A-Evolve-Training: Autonomous Post-Training of a 30B Model

Contribution

• The paper positions this as a first-at-scale publicly reported autonomous post-training run, extending prior public demonstrations from roughly GPT-2-class budgets to a 30B, multi-week campaign.The claim is deliberately scoped to one specified, auditable datapoint rather than a milestone or first autonomous match of human researchers.Source: §1 · pp. 1–2

Motivation

• Frontier post-training makes execution and infrastructure grow by orders of magnitude, turning retries, branching, and exhaustive sweeps from cheap actions into binding constraints.At 30B, each repeat is a full training run, so measurement and execution constrain the loop beyond idea generation.Source: §1 · p. 2, Table 1 · p. 2

System design

• The system uses an immutable reference substrate, eight homogeneous full-stack workers, and constitutionally bounded round-level policy updates.Workers edit recipes and data, run jobs, debug, evaluate, and report without carrying state across rounds; only the rolling search policy changes.Source: §1 · pp. 2–3
• Its central design principle is asymmetric freedom: preserve invariants needed for comparability while maximizing freedom in axes where exploration creates value.The substrate is never overwritten, and promotion occurs in the rolling policy rather than by replacing the reference workspace.Source: §1 · p. 3

Results

• 0.86 on the public leaderboard versus 0.87 for the top submission, ranking 8th among roughly 4000 entries after four autonomous rounds.The result is reported as an auditable datapoint of a frontier-scale autonomous loop closing, not as a claim of matching human researchers.Source: §1 · p. 3

Discovery

• The loop inverted its search objective after interventions raised the internal metric to record highs without improving the external target.It sought interventions that could lower the proxy while improving the external target, including domain rebalancing and reweighted checkpoint selection.Source: §1 · p. 3

LLM-as-scientist at small scale

• Small-scale LLM-as-scientist systems form hypotheses, run experiments, and write findings without human intervention, but rely on cheap experiments and freely retryable GPT-2-scale training.Recent evaluation work reports that frontier agents still struggle to reimplement known improvements even in that forgiving regime.Source: §2 · p. 4

LLM-guided evolutionary code search

• AlphaEvolve and related systems evolve program code under LLM mutation, whereas this system evolves the code and data that produce a post-trained model.Both use LLM-driven mutation, but their feedback structures and evaluation costs differ by orders of magnitude.Source: §2 · pp. 4–5

Classical AutoML and hyperparameter search

• Classical AutoML and hyperparameter optimization search coordinates in a fixed parameterized pipeline, while this loop edits the training recipe itself.Recipe-level changes include data construction, mixture weights, and trainer configuration; the reported upsample finding could not be expressed by a fixed-grid search.Source: §2 · p. 5

Where this work sits

• Figure 1 places this work alone in the multi-week per-trial regime, where the loop must be restructured around evaluation cost rather than treated as a cheap retry cycle.The figure compares what is searched against the cost of one evaluation across related approaches.Source: §2 · p. 5

Pillar 1: an immutable reference substrate

• An immutable default/ substrate is forked for every round, so recipes remain comparable against the same base checkpoint, data pipeline, training code, and evaluation setup.Winning recipes never overwrite the substrate or become the base for later rounds.Source: §3 · p. 5

Pillar 2: homogeneous, memory-free workers

• Identical memory-free workers explore assigned axes instead of exchanging specialized-role mid-states, because mid-state compounding breaks recipe comparability and corrupts selection signals.The earlier specialized-role design did not work in practice.Source: §3 · pp. 5, 7

Pillar 3: round-level evidence aggregation and policy-only promotion

• Each round promotes only a revised search policy: manifests and fixed-schema evidence summaries feed a constitutionally bounded meta agent while the substrate remains unchanged.The evidence artifact records candidate summaries, cross-candidate patterns, dev/leaderboard calibration, surprises, and rationale audits.Source: §3 · p. 7
• The rolling policy updates standing recipes, search axes, dead ends, and empirical promotion rules, while the frozen constitution enforces the system-level meta-rules.The policy can revise empirical rules such as promotion thresholds as evidence accumulates.Source: §3 · p. 7

Cross-cutting principle

• Asymmetric freedom freezes the substrate, evaluation harness, and comparison baseline while maximizing exploration over semantic mutations.This preserves runnable, comparable trials while directing search capacity toward changes that can create value.Source: §3 · p. 7

4 Results

• The campaign used N=8 workers per round on a 30B Nemotron base, ran for four rounds, and required no human intervention after initial substrate authoring.Evaluation used the NVIDIA Nemotron-Reasoning Challenge, whose benchmark consists of logical reasoning puzzles across domains including bit manipulation and algebraic equations.Source: §4 · p. 8
• 0.86 versus 0.87 for the top human submission placed the autonomous model 8th among approximately 4,000 leaderboard entries.The score was reached after four rounds, with monotonic improvement across rounds.Source: Figure 3 · p. 8, §4 · p. 8

5 Discovery: when the proxy stopped being evidence

• Dev scores rose to 0.93–0.94 and the weak equation domain reached as high as 0.82, yet external performance stayed at 0.84–0.85.These results falsified the assumption that maximizing the easiest proxy dimension would improve the external target.Source: §5 · p. 8
• Downweighting the apparently weak domain lowered its internal score without degrading the external target, showing that it was not the limiting factor for real performance.Together with the failed high-dev interventions, this provided positive and negative single-variable probes of the proxy’s failure.Source: §5 · p. 8
• The next policy asked for interventions that could lower the proxy while improving the external target, including rebalancing domains and reweighting checkpoint selection.Raw development gains became suspect unless they survived an external probe.Source: §5 · p. 9
• The conclusion is scoped to this performance band: the internal metric was not shown useless, but ceased to be a reliable causal guide beyond that range.The observation rests on multiple probes rather than a single lucky run.Source: §5 · p. 9

6.1 Scaled autonomous loops can revise the rules of their own search, not only run within them

• A single unreplicated campaign provides a verifiable revision of the search policy: the loop detected that its measurement frame had stopped being evidence and changed what workers maximized.The authors distinguish this discovery from optimization within a fixed measurement frame.Source: §6.1 · p. 9

6.2 A system-design lesson: asymmetric freedom and memory-free workers

• The working design freezes one immutable substrate, fixed evaluation, and memory-free workers while allowing freedom only where variation creates value.Re-forking preserves cross-round comparability and avoids compounding unobserved variance from shared mid-states.Source: §6.2 · p. 9
• The substrate and constitution remain immutable across rounds, while a rolling search policy alone mutates under fixed meta-rules.This split prevents goal drift and substrate contamination while allowing the system to revise its noise band and retire axis categories.Source: §6.2 · pp. 9–10

6.3 Mapping back to the operational RSI framing

• The 30B Nemotron campaign clears the operational bar once, with four self-directed rounds, no human in the loop, and a final score within one point of the top human submission.The authors treat this as evidence that autonomous frontier post-training is reachable, not that recursive self-improvement is solved.Source: §6.3 · p. 10

• The result covers one post-training task family, leaving out-of-prior findings on other domains open.Source: §7 · p. 10
• The system was run on one 30B Nemotron base, so replication across bases and sizes is needed to separate loop behavior from checkpoint-specific behavior.Source: §7 · p. 10
• The optimization claim relies on one public leaderboard, making broader benchmarking important for reducing dependence on a single metric.Source: §7 · p. 10
• The system uses an internal A-Evolve framework version, and release timing for code, substrate, and checkpoint remains unconfirmed.Source: §7 · p. 10

• The trajectory moves from internal-proxy optimization to detecting proxy–target divergence and finally changing the search policy.Source: Table 2 · p. 12