Finding What Models Rely On

An RL auditor rewarded when one model gives inconsistent outputs under content-preserving edits. Open VLMs first, then text-editing diffusion models.

Status: planned · builds on published work (BMVC 2026 oral, to appear)

The question

Our BMVC result came from an intervention we designed by hand, on a detector. Can an auditor find such interventions itself for generative and multimodal models, by being rewarded whenever one model changes its output while the content stays the same?

Evidence so far

  • Content-preserving changes can expose shortcuts. Shifting the JPEG grid by 4 pixels, with no change to content or forgery, drops DTD from 0.73 to 0.11 F1 on DocTamper (BMVC 2026).
  • A first foundation-model data point. On our ICDAR 2026 test set, Qwen3-VL-4B fine-tuned with QLoRA reaches 7.28% CER, against 3.66% for a small task-specific model. The audit is meant to explain gaps like this one; it does not yet.

Next step

  1. Audit open VLMs. An RL auditor whose actions are content-preserving edits (re-encoding, padding, re-rendering, layout changes) and whose reward is the target model’s inconsistency. Because the auditor may only use edits that keep the content fixed, it cannot score by changing the right answer, so every inconsistency it reports is a real failure rather than reward hacking.
  2. Audit text-editing diffusion models. The same auditor, with consistency measured by reading the edited text back with OCR.
  • Liu et al., Differences that matter: auditing models for capability gap discovery and rectification (AuditDM), CVPR 2026. An RL-trained auditor that maximizes disagreement between models; I audit consistency within one model.
  • Zhang et al., R-C2: cycle-consistent reinforcement learning improves multimodal reasoning, CVPR 2026. Uses cross-modal consistency as a training reward; I use within-model consistency as an audit.
  • Ribeiro et al., Beyond accuracy: behavioral testing of NLP models with CheckList, ACL 2020 (invariance tests).
  • Perez et al., Red teaming language models with language models, EMNLP 2022.
  • Tong et al., Eyes wide shut? Exploring the visual shortcomings of multimodal LLMs, CVPR 2024.

← Directions I’m exploring · Homepage