Unchanged agent instructions could appear to improve during skill selection
Source check: Source checked. Not independent scientific confirmation.
Finding: Across 16 tuned instruction sets, held-out task success averaged only 0.1 percentage points above the starting instructions. In replay tests, the same acceptance logic often selected apparent improvements even when every candidate kept the original instructions.
Technology: DeepSeek-V4.1-Flash and GPT-6 Luna; model-driven skill optimizers
Family GPT-6: profile · models page
Publication: Preprint · Test context: direct test. Catalog categories, not an evidence grade.
- AI contribution
- Learned agents performed each task under original and edited natural-language instructions. Their repeated scores supplied the null replay and the held-out comparison.
- Test and result
- The authors tested three optimizers on three benchmarks. Their no-change replay draws task-success probabilities from a fitted beta-binomial posterior, simulates each candidate as another run of the seed instructions, and applies the original acceptance rule. Small development sets had at most 30 tasks.
- Limitation / still unknown
- The no-change result concerns the studied acceptance rules and resampling design. An aggregate near-zero test result does not establish equivalence of every skill or model. The null replay assumes episodes are independent conditional on each task’s success probability; its prior is fitted using seed development and test episodes, so the replay is a model-based diagnostic rather than a newly executed unchanged-instruction optimization trial.
Evidence sources and annotation context
Evidence context (descriptions follow the cited sources; test type is a catalog classification):
Evidence sources: https://arxiv.org/abs/2610.37201v1 · https://arxiv.org/html/2610.37201v1
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: §§4–6, Table 1 and no-change replay protocol and Appendix C.3.
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 2/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.