# Encoded but Disconnected: Decomposing Vision-Language Model Failures under a Patching Null

arXiv: https://arxiv.org/abs/2610.00024 · HTML: https://mlsanity.com/paper/2610.00024/

## Key facts

- No citing papers in the map yet (data through 14 Oct 2026).
- Builds on 12 of its 17 references that are in the map.
- First posted on arXiv 1 Sep 2026.

## Builds on (most cited first, 12 of 12)

- [Locating and Editing Factual Associations in GPT](https://mlsanity.com/paper/2202.05262/) (2022)
- [Interpretability in the Wild: a Circuit for Indirect Object Identification in GPT-2 small](https://mlsanity.com/paper/2211.00593/) (2022)
- [Eyes Wide Shut? Exploring the Visual Shortcomings of Multimodal LLMs](https://mlsanity.com/paper/2401.06209/) (2024)
- [Mitigating Object Hallucinations in Large Vision-Language Models through Visual Contrastive Decoding](https://mlsanity.com/paper/2311.16922/) (2023)
- [Eliciting Latent Predictions from Transformers with the Tuned Lens](https://mlsanity.com/paper/2303.08112/) (2023)
- [Dissecting Recall of Factual Associations in Auto-Regressive Language Models](https://mlsanity.com/paper/2304.14767/) (2023)
- [OPERA: Alleviating Hallucination in Multi-Modal Large Language Models via Over-Trust Penalty and Retrospection-Allocation](https://mlsanity.com/paper/2311.17911/) (2023)
- [DoLa: Decoding by Contrasting Layers Improves Factuality in Large Language Models](https://mlsanity.com/paper/2309.03883/) (2023)
- [The Hidden Life of Tokens: Reducing Hallucination of Large Vision-Language Models via Visual Information Steering](https://mlsanity.com/paper/2502.03628/) (2025)
- [Multimodal Neurons in Pretrained Text-Only Transformers](https://mlsanity.com/paper/2308.01544/) (2023)
- [Evaluating Hallucination in Large Vision-Language Models based on Context-Aware Object Similarities](https://mlsanity.com/paper/2501.15046/) (2025)
- [To See or To Please: Uncovering Visual Sycophancy and Split Beliefs in VLMs](https://mlsanity.com/paper/2603.18373/) (2026)

## Abstract (arXiv)

Across three vision-language model architectures (LLaVA-1.5-7B, Qwen2.5-VL-7B, InternVL3-8B), we report a universal negative finding for mid-layer interpretability. On POPE -- the benchmark common to all three -- the mid layers encode the ground-truth answer in 68-91% of errors, yet this signal is not causally active for the final prediction: residual-stream patching yields 0% non-trivial flip at the layer level on all three architectures, and on two of three at the per-head level (Qwen: 0/12,600 patched forwards). The lone exception, InternVL3 layer-20 head-2, is a non-vocab, self-attending head whose effect is localized to that specific head (p < 1e-4). Despite the null, the errors separate operationally into three failure modes -- Perception Failure, Encoded-but-Disconnected, Prior-Override -- learnable above 60% on all three architectures, and the architecture's prior direction predicts which of two interventions elicits a category-specific response. We report these mitigation effects under oracle labels as evidence the categories are mechanistically real, not as a deployable method.