{
  "id": 9045545,
  "title": "Head-level lesion-symptom mapping of picture naming in vision-language models",
  "url": "https://urgent.news/2026/09/21/head-level-lesion-symptom-mapping-of-picture-naming-in-vision",
  "topic": "ai",
  "section": "AI",
  "published": "2026-09-21T00:00:00.000Z",
  "source": {
    "name": "bioRxiv",
    "slug": "biorxiv",
    "url": "https://www.biorxiv.org/content/10.64898/2026.09.15.751760v1?rss=1"
  },
  "original_language": "en",
  "account": "Researchers in artificial intelligence are using techniques similar to brain lesion studies in neuroscience to understand how language models are organized. One common approach is to eliminate attention heads and observe the impact on specific behaviors. However, it remains uncertain how much this method can reveal about the precise location of a behavior within the model. The study examined this question in the context of picture naming, a task that defines aphasia. Across six vision-language models with different language backbones and parameter scales, the degree of localization varied significantly. In the LLaVA-1.6-Vicuna-13B model, a single early attention head (layer 0, head 20) was essential for naming accuracy, which dropped from 0.99 to 0.006 when removed. However, the same head was not sufficient for accurate naming when other 1,599 heads were also ablated. Two Mistral-backbone models (LLaVA-Mistral-7B and Idefics2-8B) showed no critical head. In Qwen2.5-VL, a dominant head was present in the 7B model but absent in the 3B model, suggesting that the concentration emerges with scale rather than being consistent across the model family. The findings indicate that ablating the head responsible for disrupting naming does not definitively establish that the head computes the behavior, nor does it generalize across models or suggest a head-specific localization of the behavior.",
  "summary": "Researchers in artificial intelligence increasingly intervene on language models to study how their functions are organized, silencing weights and attention components in ways reminiscent of the brain lesions long used to map language in post-stroke aphasia. Under this program of mechanistic interpretability, a common move is to ablate an attention head and read the resulting drop in a behavior…",
  "key_points": [],
  "editors_take": null,
  "illustration": null,
  "coverage": {
    "outlets": 1,
    "also_reported_by": []
  },
  "ai_generated": true,
  "disclaimer": "Summaries, key points and the editor’s take are written by software from other outlets’ reporting and may contain errors — always check the linked original."
}