AI Science Tracker
EN

Small neuron sets shifted cognitive-impairment answers in three audio models

Johns Hopkins University; Imperial College London · 23 Oct 2026 · computer science

Discovery

Small neuron sets shifted cognitive-impairment answers in three audio models

Source check: Source checked. Not independent scientific confirmation.

Finding: Researchers screened eight audio-language models on recordings and transcripts from existing cognitive-impairment datasets. In three of five models that met the analysis criteria, disabling small neuron sets shifted the models' answers more than disabling matched random sets. Separately, fitted readouts from other neurons recovered similar classification information. The experiments distinguish influence over the model's answer from where class information is represented. They do not establish clinical diagnostic performance.

Technology: Qwen3-Omni-30B; MiniCPM-o-4.5; Kimi-Audio; activation ablation

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

AI contribution
Intervening on learned decoder activations changed output tendencies; trained probes separately measured how widely class information was recoverable.
Test and result
The study analyzed existing Pitt and PROCESS-2 materials. It separated the samples used to identify neuron sets from those used for evaluation, compared ablation with matched-size, matched-layer and activity-matched controls, and evaluated fitted readouts from selected, random and all available neurons.
Limitation / still unknown
The effects apply to selected models, corpora and prompt conditions. The work does not establish clinical use or generalized diagnostic performance, and some exploratory small-readout differences did not survive correction.
Evidence sources and annotation context

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

Evidence sources: https://arxiv.org/abs/2610.32061 · https://arxiv.org/html/2610.32061

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: ArXiv v1 front matter and metadata, speaker cohort and eligibility methods, activation interventions, controls, and independent readout comparisons.

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.