What the evidence supports
External evidence
Selected results make this a serious research direction. DeepSeek-R1 reports reasoning gains from reinforcement learning on a pretrained base. Snell et al. find that allocating inference computation according to problem difficulty improves the efficiency of search in their mathematical setting. AlphaEvolve combines language models, evolutionary program search and automated evaluation to report mathematical and infrastructure improvements. These are different mechanisms with different evidence.
They establish useful capabilities in the settings studied. A checked construction or faster kernel does not establish a general ability to solve open problems, and an improved research tool does not establish indefinitely accelerating intelligence. Rubedo has not reproduced these systems or run the experiments proposed below. Our reading map records sources and their limits.
Better task performance: more correct answers under a stated budget. New discovery: an independently checked contribution beyond documented prior work. Better discovery process: an improvement that makes subsequent, unfamiliar investigations more effective after its full cost is counted. Each requires its own evidence.
Post-training and agentic reinforcement learning
Frontier post-training can include supervised demonstrations, preference optimization, distillation, outcome rewards, process feedback and reinforcement learning with verifiable rewards (RLVR). These supply different information. Teacher demonstrations can introduce solutions; a verifier can select correct attempts; an interactive environment can supply observations unavailable in the original prompt. An agent repeatedly calling tools with frozen weights is performing inference and interaction. Learning a policy from those trajectories is a separate training step.
Open question. When does RL increase reliable access to existing strategies, and when does it expand the set of problems a model can solve? Yue et al. find improved one-attempt performance alongside limited or reduced large-sample coverage in the RLVR setups they test. A 2026 curriculum RL preprint reports broader coverage using targeted teacher guidance. Because that adds information, it is not an isolated test of reward-only learning.
Investigation. Compare a fixed base model, supervised fine-tuning, RLVR, teacher-assisted curricula and multi-turn agentic RL on controlled code or proof tasks. Hold task families, tools and inference limits fixed; account separately for training and teacher costs. Plot pass@1 and pass@k, retention and transfer to withheld task families. Pass@k measures whether at least one of k attempts succeeds; it does not tell us whether the system can select that attempt without an oracle.
Uncertainty. Failure to sample a solution cannot prove that its probability was zero. Sampling temperature, truncation, contamination, verifier quality and training stability can change the apparent boundary. Sparse rewards, long-horizon credit assignment, exploration collapse and loss of earlier competence may prevent progress before representational limits are reached.
Inference-time search and tool use
A tool-use harness is the surrounding system that manages observations, retrieval, code execution, candidate branching, retries, context and stopping. Best-of-N sampling, sequential revision and tree search spend computation differently. A verifier or search controller may contribute much of the final capability. ReAct provides an early demonstration of interleaved reasoning and action; prompting such a loop does not by itself update model weights.
Investigation. Freeze the model and compare direct answers, independent sampling, verifier-guided search and tool-assisted revision. Record tokens, tool calls, evaluator calls, latency and failures, including the verifier's own cost. Report an accuracy–cost curve and the gap between oracle success and actually selected answers. Then test whether distilling verified trajectories retains any improvement when the extra search is removed.
Drawbacks. Correlated candidates can make larger searches redundant. Repeatedly optimizing against a weak verifier selects its mistakes. Long traces can reinforce a wrong premise, tools can fail, and permissive retries can hide unreliable behavior. Compute accounting must cover the whole system.
Coordination and shared memory
Rubedo hypothesis
Independent proposals, specialist roles and reusable records may improve discovery when work can be decomposed and checked. More agents can also duplicate effort and repeat one another's errors. Controlled agent-scaling evaluations report gains and losses that depend on task structure and coordination. Shared model ancestry limits how independent apparent agreement really is.
Investigation. Compare one agent with the same total budget against independent workers, a coordinator and workers, and a shared-memory group. Test separate ablations for role specialization, communication and memory. Include both separable tasks and tasks with sequential dependencies; rotate roles and seeds. Measure verified success, duplicate work, communication overhead, error propagation and the ability of a fresh agent to reuse an artifact.
Memory design. Store candidate claims separately from accepted artifacts. Each record needs provenance, dependencies, verifier version, uncertainty and an expiry or revalidation rule. Compare private, shared and selectively shared archives under the same storage limit. Inject stale or contradictory records and measure recovery. A consensus vote cannot replace independent verification. This extends our collective memory and agent architecture questions.
Evolution, quality-diversity and curricula
Rubedo hypothesis
An evolutionary outer loop could preserve strategies that an immediate reward signal would discard, while gradient learning improves each candidate locally. Quality-diversity (QD) search retains good candidates across behavior niches: MAP-Elites makes this concrete. Candidate objects might be programs, tool workflows, memory policies, curricula or small adapters. Evolving a whole frontier model would have a very different cost and is not our proposed first experiment.
Investigation. On a small family of deceptive program-synthesis tasks, compare equal-budget random sampling, reward-only selection, novelty search and a QD archive. Keep the model and verifier fixed. Define descriptors from observable behavior, such as algorithm family, resource usage or tool sequence; test sensitivity to those choices. Compare fixed task ordering, random curricula and curricula selected for measured learning progress, with an external held-out task distribution.
What would count. Archived candidates should help solve later withheld problems, survive environment changes or improve learning efficiency. Archive occupancy and text diversity alone are insufficient. Remove the archive and replay its candidates to test whether preserved stepping stones caused later gains. See preservation and the proposed pilots.
Drawbacks. Descriptors embed human assumptions; learned descriptors can drift or collapse. A curriculum can manufacture easy progress by choosing easier tasks. Evaluation noise can preserve lucky winners, recombination can break compatible parts, and diversity consumes memory and compute. Evolutionary search is complementary only if its marginal gains justify those costs.
Verification, novelty and attribution
Correctness, novelty and significance are separate judgments. A proof checker establishes that a formal statement follows from the admitted axioms and dependencies, assuming its trusted kernel is sound. It does not establish that the statement faithfully expresses the intended open problem. Passing tests provides evidence about tested behavior, not a universal proof of program correctness.
Investigation. Use a small task suite with executable specifications or formal checks. Freeze the task split and evaluator before search. Keep final tasks inaccessible to candidate generation and curriculum selection; limit validation queries, and use a fresh audit set after repeated development. Add mutation tests, counterexamples, independent checker implementations where feasible, and manual checks of specification fidelity. Record invalid proposals and false acceptances, not only winners.
For claims about open mathematics, require the exact statement, prior status, dated artifact, proof or reproducible construction, independent expert assessment and a search of prior work. Record human problem selection, formalization, hints, proof repair and selection, alongside the model, tools, attempts and compute. Rediscovery, benchmark success, a conjecture, a numerical bound and a new theorem deserve different descriptions. We do not infer a count of AI-solved open problems or mathematicians' reactions from anecdotes.
Limit. Cheap, decisive feedback is much more available in some code and formal tasks than in physical science or ambiguous social objectives. The verifier and environment can become the bottleneck. Our evaluation question concerns both reliability and what a score actually values.
Meta-optimization and recursive improvement
Meta-optimization changes how search or learning operates: its mutation operators, resource allocation, tool interfaces, memory rules or update procedures. The Darwin Gödel Machine reports benchmark improvements from evolving coding-agent implementations. That is evidence for bounded improvement of an agent system. AlphaEvolve's reported infrastructure optimizations provide another route by improving resources used in AI development.
Open question. Can a changed discovery process reliably produce further improvements in itself, across unfamiliar tasks and successive generations, after paying for proposal generation, evaluation and deployment? Better task reward, a better harness, improved model training and sustained recursive improvement occupy different levels of this question.
Investigation. Begin with editable search controls and a frozen base model. Compare a fixed controller, hand-tuned and random-search controllers, and a learned or evolved controller. Give each the same total optimization budget, include amortization across future tasks, and evaluate several rounds on new task families with frozen external checks. Measure learning curves, regressions and the rate of accepted improvements. Disable specific changes to establish their contribution.
Limits. Faster kernels need not improve reasoning; a better harness can exhaust the base model's useful proposals. Evaluation may get more expensive as improvements get smaller. Human feedback, experiments, hardware, energy and coordination remain constraints. A system can optimize an evaluator while degrading the intended objective. No finite sequence establishes indefinite or exponential improvement. This connects to evolvability and the physical substrate.
How the tracks fit together
Proposed synthesis
A pretrained model supplies proposals; search and tools generate experience; external checks filter artifacts; a diverse archive preserves alternatives; coordination distributes work; training may consolidate useful behaviors; and a slower outer loop adjusts the process. Whether each addition pays for itself is the research question. Rubedo has selected no combined architecture.
| Connection | Concrete test | Reason to revise it |
|---|---|---|
| Search → post-training | Distill checked trajectories; test new tasks with the original inference budget. | Gains disappear without the search or fail to transfer. |
| QD → shared memory | Give a fresh agent a bounded archive; compare it with reward-ranked and random archives. | Storage grows but reusable competence does not. |
| Curricula → agentic RL | Select training tasks by learning progress; keep a fixed external transfer set. | Progress comes from task selection or extra teacher information alone. |
| Coordination → discovery | Compare decomposed work with a matched-budget single agent and independent sampling. | Communication costs exceed verified gains or errors become correlated. |
| Meta-optimization → further improvement | Run successive controller generations on fresh tasks; count all search costs. | Later rounds plateau, regress or only exploit the evaluator. |
Failure modes and research boundaries
Goodhart risks grow when selection repeatedly exploits the same proxy. A reward can overvalue style, an archive can reward meaningless novelty, and a shared memory can make an unsupported claim look established. Contamination and repeated validation reuse can create apparent generalization. Competitive agents can collude, and tool observations or memory entries can carry adversarial instructions.
For the proposed pilots, candidate code runs in a restricted sandbox with explicit time and resource limits. Keep evaluation code and final task data outside candidate write access; use least-privilege tools, provenance logs and reversible checkpoints. Human review remains necessary before expanding tool authority or adopting changes to the evaluator. These are boundaries to test, not guarantees of safety.
A useful negative result would identify where a loop stops paying: proposal quality, exploration, verification, coordination, retention or physical cost. That would sharpen the inquiry just as much as a measured improvement. The next step is the smallest experiments that distinguish those explanations.