotation Noise in Video Generation Reward Models](https://mlsanity.com/paper/2512.22170/) (2025)
- [When Vision-Language Models Judge Without Seeing: Exposing Informativeness Bias](https://mlsanity.com/paper/2604.17768/) (2026)
- [On the Adversarial Robustness of Multimodal LLM Judges](https://mlsanity.com/paper/2606.15608/) (2026)
- [Evidence-RL: Towards Evidence-intensive Visual Reasoning](https://mlsanity.com/paper/2608.08021/) (2026)
- [OmniHalluc-L: Counterfactual Benchmarking and Modality-Perturbation Reliability Calibration for Long-Form Omni Hallucination](https://mlsanity.com/paper/2606.03614/) (2026)
- [When Does a Video-Language Model Stop Watching? Reward Strength Controls the Formation and Reversal of Visual Shortcuts in Multimodal RLVR](https://mlsanity.com/paper/2606.22043/) (2026)
- [Better Call CineCrew: Consistent Ultra-Long Narrative-to-Film Generation](https://mlsanity.com/paper/2609.07720/) (2026)
- [Seeing Isn't Believing: Uncovering Blind Spots in Evaluator Vision-Language Models](https://mlsanity.com/paper/2604.21523/) (2026)
- [EG-VQA: Benchmarking Verifiable Video Question Answering with Grounded Temporal Evidence](https://mlsanity.com/paper/2606.24797/) (2026)

## Abstract (arXiv)

Agentic video-generation systems close a loop between a generator and a verifier: an LLM plans shots, calls a text-to-video model, and a multimodal judge decides whether the result satisfies the request. To diagnose where a long workflow fails, recent harnesses deliberately show the judge more than the video-the agent's execution trace, its plan, the narration it synthesized. We ask whether this auxiliary text moves the judge's verdict on purely visual requirements, holding the frames fixed. On a benchmark of 109 generated two-event clips with manual labels, in which the requested event is either visibly completed or visibly missing, a trace that reports a successful tool call makes three open-weight Qwen-VL judges (7B, 8B, 32B) accept 78--90% of the failures, up from 7--19% without text, and a contradicting trace makes them reject up to 100% of correct clips; an instruction to ``use only the frames'' does not remove the effect. Frontier closed judges are essentially unmoved on the same clips, showing that the vulnerability is a property of the judge's learned trust in tool logs rather than of the task. Plan-derived text carries no clip-specific information, so it can only shift a judge's operating point, and in a repair loop that shift becomes a cap on the true pass rate that no repair policy can exceed; the cap matches simulation to two decimals. In the loop, contamination is exploited without any adversarial agent: an honest LLM planner that always regenerates ends with a judge pass rate of 1.00 and a human-labelled pass rate of 0.28, and a pipeline in which a cheap checker writes its verdict into the trace launders that checker's errors into a stronger final judge (0.69 false accepts).