Interventional Training Reduces Causal Bypass in In-Context Introspection
Source check: Source checked. Not independent scientific confirmation.
Finding: Straight SFT can achieve endogenous signal-reporting performance while failing interventional checks; counterfactual intervention training improves causal introspection, especially OOD. Models are trained to introspect neuron, attention and sparse-autoencoder signals; intervention training uses counterfactually perturbed signals. Testing: Hundreds of introspectors across eight language models and three signal types are compared under endogenous and intervened/OOD conditions; generalization to unseen sites is limited.. Hundreds of introspectors across eight language models and three signal types are compared under endogenous and intervened/OOD conditions; generalization to unseen sites is limited. Source affiliation was not stated in the reviewed case checks and was unavailable in the local bibliographic cache.
Technology: Models are trained to introspect neuron, attention and sparse-autoencoder signals; intervention training uses counterfactually perturbed signals.
Publication: Preprint · Test context: direct test. Catalog categories, not an evidence grade.
- AI contribution
- Models are trained to introspect neuron, attention and sparse-autoencoder signals; intervention training uses counterfactually perturbed signals.
- Test and result
- Hundreds of introspectors across eight language models and three signal types are compared under endogenous and intervened/OOD conditions; generalization to unseen sites is limited.
- Limitation / still unknown
- Generalizability beyond the reported study population, data, and protocol was not established in the reviewed material.
Evidence sources and annotation context
Evidence context (descriptions follow the cited sources; test type is a catalog classification):
Evidence sources: https://arxiv.org/html/38911v1
Source checked · editorial details
The source page was checked against this entry. Independent confirmation is not assessed by this label.
Source type: paper. Paper: not linked in this catalog; this does not mean no paper exists.
Review date: 22 Nov 2026. Review scope: Primary result-bearing material for the stated method, result and evaluation, with bibliographic version information..
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 3/5 · 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.