# When Does a Video-Language Model Stop Watching? Reward Strength Controls the Formation and Reversal of Visual Shortcuts in Multimodal RLVR

arXiv: https://arxiv.org/abs/2606.22043 · HTML: https://mlsanity.com/paper/2606.22043/

## Key facts

- Cited by 2 papers in the map (data through 14 Oct 2026).
- Placed in Embodied and 3D vision-language models, one of 16 directions in Multimodal LLMs / vision-language models.
- Builds on 13 of its 14 references that are in the map.
- First posted on arXiv 20 Jun 2026.

## Placed in

- Multimodal LLMs / vision-language models / [Embodied and 3D vision-language models](https://mlsanity.com/mllm/embodied-and-3d-vision-language-models/)

## Builds on (most cited first, 13 of 13)

- [DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models](https://mlsanity.com/paper/2402.03300/) (2024)
- [MVBench: A Comprehensive Multi-modal Video Understanding Benchmark](https://mlsanity.com/paper/2311.17005/) (2023)
- [BLINK: Multimodal Large Language Models Can See but Not Perceive](https://mlsanity.com/paper/2404.12390/) (2024)
- [Reward Hacking in the Era of Large Models: Mechanisms, Emergent Misalignment, Challenges](https://mlsanity.com/paper/2604.13602/) (2026)
- [LLMs Gaming Verifiers: RLVR can Lead to Reward Hacking](https://mlsanity.com/paper/2604.15149/) (2026)
- [Countdown-Code: A Testbed for Studying The Emergence and Generalization of Reward Hacking in RLVR](https://mlsanity.com/paper/2603.07084/) (2026)
- [Understanding Language Prior of LVLMs by Contrasting Chain-of-Embedding](https://mlsanity.com/paper/2509.23050/) (2025)
- [Spurious Rewards Paradox: Mechanistically Understanding How RLVR Activates Memorization Shortcuts in LLMs](https://mlsanity.com/paper/2601.11061/) (2026)
- [VLMs Need Words: Vision Language Models Ignore Visual Detail In Favor of Semantic Anchors](https://mlsanity.com/paper/2604.02486/) (2026)
- [Adversarial Reward Auditing for Active Detection and Mitigation of Reward Hacking](https://mlsanity.com/paper/2602.01750/) (2026)
- [Do Vision-Language Models See or Guess? Measuring and Reducing Textual-Prior Reliance with a Phrasing-Controlled Benchmark](https://mlsanity.com/paper/2606.10400/) (2026)
- [From Rebound to Remedy: Understanding and Mitigating Reward Hacking via Representation Engineering](https://mlsanity.com/paper/2604.01476/) (2026)
- [Thinking with Deltas: Incentivizing Reinforcement Learning via Differential Visual Reasoning Policy](https://mlsanity.com/paper/2601.06801/) (2026)

## Built on by (most cited first, 2 of 2)

- [Where Should Language Sit in a Multimodal Model? Lessons from What Language Does to Human Perception and Cognition](https://mlsanity.com/paper/2609.07474/) (2026)
- [Don't Read the Log: Execution Traces Contaminate Verifiers in Video-Generation Agents](https://mlsanity.com/paper/2609.28564/) (2026)

## Abstract (arXiv)

Reinforcement learning with verifiable rewards (RLVR) is increasingly applied to large vision-language models (LVLMs), yet outcome-only optimization can drive a model to stop attending to the video and instead exploit linguistic priors -- a failure we call a visual shortcut. While the existence of such perception bypass is by now documented, how it forms, whether it can be undone, and when intervention still helps remain open. We treat the strength of a grounding penalty, lambda, as a control knob and characterize the formation-reversal dynamics of visual shortcuts along the training time axis. On a held-out, out-of-distribution diagnostic set, we find: (i) a sharp onset -- shortcut reliance emerges abruptly over a narrow window of optimization steps and is robust across random seeds; (ii) a monotone dose-response -- increasing lambda progressively suppresses the shortcut, and at an intermediate dose the trajectory first forms and then reverses the shortcut, exposing a hysteresis-like asymmetry between acquiring and removing it; and (iii) a critical intervention window -- applying the penalty before onset arrests shortcut formation, whereas the same penalty applied after consolidation is markedly less effective. Together these results recast visual-shortcut collapse not as a binary defect but as a controllable, time-dependent, and asymmetric process, with direct implications for when and how strongly to regularize multimodal RLVR.