Topics

Dream-RSI replays completed discovery trees to evaluate and rewrite exploration-policy code under fixed models and evaluators. A small tree shows how two policies reach different observations from the same starting information. The Gemini-3.1-Pro Lasso comparison reports 317 discovery calls versus 550 and lower mean solver runtime, but the paper attributes that mean improvement to RCV1 while the other five held-out datasets become slower. Incumbent retention protects the mean replay score only on the identical fixed history pool.

RSI-Anything reports hidden-test all-pass rates of 26.3%, 29.2% and 31.2% for static verification, fixed-world-model selection and adaptive-world-model selection with the same coder and candidate budget. Only the executed patch supplies observed targets for its predictions. Feedback can accumulate across steps before an update-and-acceptance cycle changes either component. The five applications exercise different update pathways, leaving sustained mutual improvement within one task unresolved.

AlphaEvolve searched for a reusable matrix-tiling rule, judging its choices by actual runtime on real TPUs. Half the collected input shapes guided optimization; the other half evaluated the resulting rule. The reported production benefit connects model-generated code to model-training infrastructure, while a later model’s ability to discover better algorithms remains unmeasured in the supplied case.

Mythos Preview’s 64% score comes from 129 deliberately selected human-detour moments. A separate sample of 127 strong human decisions yielded about 20% model preference, without per-model results. Agents also designed every experiment in a human-framed weak-to-strong project, recovering 97% of its defined performance gap; that result did not transfer cleanly to production-scale models. These findings show progress in research suggestions and methods, without demonstrating autonomous successor development.

DGM v1 evolves coding-agent tools and workflows around frozen foundation models. Its archive preserves functional children through score dips, and two controls support the contribution of changed modifiers and alternative branches. SWE-bench and Polyglot gains require their specific evaluation regimes; Appendix F shows how a higher score can instead come from changing the detector's evidence.

The paper reports higher aggregate pass rates with cumulative tools alongside negative old-cohort backward transfer for its deployment baselines. These results use different baselines. Reading the final row for ACC and pairing introduction and final cells for BWT explains the distinction; memory controls and a no-memory AutoGen counterexample keep the conclusion bounded.

MetaRSI distinguishes changes to the running system from changes to proposal and scheduling policies. Its same-S1 fork reports a 2.3-point advantage for the updated improver under the same next-round budget. A separate composition comparison and a five-round cumulative gap answer different questions. The lead persists while per-round gains shrink, and the fork alone does not establish net returns including meta-update costs.

The Personas agent completed substantial research engineering, but its experiments kept changing what the paper could claim. Following the transfer test, cross-model repetitions and later rival comparison explains why more experimental work did not establish the proposed contribution. The case also preserves a useful narrow finding and the limits of extrapolating to AI research or RSI generally.

OpenEvolve searches through proposed versions of an input program and evaluates them. Its edit-boundary option is meant to preserve text outside marked regions, such as a function name that callers depend on. At the captured v0.4.0 commit, the parallel controller omits that option when transferring configuration, so the worker defaults to False. One synthetic example shows the consequence and a precise diagnostic: inspect the worker's setting and check that evaluation receives both the preserved name and the permitted inner edit.

RRSI changes the procedures around a fixed model. One Harvey LAB comparison connects repeated grading feedback to the instructions and file checks that survive the search, then tests their usefulness elsewhere. Both evolved methods help, but their selection scores give the wrong ranking for cross-benchmark transfer.

ScientistTwo's reported DMC-TeCh rejection separates a candidate's improved performance from evidence for its proposed mechanism. Its feedback loops can return to experiments and code; manuscript ratings and artifact checks then assess other properties of the output.

SEAL connects two learning targets: learning from passage-derived text, and learning to generate text that works well when trained on. Candidate materials undergo separate parameter adaptations and question-answer evaluations; selected passage–edit pairs train the generator. The authors report 39.7%→47.0% accuracy with the implications prompt and 49.4%→55.6% with the rewrite prompt. A pooled 200-passage run reaches 43.8%, while a separate sequential test finds declining earlier-task performance. These results support more effective adaptation under explicit evaluation, without establishing reliable indefinite accumulation or open-ended recursive self-improvement. This article provides no independent replication results.

AlphaEvolve reports a 23% kernel speedup and a 1% reduction in overall Gemini training time. Its program database also feeds later prompts, but the deployment result leaves subsequent proposal capability unmeasured. Comparing that return path with SEAL’s learned edit generation, DGM’s evolving code and archive, and MetaRSI’s same-start fork shows what each success actually carries forward.