{"slug": "coupled-or-independent-fault-coupling-in-agentic-medical-vision-systems", "title": "Coupled or Independent? Fault Coupling in Agentic Medical Vision Systems", "summary": "A new arXiv paper (2610.40913v1) finds that when a dermatology vision-language agent's sensor input is corrupted, its tool and reasoning channels can fail together, producing 55% joint wrong agreement across all cases (71% of faulty cases) with crop-first retrieval versus 35% (43%) with full-image retrieval on PAD-UFES-20. The authors define a coupling taxonomy separating single-channel errors from joint errors that agree or disagree, and report that typographic injection produces detectable directional retrieval capture while physical corruption does not. Because binary decisions force agreement when both channels are wrong, hard-label channel comparison cannot detect joint failures, limiting disagreement recall to 0.29 for crop-first versus 0.57 for full-image, with the order effect reproduced on the independent HIBA test set.", "body_md": "arXiv:2610.40913v1 Announce Type: new \nAbstract: Vision-language agents equipped with tools expose intermediate outputs for inspection and comparison with the model's reasoning, which may offer safety benefits over single-shot models. This paper tests that potential benefit when the sensor reading itself is untrusted, so that the tool and reasoning channels share a common visual fault. We consider a dermatology agent based on segmentation, contrastive retrieval, and an iterative vision-language reasoner and define a coupling taxonomy that distinguishes single-channel errors from joint errors that agree or disagree. On PAD-UFES-20, we show that typographic injection produces detectable directional retrieval capture, whereas physical corruption does not. Crop-first retrieval preserves the payload at a readable scale and produces 55% joint wrong agreement across all cases (71% of faulty cases); full-image retrieval produces 35% (43%). Under a binary decision, if both channels are wrong, their predictions must agree, so hard-label channel comparison cannot detect joint failures, limiting disagreement recall to 0.29 (crop-first) versus 0.57 (full-image). The direction of the order effect is reproduced on the independent HIBA test set.", "url": "https://wpnews.pro/news/coupled-or-independent-fault-coupling-in-agentic-medical-vision-systems", "canonical_source": "https://arxiv.org/abs/2610.40913", "published_at": "2026-11-02 05:00:00+00:00", "updated_at": "2026-11-02 05:21:45.393560+00:00", "lang": "en", "topics": ["ai-safety", "ai-research", "computer-vision", "large-language-models", "ai-agents"], "entities": ["arXiv", "PAD-UFES-20", "HIBA"], "also_reported_by": [], "alternates": {"html": "https://wpnews.pro/news/coupled-or-independent-fault-coupling-in-agentic-medical-vision-systems", "markdown": "https://wpnews.pro/news/coupled-or-independent-fault-coupling-in-agentic-medical-vision-systems.md", "text": "https://wpnews.pro/news/coupled-or-independent-fault-coupling-in-agentic-medical-vision-systems.txt", "jsonld": "https://wpnews.pro/news/coupled-or-independent-fault-coupling-in-agentic-medical-vision-systems.jsonld"}}