Clade-Metaproductivity Metric (CMP)

• CMP is a metric that quantifies long-term self-improvement by assessing the maximum downstream utility among an agent’s descendants.
• It formalizes evaluation through expected utility using probabilistic modeling and Thompson Sampling to balance exploration and exploitation.
• Applied in the Huxley-Gödel Machine paradigm, CMP overcomes immediate performance biases to yield more robust, evolving self-improving agents.

The Clade-Metaproductivity Metric (CMP) quantifies the long-term self-improvement potential of agents in self-modifying coding systems by evaluating the empirical productivity of entire lineages ("clades") rather than focusing solely on immediate performance gains. Originally developed in the context of the Huxley-Gödel Machine (HGM) paradigm, CMP addresses the observed mismatch between single-step benchmark performance and true downstream utility—the so-called Metaproductivity–Performance Mismatch—by aggregating information from descendant agents across a growing tree of self-modifications (Wang et al., 24 Oct 2025).

1. Formal Definition and Mathematical Foundations

CMP formalizes the notion of metaproductivity for a given agent within a self-improvement tree. Let $\mathcal{T}$ denote the tree of agents, each node an agent $a \in \mathcal{T}$. The clade $C(\mathcal{T},a)$ is defined as the subtree rooted at $a$. The self-improvement policy $\pi$ determines the expansion of $\mathcal{T}$, yielding a distribution $p_\pi(\mathcal{T}_B \mid \mathcal{T}, a)$ over possible final trees $\mathcal{T}_B$ upon expanding $a$.

The Clade-Metaproductivity metric under policy $\pi$ is

$a \in \mathcal{T}$0

where $a \in \mathcal{T}$1 is the downstream utility, and $a \in \mathcal{T}$2 the selection score for descendant $a \in \mathcal{T}$3. For most practical policies where $a \in \mathcal{T}$4, this reduces to

$a \in \mathcal{T}$5

Here, CMP characterizes the expected maximum downstream utility among all descendants of $a \in \mathcal{T}$6 in its future clade, robustly capturing multi-step productivity.

2. Theoretical Motivation and Context

A motivating problem for CMP arises in self-improving agent frameworks: agents with maximized immediate benchmark scores frequently give rise to stagnating (“dead”) lineages, while initial low-scoring agents can seed progeny with markedly higher eventual utility. This Metaproductivity–Performance Mismatch challenges policies that naively select for short-term gain (Wang et al., 24 Oct 2025). CMP formalizes guidance that aligns with Huxley's evolutionary principle, which evaluates clades by their most successful descendants rather than progenitors' immediate outputs.

Additionally, the framework draws an analogy with the Gödel Machine paradigm—a theoretical construct for optimal self-improvement. Under repeatable trials and unit cost per modification (Assumption 1), access to a true CMP-oracle suffices to replicate the Gödel Machine's optimal accept/reject process: CMP serves as a Q-value in the corresponding POMDP, quantifying the long-term value of all potential program evolutions within a clade.

3. Estimation and Algorithmic Implementation

Direct computation of $a \in \mathcal{T}$7 is intractable. The HGM framework, therefore, employs an empirical estimation strategy using success/failure counts from task evaluations. Each agent $a \in \mathcal{T}$8 tracks:

$a \in \mathcal{T}$9

Aggregating over the clade $C(\mathcal{T},a)$0 yields: $C(\mathcal{T},a)$1 The resulting point estimate is

$C(\mathcal{T},a)$2

Given the stochasticity and small sample sizes in early phases, HGM applies Thompson Sampling with minimal pseudo-counts and an adaptive scheduling parameter $C(\mathcal{T},a)$3: $a$0 This balances exploration and exploitation while allocating search capacity adaptively across the self-improvement tree.

4. Role in Self-Improvement Strategies

CMP fundamentally alters the search strategy for self-improving code agents. In comparison to prior techniques (SICA, DGM), which either greedily select by immediate score or probabilistically by score and child count, HGM uses Thompson sampling over $C(\mathcal{T},a)$4, prioritizing expansion from lineages with empirically highest clade productivity. The table below summarizes subpolicy distinctions:

Subpolicy	SICA	DGM	HGM (Ours)
Selection	Alternate modify/evaluate	Alternate modify/evaluate	Adaptive (expand/evaluate schedule)
Expansion parent	Greedy (immediate score)	Prob. by score/child count	Thompson sample by $C(\mathcal{T},a)$5
Evaluation choice	Full set (new child)	Progressive subset (new child)	Thompson sample task, all candidates

By shifting emphasis to clade productivity, HGM avoids the myopic trap documented as metaproductivity–performance mismatch, resulting in more robust, non-greedy self-improvement.

5. Empirical Findings and Comparative Evaluation

Empirical results substantiate the utility of CMP for self-improving agents (Wang et al., 24 Oct 2025):

• Correlation with Downstream Utility: On SWE-Verified-60 and Polyglot, the Pearson correlation between empirical CMP and the guidance metric used by each method reveals significant improvements:

Method	SWE-Verified-60 (wtd)	SWE-Verified-60 (unwtd)	Polyglot (wtd)	Polyglot (unwtd)
SICA	0.444	0.444	0.274	0.274
DGM	0.285	0.406	0.383	0.357
HGM	0.778	0.512	0.626	0.873

• Self-Improvement and Efficiency: Under 800-task budgets, HGM achieves higher accuracy and reduced wall-clock CPU time relative to SICA and DGM.

Method	SWE-Ver-60 Acc (%)	CPU-hrs	Polyglot Acc (%)	CPU-hrs
SICA	50.0 (+10.0)	∞	25.4 (+5.1)	572
DGM	53.3 (+13.3)	1231	27.1 (+6.8)	2385
HGM	56.7 (+16.7)	517	30.5 (+10.2)	347

HGM achieves approximately $C(\mathcal{T},a)$6 faster completion on SWE-Ver and $C(\mathcal{T},a)$7 on Polyglot, while generating higher-quality agents.

• Transfer to Human-Level Generalization: On SWE-bench Lite, the HGM-discovered agent, optimized with GPT-5-mini, attains 40.1% (filtered) and 49.0% (standard)—outperforming the initial ancestor and matching or exceeding official human-engineered agent baselines.

6. Assumptions, Limitations, and Extensibility

CMP’s theoretical underpinnings and empirical instantiations in HGM rest on several assumptions:

• Assumptions for Theorem: Repeatable task evaluations, unit budget cost per modification, no evaluation cost for proofs, and utility accrued only at the experiment end.
• Estimation Noise: Because early clades are small, Thompson Sampling with pseudo-counts moderates estimation variance.
• Budget Scheduling: Expansion versus evaluation is regulated by an infinite-armed bandit rule $C(\mathcal{T},a)$8.
• Exploration-Exploitation Tradeoff: Focusing on the best descendant may underweight clades that improve slowly but steadily.
• Computational Overhead: Updating $C(\mathcal{T},a)$9 aggregates for all ancestors upon each evaluation can be costly in principle, but is relatively minor compared to the cost of LLM query cycles.
• Potential for Generalization: CMP could be coupled with richer lineage priors, such as learning curves or graph embedding–based models, to improve estimation fidelity or adapt to more complex sexual recombination schemes.

A plausible implication is that future self-improving agent architectures could utilize variants of CMP to improve generalization across agent populations or domains with hierarchical inheritance and transfer.

7. Significance and Outlook

CMP provides a lineage-wise, expectation-based criterion for long-term agent productivity—addressing the limitations of immediate performance maximization. Results obtained using CMP in HGM demonstrate accelerated, higher-quality self-modifying code agent evolution, with empirical measures rivaling top human systems on widely benchmarked software engineering datasets (Wang et al., 24 Oct 2025). CMP lays foundational groundwork for next-generation self-improving artificial agents and calls for further exploration of lineage-level productivity measures in broader domains.