The Pathology: Your Optimization Metric Is Lying to You
DGM/SICA operationalize self-improvement as tree search: an agent modifies its own codebase → evaluation → expand high-scoring nodes. The default assumption: high current benchmark score = high future self-improvement potential. HGM's first contribution falsifies this — the Metaproductivity-Performance Mismatch: the two correlate only weakly (0.28–0.44 between DGM/SICA's selection criterion and true lineage value). High-scoring nodes may stagnate quickly; ordinary nodes may spawn strong descendants. The evolutionary intuition: judge an ancestor not by how strong it is, but by how far its clade ultimately goes — the "Huxley" in the title is Julian Huxley (who introduced the clade concept in 1957, cited directly in the Nature-source paper): a double homage, to biology and to the Gödel Machine.
The fix is making this computable: CMP (Clade-Metaproductivity) = aggregate evaluation over a node and all its descendants (successes / total attempts within the clade). Selection asks not "who is strongest now" but "whose lineage is most promising."
| | DGM/SICA | HGM | |---|---|---| | Selection criterion | Current benchmark score | Clade-aggregated CMP estimate | | Uncertainty handling | None (greedy) | Beta posterior + Thompson Sampling | | Budget awareness | Fixed | Exploration scaled to remaining budget | | Generation/evaluation | Coupled | Decoupled: agent-task-pair-level evaluation, early-stopping of failing branches, async multi-CPU | | SWE-Verified-60 | 53.3% (DGM, 1231h) / SICA loops forever | 56.7% (517h) | | Polyglot | 27.1% (DGM, 2385h) | 30.5% (347h) |
The 56.7% starts from a 40% base (+16.7), and is 2.38× faster than DGM on SWE-bench and 6.86× faster on Polyglot. The heaviest footnote: an agent optimized on GPT-5-mini, evaluated with GPT-5 on SWE-bench Lite, matches the best officially verified results of human-engineered agents — "human-level" by that measure.
Theory and Honesty: The Simulation Theorem Rests on Three Strong Assumptions
Theorem 1, precisely stated: under Assumption 1 (① process quality = final agent's evaluation score; ② evaluations are repeatable — environment resets, past evaluations don't affect future ones; ③ finite, known budget), access to the true CMP suffices to simulate the Gödel Machine's behavior. So HGM (using estimated CMP) is an "approximation of the optimal self-improver" — the theoretical justification holds, but the boundary is equally clear: the original Gödel Machine assumes the agent has only one life (non-repeatable evaluation) and that proving self-improvement consumes real time and resources — exactly what Assumption 1 excludes. The paper itself admits this is "a setting tailored to" the specific problem — not a proof of general self-improvement. The final disclaimer ("cannot be extrapolated into an AGI proof") is consistent and honest.
Independent Analysis
First: a fourth sample in the "loop > X" lineage: criterion > X. The trilogy argued architectural loops carry the gains; HGM adds the selection criterion: same tree, same compute, swapping "current score" for "lineage potential" moves correlation from 0.28 to 0.78 and adds 3+ points at equal budget. That's the fourth entry: criterion > X — what you use as your compass matters more than how hard you try. Orthogonal to Metan's "meta-depth 2.5 ceiling": the ceiling limits recursion depth (how many layers you can modify), while MPM corrects direction (what to modify). Two bottlenecks of self-improving systems, one vertical, one horizontal.
Second: the sixth layer of the self-improvement stack: the direction-sense layer. The five-layer stack (Ornith weights → CoE experience → Metan code → Cordis harness runtime → Mobius knowledge storage) all answer "what substrate to modify"; HGM answers "which self-modification deserves to continue" — the meta-question of exploration. CMP is essentially a value function for the self-improvement tree, like AlphaGo's value network preceding blind expansion: evaluation capability precedes expansion capability. Biology named this long ago (global view of the fitness landscape vs local hill-climbing); HGM engineers it into a bandit.
Third: bandit-style allocation of verification bandwidth. 800 evaluations is a hard budget — HGM spends each one where information is highest via agent-task-pair evaluation, early-stopping, and adaptive sampling. This is isomorphic to ARS's risk-stratified sampling (100% high-impact + 10% sentinels): when verification bandwidth is scarce, the allocation policy is itself the method.
Fourth: Schmidhuber's thirty years and the return of evolutionary thinking. 1987 self-referential programs, 2003 Gödel Machine, 2025 engineering approximation — same person. Meanwhile the clade (1957) re-enters AI paper titles. The pattern: LLM-era "new problems" often have sixty-year-old theoretical prototypes; what was missing was compute and an evaluable environment. Lesson worth keeping: when code and evaluation become cheap, old theories' value functions suddenly become computable.
Things to watch: ① whether the ICLR camera-ready adds the series references; ② whether CMP degrades under realistic self-improvement settings (non-repeatable evaluation, unknown budget) — Assumption 1 is the entire theoretical foundation; ③ overfitting risk on SWE-bench-style benchmarks: CMP aggregates scores from the same distribution — "promising lineage" may just mean "lineage better at gaming this leaderboard."
--- Sources: arXiv:2510.21614 v3 (Wenyi Wang et al., 8 authors, KAUST/AI Plan, including Schmidhuber and DGM author Zhuge; full PDF read, all numbers checked against original tables); OpenReview T0EiEuhOOL; Schmidhuber's X announcement (ICLR 2026 Oral); code at metauto-ai/HGM.