# When Good Verifiers Go Bad: Silent Negative Transfer in Verifier-Guided VLM Training

arXiv: https://arxiv.org/abs/2606.14629 · HTML: https://mlsanity.com/paper/2606.14629/

## Key facts

- No citing papers in the map yet (data through 14 Oct 2026).
- Builds on 24 of its 27 references that are in the map.
- First posted on arXiv 12 Jun 2026.

## Builds on (most cited first, 24 of 24)

- [Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena](https://mlsanity.com/paper/2306.05685/) (2023)
- [Direct Preference Optimization: Your Language Model is Secretly a Reward Model](https://mlsanity.com/paper/2305.18290/) (2023)
- [MMMU: A Massive Multi-discipline Multimodal Understanding and Reasoning Benchmark for Expert AGI](https://mlsanity.com/paper/2311.16502/) (2023)
- [MathVista: Evaluating Mathematical Reasoning of Foundation Models in Visual Contexts](https://mlsanity.com/paper/2310.02255/) (2023)
- [Scaling Laws for Reward Model Overoptimization](https://mlsanity.com/paper/2210.10760/) (2022)
- [STaR: Bootstrapping Reasoning With Reasoning](https://mlsanity.com/paper/2203.14465/) (2022)
- [Selective Classification for Deep Neural Networks](https://mlsanity.com/paper/1705.08500/) (2017)
- [Self-Rewarding Language Models](https://mlsanity.com/paper/2401.10020/) (2024)
- [RLAIF vs. RLHF: Scaling Reinforcement Learning from Human Feedback with AI Feedback](https://mlsanity.com/paper/2309.00267/) (2023)
- [Prometheus: Inducing Fine-grained Evaluation Capability in Language Models](https://mlsanity.com/paper/2310.08491/) (2023)
- [RewardBench: Evaluating Reward Models for Language Modeling](https://mlsanity.com/paper/2403.13787/) (2024)
- [Generative Verifiers: Reward Modeling as Next-Token Prediction](https://mlsanity.com/paper/2408.15240/) (2024)
- [Prometheus 2: An Open Source Language Model Specialized in Evaluating Other Language Models](https://mlsanity.com/paper/2405.01535/) (2024)
- [Beyond Human Data: Scaling Self-Training for Problem-Solving with Language Models](https://mlsanity.com/paper/2312.06585/) (2023)
- [Learning to Reason without External Rewards](https://mlsanity.com/paper/2505.19590/) (2025)
- [Large Language Models are Inconsistent and Biased Evaluators](https://mlsanity.com/paper/2405.01724/) (2024)
- [Provably Robust DPO: Aligning Language Models with Noisy Feedback](https://mlsanity.com/paper/2403.00409/) (2024)
- [On scalable oversight with weak LLMs judging strong LLMs](https://mlsanity.com/paper/2407.04622/) (2024)
- [Towards Robust Alignment of Language Models: Distributionally Robustifying Direct Preference Optimization](https://mlsanity.com/paper/2407.07880/) (2024)
- [Self-Improving Robust Preference Optimization](https://mlsanity.com/paper/2406.01660/) (2024)
- [Theoretical Modeling of Large Language Model Self-Improvement Training Dynamics Through Solver-Verifier Gap](https://mlsanity.com/paper/2507.00075/) (2025)
- [Why is Your Language Model a Poor Implicit Reward Model?](https://mlsanity.com/paper/2507.07981/) (2025)
- [Learning from Noisy Preferences: A Semi-Supervised Learning Approach to Direct Preference Optimization](https://mlsanity.com/paper/2604.24952/) (2026)
- [Causally Robust Reward Learning from Reason-Augmented Preference Feedback](https://mlsanity.com/paper/2603.04861/) (2026)

## Abstract (arXiv)

Verifier reliability is not portable across tasks. A verifier-guided self-DPO pipeline with genuine held-out gains on MathVista (+9.6 points on self-training data, +8.0 held out) can be harmful on MMMU. The failure is invisible from the target-task self-training signal: over six learner-verifier configurations, MMMU self-training performance still improves by +3.52 points while held-out performance drops by 1.42 points. We call this silent negative transfer: a verifier validated as useful before can still show signatures of successful self-training after its induced update stops transferring to unseen data. Downstream failure alone cannot diagnose this: poor generalization may mean the verifier induces a wrong learning direction, or that the correct direction itself does not generalize beyond the self-training distribution. We separate these effects via gradient fidelity F, the alignment between verifier-induced and correct training directions, and gradient transferability T, the alignment between correct and held-out directions, giving a conservative Safe-Transfer Margin: positive first-order held-out alignment is certified when arccos(F) + arccos(T) < pi/2. We then introduce Asymmetric Acceptance Gating (AAG), a precision-first filter selecting preference pairs by the verifier's absolute confidence in the response receiving the positive update. On MMMU, AAG raises fidelity from 0.29 to 0.42. In a same-cell analysis, the correct direction stays positively aligned with the held-out direction (T = 0.31), yet raw verification rotates the induced update to negative held-out alignment (-0.13); AAG restores positive alignment (+0.22). Across all six MMMU configurations, AAG beats raw verifier-guided training. Verifier reliability thus means not that it worked before, but that it induces the right update on the task at hand and that this update stays useful beyond the data that generated it.