AI Science Tracker
EN

Audio and text share neurons that influence recognition in three models

University of Edinburgh; National Taiwan University · 20 Oct 2026 · computer science

Discovery

Audio and text share neurons that influence recognition in three models

Source check: Source checked. Not independent scientific confirmation.

Finding: In three audio-language models, researchers found neurons associated with comparable audio and text concepts. Adjusting their activity using text-derived information changed audio recognition more than matched random neurons in 42 of 72 model–attribute combinations. Disabling the identified neurons affected the target task before the other tested tasks in 10 of 12 model–task combinations. Effects differed across models and covered a limited set of acoustic attributes.

Technology: Qwen2.5-Omni; DeSTA2.5-Audio; Audio Flamingo 3

Publication: Preprint · Test context: direct test. Catalog categories, not an evidence grade.

AI contribution
The researchers located overlapping units in learned feed-forward activations, then altered or removed those activations to test effects on audio recognition.
Test and result
The authors evaluated four categories of acoustic recognition across three models, selected units on validation data and tested deactivation on separate recognition examples. They compared with random sets, unrelated recognition tasks and sets responsive to audio alone.
Limitation / still unknown
This result is scoped to the chosen models, labels and intervention definitions. Steering was weaker in DeSTA2.5-Audio, and the result does not establish a universal audio–text or human perceptual mechanism.
Evidence sources and annotation context

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

Evidence sources: https://arxiv.org/abs/2610.28189 · https://arxiv.org/html/2610.28189

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review date: 21 Nov 2026. Review scope: Original arXiv v1 abstract and submission history, HTML §§2–4.4 and robustness notes.

No follow-up recorded; this is not a claim that no follow-up exists. Follow-up review scope and date are not recorded. No source-supported relationships recorded.

Editorial assessment: Impact 2/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Open in the full catalog for filters, charts and related discoveries. Browse computer science discoveries.