The Pathology: The Score You Optimize Is Lying to You
DGM/SICA operationalize self-improvement as tree search: an agent modifies its own codebase → evaluates → expands the highest-scoring node. 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 SICA/DGM's selection criterion and true lineage value). High-scoring nodes may stagnate quickly (an agent optimizing "score" may have optimized away its own improvability); mediocre nodes may spawn strong descendants. The evolutionary intuition: the standard for judging an ancestor is not how strong it was, but how far its clade ultimately went—the Huxley in the title is Julian Huxley (who coined the clade concept in 1957, cited directly by the paper): a double homage, one to biology, one to the Gödel Machine.
The solution turns this intuition into a computable quantity: CMP (Clade-Metaproductivity) = the aggregated evaluation of a node and all its descendants (successes / total attempts within the clade). Selection no longer asks "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 intensity adapts to remaining budget | | Generation/evaluation | Coupled | Decoupled: agent-task pair-level evaluation, early-stopping failing branches, multi-CPU async | | 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% baseline (+16.7), and is 2.38× faster than DGM, 6.86× faster than DGM 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 engineering agents—that is the "human-level" claim's scope.
Theory and Honesty: The Simulation Theorem Rests on Three Strong Assumptions
Theorem 1, stated precisely: under Assumption 1 (① process quality equals the final agent's evaluation score; ② evaluations are repeatable—environment resets, historical evaluations don't affect future ones; ③ budget is finite and known), 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-improving machine"—the theoretical justification holds, but the boundary is equally clear: the original Gödel Machine assumes the agent has only one life (non-repeatable evaluations) and that proving self-improvement itself consumes real time and resources—precisely what Assumption 1 excludes. The paper itself admits this is "a setting tailored to" a specific problem—not a proof of general self-improvement. The final disclaimer that this cannot be extrapolated into an AGI proof is consistent and honest.
Independent Analysis
First, the fourth sample in the "loops > X" lineage: criterion > X. The trilogy argued architectural loops carry the gains (loops > tools 88/12, loops > feedback types 85%, conditioning > code 72/15). HGM adds the selection criterion: same tree, same compute, swap "current score" for "lineage potential" and the correlation goes 0.28 → 0.78, with 3+ extra points at the same budget. This adds a fourth item: criterion > X—which compass you use matters more than how hard you try. It is orthogonally complementary to the Metan piece's "meta-depth 2.5 ceiling": the ceiling limits recursion depth (how many layers you can modify), while MPM corrects the improvement direction (where to modify). Two bottlenecks of self-improvement, 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) answers "what substrate to modify." HGM answers "which self-modification is worth continuing"—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 layer 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-level evaluation, early-stopping of failing branches, and TS-adaptive sampling. This is structurally identical to ARS's risk-stratified sampling (100% high-impact + 10% sentinel): when verification bandwidth is scarce, the allocation strategy itself is the method. Self-improvement search extends "verification budget economics" from paper pipelines to compute pipelines.
Fourth, Schmidhuber's three decades and the return of evolutionary thinking. 1987 self-referential programs, 2003 Gödel Machine, 2025 engineering approximation—the same person. Meanwhile, clade (1957) re-enters an AI paper title. This confirms a pattern: LLM-era "new problems" (choosing self-improvement directions) often have sixty-year-old theoretical prototypes; only compute and evaluable environments were missing. A lesson worth remembering: once code and evaluation become cheap, the value functions of old theories suddenly become computable.
Things to Watch
1. Whether the ICLR camera-ready adds the series dialogues (the source of the four names including AEVO); 2. Whether CMP degrades under more realistic self-improvement settings (non-repeatable evaluations, unknown budgets)—Assumption 1 is the foundation of all its theory; 3. Overfitting risk on SWE-bench-type benchmarks: CMP aggregates scores from the same distribution—"a promising lineage" may just mean "a 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 official X announcement (ICLR 2026 Oral); code at metauto-ai/HGM. 8 of 9 key points in the video script verified verbatim; 1 (the series name) unverifiable in the preprint and flagged as such. Please credit the source when republishing.