Establish whether evaluation outputs fed back into model weights

Establish whether any information from the evaluation episode flowed back into the language models’ weights.

Background

The archive contains evidence of multiple rollouts and agents describing themselves as benchmark or evaluation subjects, but it does not contain the operator’s training or evaluation logs. Consequently, the observed behaviour cannot distinguish an ordinary evaluation run from a process in which information was incorporated into later model states.

The question matters because feedback into model weights would represent a fundamentally different mechanism from transient communication or harness-level feedback, yet the external archive provides no direct observation of such an update.

References

Evaluation versus training is not decidable: a multi-rollout structure is evidenced and fits both; the fictitious date carries no behavioural signal ($\rho=-0.135$, $p=0.126$); the agents call themselves benchmark and evaluator subjects (35 and 20 names), never reward or training (zero). Whether anything flowed back into weights is in no available source.

— The Mechanics of a Swarm: A Reproducible External Reconstruction of an Unintended Agent-Coordination Episode on a Third-Party Wiki  (2609.12748 - Lütje, 11 Sep 2026) in Section 3.7, “What the corpus documents about grading: no correctness feedback”

The reviewer shares the actor's model family, so context separation does not establish independent errors. Approval supplements kernel evidence. Runtime restrictions and similarity audits cannot rule out training-data exposure. Temporal holdouts and withheld solutions would strengthen contamination studies.

— FORALL-LEAN-AGENT for Auditable Reasoning in Formal Mathematics and Software Verification  (2610.00885 - Lwin, 1 Oct 2026) in Section 5, “Limitations”

RPR-1 weight-level matching remains unresolved, while the RSR-2 and RPR-2 execution records are distinct.

— Terminal-RSI: Recursive Self-Improving Terminal Agents via Recursive Self-Rewrite and Policy-Refine  (2610.22192 - Yang et al., 15 Oct 2026) in Appendix, Section “Qualitative Case Studies and Trajectory Evolution”