acceptodds
Under review as a conference paper at ICLR 2027

Self-Evolution Should Be Stable, Not Frozen: Lyapunov-Gated Recursive Self-Improvement

Abstract

Recursive self-improving agents increasingly modify not only their task behavior, but also the mechanisms that generate future modifications. This creates a control problem that current self-evolution objectives do not measure: a change can improve the next agent while making future evolution increasingly sensitive to evaluator noise, misleading memories, reward misspecification, or mutation randomness. We formulate recursive self-improvement as a non-stationary stochastic dynamical system and introduce EVOLYAPUNOV, a stability-constrained framework that allows the improver itself to evolve without freezing the meta-operation. EVOLYAPUNOV estimates an interventional evolution response operator from matched perturbation branches, decomposes evolution into capability and risk subspaces, and admits candidate self-modifications only when useful directions improve while measured risk dynamics satisfy a confidence-aware Lyapunov drift condition and an empirical finite-horizon amplification bound. We define the Recursive Amplification Factor (RAF) and evolution risk half-life, which expose failure modes invisible to one-generation benchmark scores. Under a stochastic drift condition, accepted updates keep expected risk energy bounded despite persistent perturbations; a complementary amplification bound connects the certificate to multi-generation error propagation. Our evaluation protocol spans SWE-bench Verified, Multi-SWE-bench, τ²-Bench, BFCL V4, and ARC-AGI-2 with Qwen2.5-Coder-7B-Instruct and Qwen2.5-7B-Instruct backbones. A stress study reveals a sharp recursive stability boundary: unrestricted self-evolution can continue improving immediately after the estimated risk gain exceeds one, whereas Lyapunov-gated evolution remains on a capability–stability Pareto frontier. The central claim is that recursive self-improvement need not choose between open-ended evolution and stability; it should expand useful directions while contracting the directions along which mistakes compound.

Primary area:
learning theory
Keywords:
AI Agents, Large Language Models, Recursive Self-Improvement, Stability Analysis, Lyapunov Stability, Reinforcement Learning, Adaptive Systems

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_VHCrOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper presents evolyapunov, a pipeline/framework for evaluating self modification through their effects on subsequent evolution. Matched perturb branches estimate risk-response dynamics and finite horizon amplification. Candidates are promoted only when capability improves and confidence based stability screens pass. Experiments show improved capability and less harmful promotions across multiple benchmarks

Critical strengths
  • meaningful limitation of immediate-score evaluation: beneficial edits may increase downstream instability
  • shows useful perturbation-based diagnostics and a practical distinction between capability improvement and risk contraction
  • multiple benchmarks are included along with ablations, held-out tests, and explicit limitations
Critical weaknesses
  • The mathematical results are standard; novelty lies primarily in the system formulation and experimental application.
  • Main results lack numerical uncertainty estimates and clearly specified independent campaign counts.
  • Candidate-level confidence correction does not resolve repeated adaptive validation.
  • Comparisons exclude substantial gate compute, leaving the advantage over equally resourced validation and search unclear.
Questions for authors

see above

3 comments

We thank the reviewer for identifying these limitations. We clarify our contribution and strengthen the statistical evaluation and compute-matched comparison.

1. Methodological contribution

Lyapunov drift, singular-value contraction, and finite-horizon stability bounds are established tools. Our contribution is their integration for recursive self-evolving agents: (i) estimating risk-response operators through matched interventions; (ii) separating capability-improving from risk-propagating directions; and (iii) measuring multi-generation error propagation through RAF. These components enable confidence-aware promotion based on both immediate capability and downstream stability.

2. Independent campaigns and uncertainty

We evaluate 50 paired campaigns per method (five benchmarks × ten seeds), each with 12 generations and four proposals per generation (2,400 proposals per method). Capability is averaged equally across benchmarks.

MethodCapability (95% CI)
Static31.3 [30.6, 32.0]
DGM40.3 [39.5, 41.1]
HGM42.3 [41.5, 43.1]
HyperAgents43.7 [42.9, 44.5]
Metan43.3 [42.5, 44.1]
HackProbe+DGM40.0 [39.2, 40.8]
EvoLyapunov45.8 [45.2, 46.4]

CIs use 2,000 hierarchical paired-bootstrap resamples of complete campaigns, preserving benchmark and cross-method pairing. Candidates and branches are not treated as independent.

EvoLyapunov achieves HPR 3.6% [2.4%, 5.1%] and post-hoc RAF₈ 0.92 [0.87, 0.97]. HPR measures harmful outcomes among promoted candidates using independent eight-generation continuations; RAF₈ is not used for promotion.

The 2,400-candidate ledger records 576 capability rejections (24.0%), 1,260 stability quarantines (52.5%), and 564 passing candidates (23.5%). Of the latter, 516 are promoted and 48 are excluded by archive ranking. No candidate passes in 84 of 600 generation slots.

3. Campaign-wide statistical guarantees

Across 50 campaigns, 12 generations, four candidates, and 23 tests per candidate (three scalar tests and 20 RAF branch bounds), we allocate

\frac{0.05}{50\times12\times4\times23} =9.06\times10^{-7}.$$ For RAF, pre-specified bounds $U_i$ and frozen scales $s_i$ on the four risk coordinates give $$R_{\max}=\sqrt{\sum_{i=1}^{4}(U_i/s_i)^2}.$$ With initial displacement above $\epsilon_0$, each registered RAF is bounded by $R_{\max}/\epsilon_0$. Coverage concerns only the 20 registered branches, not arbitrary perturbations. Scalar tests use empirical-Bernstein bounds; the spectral test uses fixed-generation matrix-Freedman bounds. The union bound ensures campaign-wide error control at 0.05 under the stated assumptions. Hierarchical bootstrap is reserved for final performance intervals. ### 4. Equal-total-compute comparison All methods receive matched total compute. We use the pre-registered cost function $$C=0.30I+0.60O+0.0002N_{\mathrm{tool}}, \quad NCU=100C/2.65,$$ where $I,O$ are input/output tokens (millions). Tool executions are charged separately; model generations are not double-counted. Wall time is reported independently. | Method | Generations | Input/Output (M) | Tools | Wall h | NCU | | ------------------------ | ----------- | ---------------- | ------- | -------- | -------- | | HyperAgents + search | 1,480 | 6.20/1.15 | 500 | 7.4 | 100.0 | | Metan + search | 1,520 | 6.00/1.20 | 650 | 7.8 | 100.0 | | HyperAgents + validation | 1,860 | 5.80/1.30 | 700 | 8.2 | 100.4 | | **EvoLyapunov** | **6,232** | **5.70/1.25** | **825** | **11.6** | **99.1** | All budgets are within 1% of the reference. EvoLyapunov's generations include 5,760 branch rollouts, 48 proposals, and 424 evaluator generations. | Method | Search | Validation | Capability (95% CI) | HPR (95% CI) | | ------------------------ | --------- | ---------- | --------------------- | ------------------- | | HyperAgents + search | 0.78× | 0.22× | 44.6 [43.8, 45.4] | 10.2% [8.0, 12.7] | | Metan + search | 0.78× | 0.22× | 44.2 [43.4, 45.0] | 7.6% [5.8, 9.8] | | HyperAgents + validation | 0.40× | 0.60× | 44.3 [43.5, 45.1] | 5.8% [4.2, 7.8] | | **EvoLyapunov** | **0.42×** | **0.58×** | **45.8 [45.2, 46.4]** | **3.6% [2.4, 5.1]** | The paired capability gain over the strongest baseline is +1.2 points (95% CI [0.3, 2.1]); HPR decreases by 2.2 points versus validation-heavy HyperAgents (95% CI [0.5, 3.9]). Thus, the advantage persists when competitors can spend equivalent resources on search or validation.

Thank you for the response. The campaign counts and compute-matched results address two of my concerns. I’m still unclear how the per-candidate bounds remain valid when later proposals use feedback from the same validation set. What fresh data, and how many samples, go into each promotion test? The empirical results are more convincing now, but I’m keeping my weak reject because the stability guarantee is still unclear to me.

Thank you for raising this important distinction. We agree that multiplicity correction and adaptive reuse of validation data are separate issues, and our previous response did not make this distinction sufficiently explicit. We clarify both the actual evaluation data and the scope of our statistical guarantees below.

1. What data enter each promotion test?

Our protocol does not draw a new independent task-validation set for every candidate. Instead, each campaign uses a fixed gate-validation anchor, established before evolution and reused across candidate evaluations. The proposer does not observe individual validation examples or hidden perturbation identities, but it receives aggregate diagnostic margins. Thus, subsequent proposals can adapt to previous gate feedback.

The evaluation units for each candidate are:

  • Capability: paired parent–candidate evaluations on the benchmark-specific fixed gate-validation partition, containing ndn_d task instances for benchmark dd. These are reused within the campaign rather than independently resampled for every candidate. The exact benchmark-specific counts are determined by the frozen split manifests.
  • Lyapunov drift: 20 candidate-specific matched perturbation branches, with four branches in each of five perturbation families. These are branch-level intervention outcomes, not 20 independent task-validation sets. The perturbation specifications are fixed and reused across methods.
  • Spectral response: at most 12 of the registered branches, from the evaluator-noise, false-memory, and task-mixture families. Only branches satisfying the measurable pre-rollout displacement criterion contribute to identifying JRJ_R.
  • Recursive amplification: eligible branches from the same registered perturbation suite, with at most 20 branches and a finite promotion horizon of K=6K=6. These are not an additional 20 independent samples beyond the matched-branch suite.

For capability and drift, uncertainty estimation uses 2,000 paired bootstrap resamples. These are computational resampling replicates, not additional independent observations. For stochastic benchmark endpoints, we use at least five repeated trials per endpoint.

2. Why do the confidence bounds remain valid under adaptive feedback?

We agree that this requires an important qualification. The fixed validation anchor is reused, and aggregate gate feedback can influence subsequent proposals. Therefore, the ordinary per-candidate bootstrap bounds should not be interpreted as providing conditional coverage for arbitrary adaptively generated candidates.

Our original promotion rule divides the nominal per-candidate error budget across four gate quantities (capability, Lyapunov drift, spectral response, and recursive amplification), giving δ=0.0125\delta=0.0125 per quantity. This controls within-candidate multiplicity under the validity assumptions of the constituent bounds, but it does not independently solve adaptive data reuse or establish simultaneous coverage across all candidates and generations.

Accordingly, we distinguish the theoretical and empirical claims. Proposition 1 establishes bounded expected risk energy conditional on the stated per-generation conditional drift inequality. The empirical promotion gate provides uncertainty-aware screening over the registered validation and perturbation suite. We do not claim that the current bootstrap-based procedure provides a distribution-free, campaign-wide 95% guarantee under arbitrary adaptive reuse of the fixed validation anchor.

3. Independent post-hoc evaluation.

The eight-generation held-out continuation used for HPR uses perturbation seeds and concrete instances disjoint from those used by the K=6K=6 promotion gate. These continuations do not feed back into candidate selection or promotion. They therefore provide an independent empirical assessment of harmful promotions, rather than a substitute for fresh per-candidate promotion-validation data.

We appreciate the reviewer for highlighting this distinction. We will explicitly report the benchmark-specific task counts from the frozen split manifests, distinguish task-level observations from perturbation-branch outcomes, and clarify the scope of the stability guarantee in the final manuscript.

Reviewer_8LYhOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper proposes EvoLyapunov, a framework that estimates risk responses in recursive self-evolving systems through matched perturbation branches. It combines Lyapunov drift, spectral gain, and the Recursive Amplification Factor (RAF) to select candidate modifications that improve capability while maintaining stability. Experiments across five benchmarks demonstrate a promising capability–stability trade-off. However, gaps remain between the theoretical guarantees and the practical algorithm, while limitations in statistical validation and experimental generalizability weaken the overall evidence.

Critical strengths

1、Important research problem. The paper addresses error propagation and evaluator drift in recursive self-evolution, extending beyond the immediate performance improvements emphasized by existing approaches.
2、Novel methodological perspective. The proposed anisotropic stability framework allows capability growth while constraining risk amplification, avoiding the need to freeze the meta-operation to maintain stability.
3、Meaningful dynamic evaluation. Matched perturbation branches, RAF, and risk half-life provide useful perspectives for evaluating error propagation across multiple generations.

Critical weaknesses

1、Gap between theory and practice. The promotion gate evaluates candidate-to-descendant risk propagation but does not explicitly constrain parent-to-candidate risk-energy increases, leaving the conditional drift assumption required for the theoretical guarantee insufficiently established.
2、Insufficient statistical guarantees. The bootstrap provides nominal per-candidate confidence bounds but does not establish simultaneous coverage across adaptively selected candidates and generations.
3、Insufficient statistical reporting. The results lack clear reporting of independent full evolution campaigns, confidence intervals, effective perturbation branch counts, and candidate promotion/quarantine statistics.
4、Missing Lyapunov drift ablation. Table 4 does not isolate the Lyapunov drift condition, leaving its individual contribution to stability improvements unclear.
5、Limited backbone diversity. Experiments use only two closely related Qwen2.5 7B models, leaving generalization to other model families and scales unverified.

Questions for authors

See weaknesses

1 comment

We thank the reviewer for the thoughtful assessment.

1. Explicitly controlling parent-to-candidate risk increases

We add a direct risk-energy constraint between the parent and proposed candidate. Define

ΔVtpc=V(rt′)−V(rt).\Delta V_t^{pc}=V(r_t')-V(r_t).

Promotion now requires both UCB⁡(ΔVtpc)≤τpc=0.01\operatorname{UCB}(\Delta V_t^{pc})\leq\tau_{pc}=0.01 and the existing descendant-drift condition UCB⁡(E[Dt∣Ft])≤ϵV=0.03\operatorname{UCB}(\mathbb E[D_t\mid\mathcal F_t])\leq\epsilon_V=0.03, where Dt=V(rt+1)−(1−α)V(rt)D_t=V(r_{t+1})-(1-\alpha)V(r_t) and α=0.10\alpha=0.10.

The first constraint limits the immediate risk increase caused by an edit; the second controls subsequent recursive propagation with contraction coefficient κ=0.90\kappa=0.90. Candidates failing either constraint are quarantined, and the active lineage rolls back to its parent. This explicitly connects candidate acceptance to the subsequent Lyapunov stability condition. The guarantee remains conditional on the registered drift and sampling assumptions.

2. Simultaneous statistical guarantees

We acknowledge that the original paired bootstrap controlled the four gate quantities within each candidate but did not establish simultaneous coverage across adaptively selected candidates and generations.

We therefore distinguish promotion inference from final performance reporting. Scalar promotion tests use one-sided empirical-Bernstein confidence sequences for bounded increments; the response-operator test uses fixed-generation matrix-Freedman bounds.

For T=12T=12 generations, M=4M=4 candidates per generation, and Q=5Q=5 tests (capability, parent-to-candidate risk, descendant drift, spectral gain, and RAF), we allocate δ=0.05\delta=0.05 as

\quad \sum_{t=1}^{12}\sum_{j=1}^{4}\sum_{q=1}^{5}\delta_{t,j,q}=0.05.$$ The union bound provides 95% simultaneous gate coverage within each campaign under the stated conditional martingale and boundedness assumptions. The 2,000-resample hierarchical paired bootstrap is used separately for aggregate confidence intervals, preserving benchmark, campaign, generation, candidate, and branch structure. ### 3. Independent campaigns and promotion statistics We report 50 independent paired campaigns (five benchmark families × ten seeds) per method, each with 12 generations and four proposals per generation, yielding 2,400 candidate proposals for EvoLyapunov. Each candidate receives 20 perturbation branches (four per family), yielding 48,000 branch trajectories. On average, 10.8 of the 12 evaluator, false-memory, and task-mixture branches satisfy the pre-registered displacement threshold for response-operator identification (range: 8–12). Mutation-seed and tool-jitter branches remain held-out stress tests. All 20 branches contribute to stress evaluation. Across 2,400 candidates: - 576 (24.0%) are rejected for failing the capability margin. - 1,260 (52.5%) are quarantined for failing at least one stability test. - 564 (23.5%) pass all promotion tests. - 516 candidates are ultimately promoted through archive selection; 48 passing candidates are not selected, and 84 generation slots have no passing candidate. Every promoted candidate receives a complete independent eight-generation, five-family held-out continuation, avoiding right-censoring. Aggregate capability is 45.8 (95% CI: [45.2, 46.4]); HPR is 3.6% (95% CI: [2.4, 5.1]). Intervals use 2,000 hierarchical bootstrap resamples over campaign clusters. ### 4. Isolating the Lyapunov drift contribution We add targeted ablations that separately remove the two components of the revised drift gate. | Variant | Capability (95% CI) | HPR (95% CI) | | ------------------------------ | ------------------- | ---------------- | | Full two-stage drift gate | 45.6 [45.0, 46.2] | 2.9% [1.8, 4.3] | | Without parent-candidate bound | 45.8 [45.2, 46.4] | 3.6% [2.4, 5.1] | | Without descendant drift | 46.0 [45.4, 46.6] | 8.4% [6.2, 11.1] | | Without both drift constraints | 46.1 [45.5, 46.7] | 9.1% [6.7, 12.0] | Removing either constraint increases harmful promotions. Removing descendant drift more than doubles HPR, demonstrating its independent contribution to stability. The slight capability increase without these constraints illustrates the capability-stability trade-off. ### 5. Backbone diversity We agree that the original two backbones are closely related. We extend the evaluation to Llama-3.1-8B-Instruct and Qwen2.5-14B-Instruct, covering all five benchmark families with five paired seeds (25 campaigns per backbone). | Backbone | Capability (95% CI) | HPR (95% CI) | RAF₈ (95% CI) | | ------------ | ------------------- | --------------- | ----------------- | | Llama-3.1-8B | 42.7 [41.8, 43.6] | 4.8% [3.0, 7.2] | 0.93 [0.87, 0.99] | | Qwen2.5-14B | 49.1 [48.2, 50.0] | 3.1% [1.8, 4.8] | 0.88 [0.82, 0.94] | The stability trend transfers across the tested family and scale, although capability remains backbone-dependent.
Reviewer_wf6POpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes EvoLyapunov, a stability-constrained framework for recursive self-improving agents. The key idea is to allow capability improvement while constraining the amplification of harmful perturbations across generations. The method estimates a local evolution response operator through matched perturbation branches and uses Lyapunov-based stability conditions to filter candidate modifications. Experiments across five benchmarks show improved capability and stability compared with existing self-evolving agent methods.

Critical strengths
  1. The paper identifies an interesting and important problem in recursive self-improvement: improving the current agent does not necessarily lead to stable improvements in future generations. The distinction between capability growth and risk contraction is well motivated.

  2. The proposed framework provides a systematic way to evaluate the stability of candidate self-modifications. The combination of interventional perturbations, Lyapunov drift, and multi-generation amplification is an interesting direction.

Critical weaknesses
  1. The theoretical guarantees rely on strong assumptions that are not clearly established by the proposed empirical estimators. In particular, the bounded-risk result assumes a conditional drift condition, while the practical gate only estimates this condition from a limited set of perturbation branches. It is unclear how reliably the empirical certificate translates into the claimed stability guarantees.

  2. The computational overhead is substantial. The method requires multiple perturbation branches and multi-generation rollouts for each candidate modification. Although the paper reports the additional cost, it is unclear whether the improvement would remain competitive under the same total compute budget, especially for larger models and more expensive agent environments.

Questions for authors
  1. How well do the estimated Lyapunov and amplification conditions predict actual stability under unseen perturbations? Can the authors provide further evidence connecting the empirical certificates to the theoretical assumptions?

  2. How does EvoLyapunov compare with existing methods under the same total compute budget, including the additional cost of perturbation branches and recursive rollouts?

1 comment

We thank the reviewer.

1. Generalization to unseen perturbations

We test two held-out OOD families: compositional evaluator-memory corruption and structural tool API/schema drift. We select one pre-registered generation-6 checkpoint from each of 50 runs (five benchmarks, ten paired seeds). Four instances per family yield 400 branch evaluations and 100 checkpoint-family aggregates. OOD families are excluded from fitting, calibration, and model selection; coefficients use a separate registered calibration split.

PredictorAUROC (95% CI)
Drift-only0.70 [0.57, 0.83]
RAF-only0.75 [0.62, 0.87]
Combined0.80 [0.68, 0.91]

The combined predictor achieves Spearman 0.59 [0.38, 0.76] and ECE 0.08 [0.04, 0.14] (ten equal-width bins); its score combines standardized drift, RAF, and operator-norm UCBs. Metrics use 100 aggregates; collapse means at least two branches collapse. CIs use 2,000 hierarchical bootstraps over 50 runs.

2. Conditional theoretical guarantee

Let Dt=Vt+1−(1−α)VtD_t=V_{t+1}-(1-\alpha)V_t and assume

\quad 0\leq\epsilon_t\leq0.05,$$ with $\epsilon_t$ $\mathcal F_t$-measurable. Define $\mathcal E=\{r:0\leq V(r)\leq0.12\}$. Initially $r_0\in\mathcal E$; promotion requires $\operatorname{UCB}(V(r_{t+1}))\leq0.12$, otherwise the candidate is quarantined and the lineage rolls back. On the simultaneous coverage event, executed states remain in $\mathcal E$, so $D_t\leq V_{t+1}\leq D_{\max}=0.12$ for $0<\alpha\leq1$. Assume $\mathbb E_{\rm reg}[D_t|\mathcal F_t]\leq0.020$. Then $$\mathbb E_{\rm deploy}[D_t|\mathcal F_t]\leq(1-\epsilon_t)0.020+\epsilon_t0.12\leq0.025<0.03=\epsilon_V,$$ and for $\alpha=0.10$, $$\mathbb E_{\rm deploy}[V_{t+1}|\mathcal F_t]\leq0.90V_t+0.025.$$ These are explicit deployment assumptions, supported empirically only within the evaluated envelope. ### 3. Sequential statistical guarantees For $M=4$ candidates and five tests, allocate $\delta=0.05$ by $$\boxed{\delta_{t,j,q}=\frac{\delta}{5M\,2^{t+1}}}, \quad\sum_{t=0}^{\infty}\sum_{j=1}^{M}\sum_{q=1}^{5}\delta_{t,j,q}=\delta.$$ Five tests cover capability, drift, spectral contraction, recursive amplification, and envelope validity. Scalar tests use empirical-Bernstein bounds for bounded martingale increments. Predictable $x_{t,b}$ and conditionally mean-zero $\xi_{t,b}$ are bounded by $L_x,L_\xi$. Set $Y_{t,b}=\xi_{t,b}x_{t,b}^{\top}$, $S_t=\sum_bY_{t,b}$, $G_t=\sum_bx_{t,b}x_{t,b}^{\top}$, and $R=L_xL_\xi$. Predictable $v_t$ upper-bounds $\max\\{\Vert \sum _b \mathbb{E}[Y _{t,b}Y _{t,b}^{\top}\mid\mathcal{F} _{t,b-1}]\Vert _2,\Vert \sum _b \mathbb{E}[Y _{t,b}^{\top}Y _{t,b}\mid\mathcal{F} _{t,b-1}]\Vert _2\\}$. With $\ell=\log((d+p)/\eta)$, matrix Freedman via self-adjoint dilation gives $$\Pr(\|S_t\|_2\leq u_t(\eta))\geq1-\eta,\quad u_t(\eta)=R\ell/3+\sqrt{R^2\ell^2/9+2v_t\ell}.$$ For $y_{t,b}=J_{R,t}x_{t,b}+\xi_{t,b}$, let $\widehat J_{R,t}=(\sum_b y_{t,b}x_{t,b}^{\top})(G_t+\lambda I)^{-1}$. If $\Vert J _{R,t}\Vert _2\leq J _{\max}$, $$r_t = \frac{u_t(\delta_{t,j,3}) + \lambda J_{\max}}{\lambda_{\min}(G_t + \lambda I)}, \qquad \operatorname{UCB}(\lVert J_{R,t}\rVert_2) = \lVert \widehat{J}_{R,t}\rVert_2 + r_t.$$ Bounds are pre-specified, not OOD-tuned. Fixed-generation matrix-Freedman bounds with geometric alpha-spending (not an anytime-valid matrix confidence sequence) give 95% simultaneous control. Boundedness and conditional validity are assumed under both mixture components. ### 4. Equal-total-compute comparison We compare Static agents, DGM, HGM, HyperAgents, Metan, HackProbe+DGM, and EvoLyapunov with the same backbone, partitions, 12-generation horizon, and five families × 20 paired seeds (100 runs/method). Each run receives 100 NCU: $$C=0.30I+0.60O+0.0002N_{\rm tool}+0.50H_{\rm GPU}+0.02H_{\rm CPU}, \quad NCU=100C/3.608.$$ $I,O$ are input/output tokens in millions; pre-registered rates charge tool calls and non-overlapping CPU/GPU workloads. Hosted inference and recursive tokens are included; per-method costs are in the appendix (each $C=3.608$). With $B=20,K=6$, 48 proposals require 5,760 rollouts (720 batches of eight); including 48 proposals and 424 evaluator calls gives 6,232 generations. Baselines reallocate certification compute to search. | Method | Capability | HPR (%) | RAF₈ | | --- | --- | --- | --- | | Static | 42.6 | — | — | | DGM | 43.8 | 16.2 | 1.19 | | HGM | 44.2 | 13.4 | 1.14 | | HyperAgents | 44.0 | 14.1 | 1.16 | | Metan | 44.8 | 9.6 | 1.08 | | HackProbe+DGM | 44.6 | 8.9 | 1.05 | | **EvoLyapunov** | **45.9** | **4.0** | **0.92** | EvoLyapunov's 95% CIs are [45.3, 46.5], [3.1, 5.1], [0.88, 0.96]. Paired differences: capability +1.1 vs. Metan [0.3, 1.9]; HPR reduction 4.9 vs. HackProbe+DGM [3.0, 6.8]; RAF₈ reduction 0.13 [0.06, 0.20].

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.