agentic-systems · 2026-08-12 · Tier 2

Mendel Gödel Machine: self-modification from the archive, not from the last failure

Mendel Gödel Machine: self-modification from the archive, not from the last failure

Source: HuggingFace Daily Papers, 2026-08-12 · arXiv 2608.07645 · raw

Authors: Changzhi Liu (UESTC), Yilun Liu, Sikuan Yan, Volker Tresp, Yunpu Ma (LMU Munich / Munich Center for Machine Learning).

TL;DR. Self-improving coding agents rewrite their own source code. Every published version of the loop derives each rewrite from one failure trajectory on one task, which throws away everything the agent's archive of past attempts already knows. MGM adds two self-modification operators that read the archive instead. Reaction-norm mutation edits an agent using its trajectories across multiple tasks at once, so a recurring failure can be told apart from a one-off. Cross-lineage hybridization edits an agent using the trajectory of a reference agent from a different lineage on the same task, so a behavior that worked elsewhere can be transferred. The paper proves faster convergence under an additive fitness-landscape model, confirms it in controlled surrogate simulation, and reports consistent gains on SWE-bench and Polyglot.


flowchart LR
  A[(Archive of past<br/>agent variants<br/>+ trajectories)] --> B{Choose<br/>mutation type}
  B -->|baseline| C[Clonal mutation<br/>one failure, one task]
  B -->|new| D[Reaction-norm mutation<br/>one agent, many tasks]
  B -->|new| E[Cross-lineage hybridization<br/>two lineages, one task]
  C --> F[Edited agent]
  D --> F
  E --> F
  F --> G[Evaluate on<br/>SWE-bench / Polyglot]
  G --> A
  D -.distinguishes.-> H[Genotype defect<br/>vs transient error]
  E -.transfers.-> I[Behavioral traits<br/>across lineages]
  classDef input fill:#dbeafe,stroke:#3b82f6,color:#1e3a8a
  classDef decision fill:#fef3c7,stroke:#f59e0b,color:#78350f
  classDef output fill:#d1fae5,stroke:#10b981,color:#065f46
  classDef aux fill:#e0e7ff,stroke:#6366f1,color:#312e81
  class A input
  class B decision
  class C,D,E,F,G output
  class H,I aux

The argument

The genetics framing is not decoration, it is the actual claim. A reaction norm in biology is the pattern of phenotypes one genotype produces across a range of environments. Here the "environments" are tasks and the "phenotype" is the success-or-failure profile. The point is diagnostic: if an agent fails one task, the failure could be a defect in the agent or an accident of that task. If the same failure mode recurs across several tasks, it is a property of the agent. Conditioning a rewrite on a single trajectory cannot make that distinction, so single-trajectory loops spend edits fixing noise.

Cross-lineage hybridization targets the other waste. When several agents from different evolutionary branches attempt the same task, their divergent trajectories are a natural controlled comparison. A failing agent can be edited using a successful sibling's trajectory as the contrast, and the paper notes the case of comparing two failing trajectories to identify complementary weaknesses, which is the more interesting version because it does not require a success to learn from.

The theoretical result is under an additive fitness landscape, which is a strong assumption: it means an agent's fitness decomposes into independent per-component contributions. That is precisely the regime where recombining components from two lineages should help, so the proof is somewhat load-bearing on an assumption that favors the method. The controlled surrogate simulation is the honest part of the evidence; SWE-bench and Polyglot are the realistic part.

How this relates to what the wiki already knows

It is the first paper to improve the self-modification operator rather than the loop around it. The 08-11 harness-evolution cluster had three papers about the loop: Ouroboros, the self-developing coding agent whose tools, prompts and core implementation improve through reviewed commits, scoring 86.74% on Terminal-Bench 2.1 with a 161-day live deployment; Evo-Bench, which isolates harness-evolving skill from base-model strength and finds up to 16.6 absolute points but reports early saturation; and A²E, which finds no model-harness combination wins across all task types. All three take "the agent edits itself given a failure" as a primitive. MGM opens that primitive up and asks what evidence the edit should be conditioned on. That is a genuinely different axis and it means the two lines compose rather than compete.

And it is a candidate explanation for Evo-Bench's early saturation, which this wiki called "the single most important unexplained result" on the self-evolving page. If each cycle conditions on one trajectory, the edits are high-variance and largely corrective of noise, so the improvement curve should flatten once the easy single-trajectory-visible defects are gone. MGM's whole premise is that the archive contains signal that single-trajectory conditioning cannot see. The falsifiable version: run Evo-Bench with MGM's operators and see whether the saturation point moves. The paper does not do this, and it is the experiment that would matter most.

Its convergence claim sits awkwardly next to the day's benchmark results. MGM reports "consistent improvement in performance, efficiency and generalizability" on SWE-bench, and SWE-Bench ProMax (08-11) cited an audit finding nearly 60% of unsolved SWE-bench Verified instances contain flawed tests, with frontier models able to reproduce gold patches verbatim from training data. An evolutionary loop optimizing against a partly broken fitness signal will happily evolve toward the broken part. Nothing in MGM's design detects that, and its archive-based operators arguably amplify it, because a defect that reliably pays off across many tasks is exactly what a reaction-norm mutation is built to lock in.

Cost remains unpriced, which is now the standing complaint on this entire page. Reaction-norm mutation requires trajectories on multiple tasks per edit and cross-lineage hybridization requires maintaining multiple lineages. Both are strictly more expensive per modification step than clonal mutation. Faster convergence in cycles is not the same as faster convergence in dollars, and the abstract reports the former. This is the same omission the concept page recorded against Ouroboros, Evo-Bench and A²E, and the same one SkillZip (08-12) partially closes on the memory side.

Gaps

  • No token or dollar accounting, so "faster and better convergence over single-trajectory baselines" is measured in the currency that flatters the method.
  • The additive fitness landscape assumption is the one under which recombination provably helps. Real agent code has strong interactions between components; the theory says nothing about the non-additive case, and no empirical check of additivity is reported.
  • Lineage diversity is not characterized. Cross-lineage hybridization needs genuinely different lineages, and nothing reports whether the population stays diverse or collapses, which is the classic failure of evolutionary search.
  • Evaluated on SWE-bench, whose validity was challenged on this wiki one day earlier. Polyglot partially offsets this.

Related