Reasoning or Pattern Exploitation? Mechanistic Insights from RL-Trained Language Models
Source check: Source checked. Not independent scientific confirmation.
Finding: In 100 paired boards, patching layer 25 redirects predictions to the donor-optimal move 66 times versus 10 unpatched and 12 unrelated-donor controls. A GRPO-trained Qwen2.5-1.5B model is analyzed with sparse autoencoders and activation-patching interventions on Tic-Tac-Toe boards. Testing: Paired board interventions use unpatched, unrelated-board and earlier-layer controls; result is limited to one small model and Tic-Tac-Toe.. Paired board interventions use unpatched, unrelated-board and earlier-layer controls; result is limited to one small model and Tic-Tac-Toe. Source affiliation was not stated in the reviewed case checks and was unavailable in the local bibliographic cache.
Technology: A GRPO-trained Qwen2.5-1.5B model is analyzed with sparse autoencoders and activation-patching interventions on Tic-Tac-Toe boards.
Publication: Preprint · Test context: direct test. Catalog categories, not an evidence grade.
- AI contribution
- A GRPO-trained Qwen2.5-1.5B model is analyzed with sparse autoencoders and activation-patching interventions on Tic-Tac-Toe boards.
- Test and result
- Paired board interventions use unpatched, unrelated-board and earlier-layer controls; result is limited to one small model and Tic-Tac-Toe.
- Limitation / still unknown
- In 100 paired boards, patching layer 25 redirects predictions to the donor-optimal move 66 times versus 10 unpatched and 12 unrelated-donor controls. Paired board interventions use unpatched, unrelated-board and earlier-layer controls; result is limited to one small model and Tic-Tac-Toe.
Evidence sources and annotation context
Evidence context (descriptions follow the cited sources; test type is a catalog classification):
Evidence sources: https://arxiv.org/html/39182v1
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.