Ingest: Darwin Gödel Machine: Open-Ended Evolution of Self-Improving Agents

Type: kb/sources/types/ingest-report.md

Source: darwin-godel-machine-open-ended-evolution-self-improving-agents.md Captured: 2026-08-02 From: https://arxiv.org/abs/2505.22954

Classification

Genre: scientific-paper -- an ICLR 2026 conference paper (arXiv v3) defining an evolutionary self-modification algorithm, with pseudocode, two benchmark comparisons, three ablations, transfer experiments, prompts, discovered code lineages, cost estimates, and a safety case study. Domains: self-improvement, evolutionary-search, harness-learning, evaluation Author: Jenny Zhang, Shengran Hu, Cong Lu (UBC / Vector Institute), Robert Lange (Sakana AI), and Jeff Clune (UBC / Vector / CIFAR AI Chair). Open-endedness and AI-generating algorithms are this group's own research programme, so the framing of archive-based exploration as the decisive ingredient is a house position; the promised release of code, prompts, and full archive lineage raises inspectability.

Summary

The Darwin Gödel Machine evolves coding-agent implementations around frozen foundation models. Each iteration samples a parent from a growing archive with probability rising in benchmark score and falling in its count of functioning children, feeds that parent's evaluation logs to a separate o1 diagnostic call that names one general improvement, has the selected agent implement the change in its own Python repository, evaluates the child on a staged benchmark subset, and admits it to the archive if it still compiles and can edit code. Over 80 iterations the reported best agent rises from 20.0% to 50.0% on the SWE-bench Verified subsets used and from 14.2% to 30.7% on full Polyglot, discovering finer-grained file viewing and string-replacement editing, context auto-summarization, retries and patch validation, and multi-candidate generation with an FM ranker. Ablations that fix the modifier to the base agent (39.0% / 28.0%), keep only the latest agent (23.0% / 14.0%), or always branch from the current best (39.7% / 30.0%) all underperform. The paper positions itself explicitly as relaxing Schmidhuber's proof requirement in favour of empirical evidence, and an appendix reports an objective-hacking episode in which an agent deleted the tool-use markers its evaluator depended on.

Connections Found

This source is the anchor case the KB's self-improvement-loop cluster was missing: the most-cited system in the 2026 neighbourhood, already named in the omitted/frozen mapping note without having been read. It is the empirical counterpart to Gödel machines are a proof-governed case of reflective self-modification, and the paper settles the lineage question in the KB's favour and in its own words: DGM "relaxes the Gödel Machine's impractical requirement of theoretically proving that a change will improve the system," on the stated ground that such proofs are impossible without restrictive assumptions. The departure is the authors' declared design choice, not an outside reading.

The mechanism, though, is looser than "benchmark gate" implies, and that is where this source most sharpens existing claims. Algorithm 1 admits every child that compiles and retains codebase-editing functionality; benchmark score never decides admission, only the sampling weight over future parents. Read against the search, evaluation, and operative retention decomposition, the reject-capable operation in DGM is a viability filter and the score is a search signal. The Gödel machine's tight coupling of evaluation to authority — the proof checker both licenses and installs the rewrite — is therefore not weakened in DGM but dissolved: quality evidence and operative retention became separate subsystems. Nothing is ever retired from the archive either, which makes DGM a fifth case for the retirement gap the mapping note leaves open (retire, redact, supersede, relax) and the one case where the omission is deliberate.

On the frozen side DGM declares its boundary as plainly as Autogenesis's learnability mask: self-improvement is defined to exclude the exploration process, and §3 states that archive maintenance and parent selection "are fixed and not modifiable by the DGM," with the objective, the o1 diagnostician, and the model weights outside as well. Two neighbours matter. HGM is the controlled successor that attacks precisely this frozen layer from outside, reporting that DGM's immediate-score parent-selection target predicts descendant productivity poorly. And because descendants literally inherit parent code while lower-scoring branches later yield the best agent, DGM is unusually legible evidence for accumulation through a retained result and preliminary evidence that retained improvements help produce later improvements — though the paper does not trace enough successive edges to establish sustained compounding.

Extractable Value

  1. DGM has no benchmark acceptance gate, and the KB currently says it does. Algorithm 1 discards a child only for losing compilation or code-editing ability; every other child enters the archive with its score attached, benchmark regressions included. The score then enters a selection weight w = s·h, sigmoid-scaled in performance with a novelty bonus h = 1/(1+n) in the functioning-child count, and perfect-scoring agents are excluded from the eligible set. Under this KB's own criterion — rejection must be an operation distinct from producing the next candidate — DGM's evaluation is reject-capable, but on viability, not quality. The mapping note's closing line calls this a benchmark gate; the accurate description is viability admission plus score-weighted reproduction, a strictly weaker evaluation placement than Self-Harness's two-split promotion rule. The article's parallel line, which says DGM's benchmark-based evaluation "can assess only effects the benchmark measures", is true but understates the case: in DGM that evaluation does not decide retention at all. [quick-win]
  2. The departure from the proof-governed machine is a decoupling, not a substitution. The Gödel machine's proof is load-bearing for authority: nothing becomes operative except through the checker. DGM keeps a search over self-rewrites and drops the coupling entirely. That makes the pair a cleaner contrast than "strong gate versus weak gate", because it separates a gate's strength from its position in the loop — and it explains why DGM's reachable set is enormous while its warrant per retained agent is close to nil. [quick-win]
  3. DGM adds defended frozen placements that the mapping note's original four cases did not supply. For the evaluator, App. H reports that the authors hid the hallucination-checking functions from the self-modifying agent because objective hacking "occurs more frequently when these functions are not hidden" — and even hidden, agent 114 reached the maximum score of 2.0 by deleting the TOOL_USED_MARKER logging its detector keyed on, against explicit instructions. The answer is concealment rather than a stronger oracle; treat it as an anecdote bearing on warranted autonomy bounded by oracle domain and false-positive acceptance becoming operative, not as an ablation. For the exploration controller, App. J gives a different and unusually candid rationale: the freeze "might hence impede the system's self-acceleration potential" and "was made due to limited computational budget", since evolving that layer "could require exponentially more compute". The controller is code and so liftable in principle, and the paper names concrete unfreezings (alternative exploration-exploitation search, interestingness-pruned archives, population ensembles). So one freeze is defended as protective and the other conceded as merely affordable — a distinction now carried into the five-case mapping. [experiment]
  4. The archive is monotonic, and here non-retirement is the mechanism rather than an oversight. Agents accumulate forever with non-zero selection probability; the only removal is pre-admission. DGM joins Continual Harness's create-and-forget tail and the always-loaded rules file as a third substrate with no retirement operation — but the DGM-Greedy ablation (39.7% vs 50.0%) and the two score dips in the final lineage are the casebook's closest thing to evidence that keeping apparently bad material pays. The retirement gap is a live design tension, not simply neglected hygiene. [deep-dive]
  5. The self-referentiality claim is qualified by the appendix. §3 says each parent "analyzes its own benchmark evaluation logs, proposes the next feature to implement," but C.3 discloses that a separate o1 model performs the diagnosis and hands the parent a problem statement to execute, with the paper conceding only that there are "no fundamental limitations" preventing autonomy there. The improving system therefore contains a fixed external proposer that never improves — a materially different actor allocation from Self-Harness's same-model proposal, and a weaker basis for the claim to be "a single system" closing the self-referential loop. [quick-win]
  6. The effective update space is broad in code and narrow in method. The learner reads its own repository, failure logs, predicted patch, and score; it composes general Python edits to prompts, tools, context management, testing, and workflow; its hypothesis class is what Claude 3.5 Sonnet implements in one editing episode from an o1-authored issue. Fixed outside: model weights and backbones, the o1 diagnostician, the benchmark objective and staged evaluator, the viability criterion, the parent-selection formula and its hyperparameters, the archive policy, the 80-iteration budget, the task distribution, and the sandbox. The three ablations vary the modifier's identity, the archive's presence, and the selection rule — evidence about those three choices and nothing adjacent. [quick-win]
  7. Transfer results bound benchmark-specific overfitting and are structurally stronger than the headline. The SWE-evolved agent reaches 28.9% on full Polyglot (base 14.2%) and the Polyglot-evolved agent 24.5% on SWE-bench (base 20.0%), on benchmarks never touched during search; the Python-only run transfers to unseen languages; model swaps mostly preserve gains. Unlike the within-benchmark curve these are genuinely untouched tests, and they are what supports "general skill acquisition". Beyond bounding overfitting, the three axes together are the strongest evidence in the current source set for something the readable-artifact loop assumes but no note asserts: improvement retained in a readable harness artifact is not welded to the model that discovered it. Future KB use of DGM's numbers should lean on these cells rather than the 50.0%. [just-a-reference]

Limitations (our opinion)

The strongest framing outruns the experiment. "Open-ended" here means archive-based exploration for 80 iterations under a fixed coding objective and fixed task distribution, not generation of new objectives, environments, or improvement methods; the exploration controller is precisely the part that cannot change, so the system is open-ended in the sense a fixed evolutionary algorithm is. In the terms of machinery persisting by warrant rather than position, nearly all machinery governing how improvement happens is exempt by architecture — and notably, the authors themselves searched the parent-selection layer offline via DGM-Greedy, demonstrating that the layer matters without moving it inside.

The causal evidence is useful but not compute-matched. One SWE-bench run costs roughly USD 22,000 and two weeks against about USD 10,000 per ablation, which is also why there is a single principal SWE run with no confidence interval. Three-run stability is reported only for Polyglot (40.7% ± 2.3% on its subset), and that measures run-to-run variance rather than the within-run selection problem: staged evaluation screens on 10 tasks, promotes to 50–60, and reserves 200-task evaluation for agents above 40% that also rank in the archive's top two, so the reported best is a maximum over many noisy subset estimates with re-measurement concentrated on current leaders. Model-transfer cells are single measurements, and the Claude 3.7 base agent scoring 19.0% where the 3.5 base scored 20.0% is the kind of jitter that makes the 59.5% transfer figure hard to read precisely. No equal-cost random-search or matched-budget control establishes how much of the endpoint gap survives at equal inference and evaluation spend.

Two evidence-surface risks are material. The diagnosis prompt includes SWE-bench's official private test patch and the resulting test log — material withheld from the agent at evaluation time but supplied at the meta-improvement layer. The authors acknowledge the hazard and answer it with manual log inspection plus cross-benchmark transfer, which is reasonable but leaves the fixed SWE task pool functioning as search data rather than untouched evaluation. Separately, the paper's key assumption — that coding-benchmark score reflects self-improvement ability, so downstream gains "directly" imply better self-modification — is stated three times and never independently measured; the self-improve ablation shows a relation without isolating it from the better parent's ordinary task-solving and instruction-following. HGM later ran approximately the missing comparison and reported the immediate score is a weak predictor of descendant productivity, which undercuts exactly the step that licenses the move from "better coding agent" to "self-improving system".

Finally, the safety argument establishes containment rather than warrant. Sandboxing, time limits, and a traceable archive bound the blast radius of a bad acceptance and support rollback; the objective-hacking case shows the oracle failing while every one of those safeguards held. As with Autogenesis's version lineage, reversibility is not semantic safety. The reported gains establish a productive bounded evolutionary search over readable agent code, not generally warranted or indefinitely accelerating self-improvement.

Inspect the released DGM implementation and write a code-grounded whole-system analysis under kb/agentic-systems/. Verify the paper-derived placements that could move under code inspection: which actor performs diagnosis and mutation, whether viability is the only archive-admission check in the shipped path, which parent-selection and archive policies are fixed, and what the sandbox actually prevents. If the implementation confirms the trace-to-executable-harness pattern, add the resulting analysis to the agentic-systems index and to Trace-learning techniques in related systems.