Classifier training concentrated Qwen3 hallucination signals in middle layers
Source check: Source checked. Not independent scientific confirmation.
Finding: Researchers fitted linear probes at different stages of classifier-head fine-tuning and found that hallucination labels became readable in a band of Qwen3's upper-middle layers before appearing across later layers. Combining the later layers gave no improvement over a single-layer probe in held-out comparisons. Removing 31% of the layers above the signal-forming band and a stabilization margin preserved detection scores within seed variability in Qwen3-4B and Qwen3-8B.
Technology: Qwen3-4B; Qwen3-8B; linear probes
Publication: Preprint · Test context: direct test. Catalog categories, not an evidence grade.
- AI contribution
- Classifier-head training and fitted linear probes expose where task labels are decodable in the learned models. Comparisons with sequence-to-sequence tuning, fusion probes and layer truncation test whether late-layer signals differ or are redundant for the classification.
- Test and result
- The authors tracked linear probe performance across network layers and fine-tuning checkpoints. Controlled fusion tests on a six-corpus held-out split looked for complementary upper-layer information. Three-seed depth-truncation comparisons tested what happens when different layers are removed.
- Limitation / still unknown
- The findings concern decodable binary hallucination labels in these task-trained models. They do not imply the same depth in other tasks, and measured early-exit scores varied on transfer backbones.
Evidence sources and annotation context
Evidence context (descriptions follow the cited sources; test type is a catalog classification):
Evidence sources: https://arxiv.org/abs/2610.27711 · https://arxiv.org/html/2610.27711
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 metadata and full-text HTML §§3–6, readout and transfer appendix context.
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.