How can feedback lead to better scientific AI?
Selected advances in recursive self-improvement and the scientific feedback it needs.
Selection reviewed . Focused coverage of relevant primary sources.
RSI research · Research agents
MASS trains a shared model on selected agent trajectories, then uses the updated model to execute, optimize and evaluate the next round. It connects learning from task experience with later self-improvement.
Recursive Self-Improvement through Multi-Agent Self-Supervision
Source version checked . View the checked source.
Related method · Materials search
CISE uses calibrated reward intervals to guide materials search. It addresses a practical dataset question: how can feedback prevent promising-looking but infeasible candidates from steering later trials?
Evidence boundary. The coverage results depend on independence and covariate-shift assumptions. This is evidence about feedback reliability in materials search; recursive revision of the improvement procedure was outside the study.
Three materials-search tasks compare interval-based feedback with proxy-score baselines, with an equal-budget comparison on the photovoltaic task. All returned candidates with completed high-fidelity evaluations were feasible in the reported experiments.
Reliable Self-Evolution with Imperfect Proxy Rewards
Source version checked . View the checked source.
RSI research · Research agents
AIDE² evolves research-agent code. Its separate ignition test puts an evolved agent back into the role of improving agents—the return path that matters for recursive self-improvement.
Evidence boundary. The main evolution run uses a fixed outer improver. The ignition test does not establish a stronger or more efficient improver; scientific transfer remains limited.
The ignition test compares the evolved and original improvers. Separate held-out evaluations include one weather-prediction optimization task.
Recursive self-improvement of AI research agents
Source version checked . View the checked source.
Related method · Scientific workflows
ScienceBuddy couples harness edits and model learning. It is useful for examining the boundary between a scientific system that keeps improving and a demonstrated recursive improvement advantage.
Evidence boundary. The auxiliary reflector remains fixed. Better scientific task performance does not establish that the improvement mechanism became stronger or isolate a recursive benefit.
The study evaluates coupled model–harness cycles and separate fixed-model and fixed-harness settings.
ScienceBuddy: Recursive-in-Recursive Self-Improvement for Interactive Scientific Agents
Source version checked . View the checked source.
RSI research · General agent tasks
MetaSkill-Evolve edits the meta-skills that guide its improvement pipeline. The revised meta-skills are reused in later rounds, including subsequent edits to those meta-skills, while the base model stays frozen.
Evidence boundary. These are not scientific-discovery tasks. The baseline descriptions differ on branching and sharing, so the reported gain cannot be cleanly attributed to recursion alone.
The paper evaluates two-timescale evolution on OfficeQA, SealQA and ALFWorld, with comparisons against single-level evolution.
MetaSkill-Evolve: Recursive Self-Improvement of LLM Agents via Two-Timescale Meta-Skill Evolution
Source version checked . View the checked source.
RSI research examines the reuse of improved capabilities in later improvement rounds. Related methods address supporting needs such as reliable feedback. Each note states the task domain and evidence limits. Source checks are AI-assisted; independent replication and peer review remain separate steps.
Selection updated . Maintained by DUAL Lab.
Discovery tools collect candidates; publication follows a separate source review. The dates here describe the saved snapshot. Collection is selective, and feed access can leave gaps.
Latest 100 results per focused query on recursive improvement and scientific research agents, plus any configured paper-ID lookup. Candidate discovery is selective, not an exhaustive historical archive or automatic publication.
Recent entries from the official RSS feed, filtered for relevant topics.
Recent entries from the official RSS feed, filtered for relevant topics.
Version discovery runs separately from topic searches. A newer version is queued for review; the explanation stays tied to the version linked above.
A task, dataset or simulator can become a new resource for others to build on.