Summary
A bioRxiv preprint tested picture naming in six vision-language models by silencing each attention head in turn. The results ranged from one strongly necessary early head to no critical head, showing that disruption and computation are not the same.
A bioRxiv preprint examining six vision-language models found that the internal organisation of picture naming varies substantially between systems. In one model, removing a single early attention head almost eliminated naming accuracy. In two others, no individual head was critical enough to produce the same effect.
The result bears on a common method in mechanistic interpretability, the study of how model components contribute to behaviour. Researchers often silence a weight, neuron or attention head and treat a resulting performance drop as evidence that the component implements the behaviour. The new analysis argues that this conclusion can be too strong: a component may be necessary for a task without being the place where the task is actually computed.
How the study tested naming
The researchers used single-word picture naming, in which a vision-language model receives an image and produces the word describing it. They ablated, or silenced, each attention head one at a time and measured naming accuracy before and after the intervention.
Attention heads are parallel processing components within a transformer layer. They help the model weigh relationships among representations, such as visual features and language tokens. The experiment treated the loss of naming as an AI analogue of anomia, the naming impairment associated with aphasia, and drew on the logic of lesion-symptom mapping used in neuroscience.
The distinction between necessity and sufficiency is central. If removing a head disrupts naming, that head is necessary under the tested conditions. But retaining that head alone may not preserve naming if other components are also required.
One model showed a dominant early head
The clearest single-head effect appeared in LLaVA-1.6-Vicuna-13B. Removing head 20 in layer 0 reduced naming accuracy from 0.99 to 0.006. This identifies the head as strongly necessary for the measured behaviour in that model and task.
The same experiment also showed why this does not by itself identify the head as the location where naming is computed. When the researchers retained that head but ablated the other 1,599 heads, accuracy still fell to 0%. The head was therefore necessary, but it was not sufficient on its own.
The broader pattern varied across the six models, which used three language backbones and a range of parameter scales. LLaVA-Mistral-7B and Idefics2-8B had no critical head: silencing any single head did not produce a comparable dominant effect. All six models showed some dependence on early layers, but the strength of that dependence differed.
Scale also mattered within the Qwen2.5-VL family. The 7B model had a dominant head, while the 3B model did not. The authors report that this concentration was not explained by attention type, suggesting that the organisation of the behaviour can emerge differently even within a related model family.
Why the result matters for AI interpretability
The study places a specific limit on head-level lesion analysis. Finding the head whose removal causes the largest behavioural change can reveal a dependency, but it does not necessarily provide a complete map of the computation. In some models, picture naming may rely on information distributed across multiple heads or layers rather than being localised to one component.
That distinction matters because mechanistic explanations are intended to describe how a model produces an output, not only which interventions damage it. A successful ablation can therefore be an important clue while still requiring additional tests to determine how the affected behaviour is implemented.
The work is a bioRxiv preprint, and its experiments concern six models performing one picture-naming task rather than human language or clinical aphasia. The competing-interest statement reports that Roger D. Newman-Norlund and Julius Fridriksson have financial interests in ALLT.AI, LLC; the remaining authors declared no competing interests.