preprint · 2026-11-03
From AutoResearch to Continual AutoResearch: Do Agents Learn Across Tasks?
On this page
Why it matters here
Qin and colleagues compare persistent and reset research agents on matched task streams without updating model weights. They report that retained state helps agents produce useful solutions earlier, while final hidden-test gains vary by agent and task. Their analysis separates initialization, search and final selection; a relevant-versus-shuffled memory comparison further shows that retaining experience does not guarantee useful transfer. This is bounded evidence about reuse of research state, rather than an established improvement in the mechanism for producing future improvements.
AI-R&D capabilityResearch autonomyEvaluation integrity
What to keep in mind
- Author-reported preprint results, not independent replication or a demonstration that an improvement mechanism became better at producing future improvements. Model weights remain fixed.
- Persistent and reset conditions are matched within each native agent; comparisons across agents confound the model and its harness. Short task streams and mostly single matched run pairs limit generalization.
- Hidden-test trajectories are evaluated after each run. They are not feedback available to the agent, and best-so-far curves are distinct from the agent's final selected solution.
- The reported efficiency metric combines timing and normalized task quality under a wall-clock budget. It does not establish matched token, monetary, hardware or full prior-learning costs, human productivity or a universal RSI score.
- The relevant-versus-shuffled insight-card ablation covers the Energy stream. Selected traces illustrate reuse, anchoring and selection failures; they do not establish a causal decomposition across all tasks.
- Original methods, appendices and selected figure pixels were checked; no scores were digitized or experiments repeated. Detailed checkpoint/harness/reasoning settings and complete raw run provenance remain unverified. The paper promises code with the camera-ready version; executable artifact availability and rights were not established here.