# Readout is not Recovery: Dissociating Coordinate Emission from Visual-Corruption Repair in Vision-Language Models

arXiv: https://arxiv.org/abs/2609.32757 · HTML: https://mlsanity.com/paper/2609.32757/

## Key facts

- No citing papers in the map yet (data through 14 Oct 2026).
- Placed in Visual token pruning and efficiency, one of 16 directions in Multimodal LLMs / vision-language models.
- Builds on 18 of its 21 references that are in the map.
- First posted on arXiv 26 Sep 2026.

## Placed in

- Multimodal LLMs / vision-language models / [Visual token pruning and efficiency](https://mlsanity.com/mllm/visual-token-pruning-and-efficiency/)

## Builds on (most cited first, 18 of 18)

- [Microsoft COCO: Common Objects in Context](https://mlsanity.com/paper/1405.0312/) (2014)
- [Layer Normalization](https://mlsanity.com/paper/1607.06450/) (2016)
- [Benchmarking Neural Network Robustness to Common Corruptions and Perturbations](https://mlsanity.com/paper/1903.12261/) (2019)
- [Qwen3-VL Technical Report](https://mlsanity.com/paper/2511.21631/) (2025)
- [Locating and Editing Factual Associations in GPT](https://mlsanity.com/paper/2202.05262/) (2022)
- [Root Mean Square Layer Normalization](https://mlsanity.com/paper/1910.07467/) (2019)
- [Interpretability in the Wild: a Circuit for Indirect Object Identification in GPT-2 small](https://mlsanity.com/paper/2211.00593/) (2022)
- [Kosmos-2: Grounding Multimodal Large Language Models to the World](https://mlsanity.com/paper/2306.14824/) (2023)
- [Attention is not Explanation](https://mlsanity.com/paper/1902.10186/) (2019)
- [Shikra: Unleashing Multimodal LLM's Referential Dialogue Magic](https://mlsanity.com/paper/2306.15195/) (2023)
- [Towards Automated Circuit Discovery for Mechanistic Interpretability](https://mlsanity.com/paper/2304.14997/) (2023)
- [Towards Faithfully Interpretable NLP Systems: How should we define and evaluate faithfulness?](https://mlsanity.com/paper/2004.03685/) (2020)
- [Ferret: Refer and Ground Anything Anywhere at Any Granularity](https://mlsanity.com/paper/2310.07704/) (2023)
- [Is Attention Interpretable?](https://mlsanity.com/paper/1906.03731/) (2019)
- [Kimi-VL Technical Report](https://mlsanity.com/paper/2504.07491/) (2025)
- [Pix2seq: A Language Modeling Framework for Object Detection](https://mlsanity.com/paper/2109.10852/) (2021)
- [Benchmarking Robustness in Object Detection: Autonomous Driving when Winter is Coming](https://mlsanity.com/paper/1907.07484/) (2019)
- [Robust Object Detection under Occlusion with Context-Aware CompositionalNets](https://mlsanity.com/paper/2005.11643/) (2020)

## Abstract (arXiv)

VLM bounding-box localization is both language generation and spatial commitment. Parseable fields such as bbox2d make localization easy to score, but dimensions that emit coordinate tokens need not repair localization after visual evidence is damaged. We study this readout/recovery separation in Qwen3-VL-4B-Instruct on single-object COCO grounding. We compare clean coordinate-token readout rankings with corruption-derived repair rankings, using object-mask endpoint replacement for recovery and clean-input flooring for depth localization. In Qwen3-VL, coordinate-token rankings are inert through layer 24, load-bearing from layers 32-35, and peak at layer 34; corruption-derived rankings harm layers 16-24 but become beneficial near layer 35/final. A Kimi-VL-A3B diagnostic shows a matching output-proximal transition despite a different box format. Object-mask recovery separates rank budgets: k=250 shows necessity, k=500 shows Top-k restoration above random, and k=d/2 is largely capacity-driven. Partial-occlusion sweeps reveal that high-overlap coordinate-token sets can hurt at k=1000 and help mainly at half-width, while population corruption-derived sets provide no reliable fixed repair set. Edge-attribution patching shows coordinate-token paths are high precision but low recall for detection recovery, and RMSNorm quasi-layer controls do not close the endpoint-repair gap. Endpoint coordinate triage is therefore a useful circuit prior, but occlusion recovery requires a separate benchmark.