The Pathology: Your Optimization Score Is Lying to You
DGM/SICA operationalize self-improvement as tree search: an agent modifies its own codebase → evaluates → expands high-scoring nodes. The default assumption: a high current benchmark score = high future self-improvement potential. HGM's first contribution falsifies this — the Metaproductivity-Performance Mismatch: the two are only weakly correlated (the selection criteria of SICA/DGM correlate with true lineage value at just 0.28–0.44). High-scoring nodes may stall quickly (an agent that optimizes the score may lose improvability); mediocre nodes may birth strong descendants. The evolutionary intuition: judge an ancestor not by how strong it is, but by how far its clade ultimately travels — the Huxley in the title is Julian Huxley (who coined the clade concept in 1957, cited directly by the paper): a double tribute, one to biology, one to the Gödel Machine.
The solution makes this computable: CMP (Clade-Metaproductivity) = aggregate 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 has the brightest future."
| | 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 adjusted to remaining budget | | Generation/evaluation | Coupled | Decoupled: agent-task pair evaluation, early-stopping failed 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% base (+16.7), and is 2.38× faster than DGM, 6.86× faster on Polyglot. The heaviest footnote: an agent optimized with GPT-5-mini, evaluated with GPT-5 on SWE-bench Lite, matches the best officially verified results of human engineering agents — "human-level" by that measure.
Theory and Honesty: The Simulation Theorem Rests on Three Strong Assumptions
Theorem 1, stated precisely: under Assumption 1 (① process quality = final agent's evaluation score; ② evaluations are repeatable — environment resets, past evaluations don't affect the future; ③ budget is finite and known), accessing 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 — theoretically justified, but with clear boundaries: the original Gödel Machine assumes the agent has one life (non-repeatable evaluation) and that proving self-improvement costs real time and resources — exactly 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 caveat ("cannot be extrapolated into an AGI proof") is consistent with this. Honest.
Independent Observations
One: a fourth sample in the "loop > X" lineage: criterion > X. The trilogy argued that architectural loops carry the gains (loop > tools 88/12, loop > feedback type 85%, conditioning > code 72/15). HGM adds the selection criterion: same tree, same compute — swapping "current score" for "lineage potential" moves correlation 0.28→0.78 and adds 3+ points at equal budget. This adds a fourth item: criterion > X — what you use as a compass matters more than how hard you try. Orthogonal to the Metan piece's "meta-depth 2.5 ceiling": the ceiling limits recursion depth (how many layers you can change), MPM corrects the direction (what to change). Two bottlenecks of self-improvement, one vertical, one horizontal.
Two: a 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 change." 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 had a name for this long ago (global view of the fitness landscape vs. local hill-climbing); HGM engineers it into a bandit.
Three: bandit-izing the 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 of failed branches, and TS-adaptive sampling. Structurally identical to ARS's risk-stratified sampling (100% high-impact + 10% sentinels): 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.
Four: Schmidhuber's thirty years and the return of evolutionary thinking. 1987 self-referential programs, 2003 Gödel Machine, 2025 engineering approximation — the same person. Meanwhile the clade (1957) reappears in an AI paper's title. This confirms a pattern: LLM-era "new problems" (choosing self-improvement directions) often have sixty-year-old theoretical prototypes; only compute and an evaluable environment were missing. Lesson worth remembering: 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 citations (AEVO and three other names); ② whether CMP degrades under more realistic self-improvement settings (non-repeatable evaluation, unknown budget) — Assumption 1 is the foundation of all its theory; ③ overfitting risk on SWE-bench-style 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: metauto-ai/HGM. Of the 9 key points in the video script, 8 were verified verbatim; 1 (the series name) could not be verified in the preprint and is flagged as such. Reposting requires attribution.