PaperlayerView exact revision on arXiv

Source-linked AI summary

OptiCom : A Unified Framework for State-Conditioned Composition in LLM-Driven Optimization

Chenxing Wei, Sichen Liu, Lizhao Liu, Ningyuan Sun, Chen Bingzhou, Ying He, Bo Jiang, Fei Yu, Yao Shu

arXiv:2609.37221v2cs.AI

TL;DR

Optimization mechanisms can have different utility in different states, complicating their coordination as candidate conditions change. OptiCom unifies optimizer components and adapts their composition; across 32 benchmark groups, it achieved the lowest average rank among 14 evaluated configurations (1.72 for five-run Max).

  • Problem

    Mechanism utility varies across optimization states, making coordination of acquisition, revision, modification, evaluation, and memory decisions a central challenge.

    Source: §1 · p. 2

  • Method

    OptiCom represents optimizers as configurations and uses a state-conditioned controller to compose Action Packages, alongside a slower adapter that refines selection preferences.

    Source: §1 · p. 2

  • Results

    Across 32 benchmark groups, OptiCom achieved the lowest average rank among 14 evaluated configurations (1.72 for five-run Max).

    Source: §6.2 · p. 7

  • Takeaways & Limitations

    Evaluations and targeted ablations support OptiCom’s effectiveness under the tested iteration limits.

    Source: §7 · p. 10

    The selection analysis is conditional on explicit model-capability assumptions at each reached history.

    Source: §4.2 · p. 5

Abstract

from arXiv · show

Large language models (LLMs) are increasingly deployed to solve complex scientific and practical problems via iterative optimization. However, dynamically coordinating diverse search mechanisms as candidate quality, failure modes, and resource budgets evolve remains a critical open challenge. Targeted empirical diagnostics reveal that mechanism effectiveness is highly state-dependent. Motivated by this, we analyze how individual decisions drive final outcomes, decomposing the expected terminal improvement under a shared budget into cumulative decision opportunities minus cumulative selection losses. Guided by this opportunity-loss theoretical foundation, we propose OptiCom, a unified framework that represents LLM-driven optimizers within a shared configuration space: C=(A,Q,O,E,M,S), corresponding to artifact, query, operator, evaluation, memory, and strategy. Operating within this space, a fast LLM-based Optimization Controller dynamically composes immediate mechanisms through structured Action Packages, while a slower Strategy Adapter refines long-term selection preferences, operator weights, and templates based on accumulated trajectory feedback. Comprehensive evaluations across 32 benchmark groups demonstrate the superiority of framework: OptiCom achieves an average Max-score rank of 1.72 among 14 evaluated configurations, securing the top score in 23 groups. Ultimately, these results highlight the broad applicability and high extensibility of OptiCom as a general-purpose paradigm for robust LLM test-time scaling.

1 INTRODUCTION

LLM-driven optimization must search for better candidates under constraints and limited resources, while its useful mechanisms shift with the optimization state. The paper proposes a shared representation and state-conditioned framework, guided by an analysis of terminal improvement as opportunities minus selection losses.

1 INTRODUCTION

  • Optimization mechanisms serve different roles as states change, from repairing invalid candidates to exploring stagnant solutions or making precise updates to high-quality candidates.Existing approaches illustrate distinct ways to organize proposals, feedback, and experience into improvement loops, including OPRO, TextGrad, FunSearch, AlphaEvolve, Reflexion, and ExpeL.Source: §1 · p. 2
  • Expected terminal improvement under a shared budget is decomposed into cumulative decision opportunities minus cumulative selection losses, motivating both valuable alternatives and improved selection.The paper states a sufficient condition for positive terminal improvement under a near-optimal selection assumption.Source: §1 · p. 2
  • Among 14 configurations, OptiCom achieves average ranks of 1.72 for five-run Max and 2.08 for five-run Mean across 32 benchmark groups.The introduction also reports a lowest-ranked overall configuration, but does not provide the group-level win count here.Source: §1 · p. 2
  • OptiCom represents optimizers with six functional dimensions and adapts their component compositions through state-conditioned Action Packages and two-timescale strategy adaptation.The framework instantiates the theory with a task-specific execution harness.Source: §1 · p. 2

2 A SHARED CONFIGURATION SPACE FOR LLM-DRIVEN OPTIMIZERS

The AQOEMS configuration space represents optimizer responsibilities and coordinating rules as six functions, not necessarily separate modules or sequential steps. It also provides a lens for relating existing approaches and motivates OptiCom's dynamically reshaped mechanism composition.

2.1 OPTIMIZATION CONCEPTS AND THE AQOEMS REPRESENTATION

  • AQOEMS represents an optimizer's artifact, query, operator, evaluation, memory, and strategy functions, whose implementations may be shared, omitted, or state-adaptive.Even a fixed-looking configuration can encode adaptive rules in its strategy.Source: §2.1 · p. 3

2.2 EXISTING OPTIMIZERS THROUGH THE AQOEMS LENS

  • AQOEMS frames existing methods as candidate–feedback loops spanning feedback-guided revision, evolutionary search, and shared abstractions or meta-optimization.The section gives representative mappings across these overlapping directions.Source: §2.2 · pp. 3–4
  • Unlike static pipelines or monolithic optimization APIs, OptiCom dynamically reshapes mechanism composition during execution and exposes extensible interfaces for new mechanisms.The described interfaces cover candidate, feedback, memory, and action.Source: §2.2 · p. 4

3 MOTIVATING STATE-CONDITIONED OPTIMIZATION

Heilbronn Triangle diagnostics show that mechanism effectiveness varies across optimization states. The observed comparisons motivate state-conditioned selection and complementary composition, but do not establish universal preferences or a general interaction effect.

3 MOTIVATING STATE-CONDITIONED OPTIMIZATION

  • In s1 and s2, environment probing succeeds at 80.0% and 76.7%, versus 50.0% and 46.7% for self-only querying, while its s4 rate is lower, 26.7% versus 33.3%.These are state-specific descriptive comparisons.Source: §3 · p. 4
  • For s1 and s2, repair succeeds at 73.3% and 76.7%, but under stagnation it reaches 20.0% while recombination reaches 63.3%.The reported success measure is resolution of the starting condition.Source: §3 · p. 4
  • In s2 situations, joint query-plus-operator modification succeeds in 83.3% of trials, versus 76.7% for either single change and 46.7% for the reference.Among successful trials, the joint change averages 2.1 iterations, versus 2.9, 3.6, and 5.2 for query-only, operator-only, and reference, respectively.Source: §3 · p. 5

4 THEORETICAL ANALYSIS OF ADAPTIVE OPTIMIZATION

The analysis characterizes terminal improvement for general optimization policies under a shared budget, then bounds selection losses under a conditional selection assumption. It motivates exposing valuable alternatives while using feedback and experience to improve their selection.

4.1 CUMULATIVE OPPORTUNITY–LOSS DECOMPOSITION

  • The budgeted-process identity expresses expected terminal improvement over a reference policy as the expected sum of per-decision opportunities minus selection losses.The decomposition characterizes a complete run without assuming individual decisions improve the incumbent.Source: §4.1 · p. 5, §4 · p. 5
  • Expanding the available actions cannot lower the best one-decision value when existing action values and costs are preserved, but unchanged selection leaves the net opportunity-minus-loss unchanged.The implication is that catalog expansion should be paired with selection that exploits added choices.Source: §4.1 · p. 5

4.2 SELECTION QUALITY AND GLOBAL IMPROVEMENT

  • Under the conditional selection assumption and bounded terminal utility, the paper bounds selection loss and derives a lower bound on terminal improvement.The assumption limits the probability that selected actions fall more than ϵt below the best available action to δt.Source: §4.2 · p. 5
  • The design principles are to expose valuable alternatives and improve selection using diagnostic feedback and accumulated experience, while allocating decision effort by state and accounting for costs.The analysis assumes a general optimization policy, not specifically OptiCom.Source: §4 · p. 5, §4.2 · p. 6

5 THE OPTICOM FRAMEWORK

OptiCom composes optimization mechanisms in a state-conditioned loop, with a fast Controller choosing immediate actions and a slower Strategy Adapter revising guidance from accumulated experience.

Composing optimization mechanisms

  • The continuous, state-conditioned loop uses immediate mechanism composition and slower guidance revision to expand action opportunities while mitigating selection losses.Both components operate through LLM API calls with fixed model parameters.Source: Figure 2 · p. 6

State-conditioned execution and evaluation

  • The Optimization Controller selects and composes query, operator, evaluation, and memory mechanisms through a structured Action Package.At each iteration it observes the incumbent, alternatives, recent feedback, progress history, and remaining budget; the task-specific harness returns benchmark checks and measurements.Source: §5 · p. 6

Adapting strategy from accumulated experience

  • The Strategy Adapter reviews accumulated successes, failures, and progress, then revises mechanism preferences and prompt templates for later Controller decisions.These updates change guidance without modifying model weights or resetting the archive.Source: §5 · p. 7

6 EXPERIMENTS

Across 32 benchmark groups, OptiCom achieves the best average ranks among 14 evaluated configurations. Experiments also examine component contributions, cross-backbone behavior, adaptation frequency, and token use during continued search.

6.1 EXPERIMENTAL SETUP

  • The evaluation spans 32 benchmark groups and compares OptiCom with 13 method-inspired profiles, using a 30-iteration cap across five runs.The common iteration cap does not equalize candidate counts, API calls, tokens, or elapsed time.Source: §6.1 · p. 7

6.2 MAIN RESULTS

  • OptiCom ranks first across 32 groups among 14 configurations, averaging 1.72 for five-run Max and 2.08 for Mean, versus EvoX at 6.33 and 6.35.It attains the highest Max score on 23 groups, including two ties.Source: §6.2 · p. 7, Figure 3 · p. 8

6.3 ABLATION STUDIES

  • Removing the Strategy Adapter lowered mean scores by 0.240 on Heilbronn Triangle and 0.050 on HotpotQA.Full OptiCom had the highest observed mean and lowest standard deviation among the core variants.Source: §6.3 · p. 8, Table 3 · p. 8
  • Broader action selection improved means by 0.016 and 0.023 over Operator-only adaptation on Heilbronn Triangle and HotpotQA; removing memory lowered them by 0.060 and 0.018.These comparisons support benefits from broader action selection and experience reuse in the tested settings.Source: §6.3 · p. 8, Table 3 · p. 8
  • Event-triggered adaptation reached a mean of 0.953 with 10 Adapter calls, versus 0.954 with 29 calls for every-iteration adaptation on Heilbronn Triangle.Adapter calls fell 65.5%, while total API tokens decreased only from 877K to 865K, approximately 1.37%; adaptive means across three initial compositions ranged from 0.951 to 0.956, versus 0.813 to 0.857 for fixed compositions.Source: §6.3 · p. 8
  • Across evaluated backbones, enabling the Strategy Adapter improved mean scores and compressed variance; GPT-5.5 rose from 0.796 to 0.987 on Heilbronn Triangle with standard deviation ±0.002.For Doubao, the reported Heilbronn gain was +0.240.Source: §6.3 · pp. 8–9, Table 4 · p. 9

6.4 TOKEN-AWARE OPTIMIZATION CASE STUDY

  • On Heilbronn Triangle, the score rose from 0.774 at iteration 2 to 0.9608 at iteration 13 after a six-iteration plateau.The trajectory illustrates that early stagnation can precede later improvement.Source: §6.4 · p. 9, Figure 4 · p. 9
  • After iteration 13, the score remained flat as token use rose from roughly 293K to 865K, with exploration costs reaching about 50K tokens per iteration.The paper distinguishes the value of exploration for breaking early plateaus from its marginal cost near the performance frontier.Source: §6.4 · pp. 9–10, Figure 4 · p. 9

7 CONCLUSIONS AND LIMITATIONS

OptiCom formalizes a shared, state-conditioned framework for coordinating optimization mechanisms, rather than introducing new individual heuristics. The appendices clarify how its functional representation relates to existing systems and where mappings or theoretical claims should not be overread.

7 CONCLUSIONS AND LIMITATIONS

  • OptiCom’s evaluations and targeted ablations support its effectiveness under tested iteration limits, while task-specific evaluation limits and incomplete run-level records constrain the conclusions.The token-aware case study also highlights the cost of continued search without further improvement.Source: §7 · p. 10

A.1 EXPANDED RELATED FRAMEWORKS

  • Unlike fixed-pipeline LLM4AD, benchmark-platform SkyDiscover, and monolithic universal optimization APIs, OptiCom dynamically composes mechanisms according to optimization state.The cited comparisons characterize LLM4AD as domain-centric, SkyDiscover as an evaluation environment, and universal APIs as often relying on a single search strategy.Source: §A.1 · p. 15

A.2 SCOPE OF THE CONTRIBUTION

  • The contribution is an executable state-conditioned coordination architecture, not new individual heuristics; its opportunity-loss decomposition specializes existing performance-difference identities rather than claiming a new theorem.The framework’s interfaces can register new mechanisms, though highly idiosyncratic legacy optimizers may require minor structural alignment.Source: §A.2 · p. 15

A.3 AQOEMS RESPONSIBILITIES AND MAPPING CRITERIA

  • AQOEMS separates optimizer responsibilities without requiring six modules or a fixed execution order, and mechanisms can vary with candidate state, failures, experience, and budget.Functions may share implementations or be absent; a fixed configuration can still contain conditional or adaptive rules.Source: §A.3 · pp. 15–17

A.4 COMPLETE AQOEMS MAPPINGS

  • Functional mappings describe published mechanisms but do not establish that experimental profiles reproduce original implementations; the registry distinguishes evaluated baseline profiles from conceptual references.The table’s B/C labels distinguish those categories, not implementation equivalence.Source: §A.4 · p. 17, Table 6 · p. 17
  • Task-adapted configurations may change artifact or evaluator while retaining selected mechanisms, and nested AQOEMS representations distinguish outer-level optimizer search from inner optimization.For example, the mapping of metaTextGrad is at the outer level, where optimizers are artifacts; its inner process can be represented separately.Source: §A.4 · p. 19

A.5 IMPLEMENTATION DIFFERENCES AND ALIGNMENT CHECKS

The appendix distinguishes preserving an optimizer’s defining decision rules and information flow from reproducing every implementation detail. It also specifies the Heilbronn diagnostic setup and the distinct starting situations used to assess state resolution.

A.5 IMPLEMENTATION DIFFERENCES AND ALIGNMENT CHECKS

  • Optimizer alignment concerns preserving defining mechanisms and their information dependencies, not identical actions; comparisons therefore apply to evaluated instantiations, not every implementation of each method.The checks do not certify equivalence to official implementations, and benchmark-pipeline alignment alone does not establish optimizer equivalence.Source: §A.5 · pp. 19–20, Table 7 · p. 20

B.1 TASK, STATES, AND REFERENCE CONFIGURATION

  • The Heilbronn study evaluates programs returning 11 points in a triangle, with a reference configuration and one-mechanism-at-a-time interventions across five state families.Each mechanism–situation pair receives ten trials under a maximum ten-iteration horizon; displayed state rates aggregate three situations.Source: §B.1 · pp. 20–21

B.2 STATE DEFINITIONS AND SUCCESS CRITERIA

  • The diagnostic study defines state-resolution criteria for its Heilbronn optimization states.Table 8 summarizes the criteria.Source: Table 8 · p. 21

B.3 THREE SITUATIONS PER STATE

  • The 15 starting situations distinguish failure modes, including local-perturbation and structure-restricted plateaus, expensive construction, and initialization sensitivity.The table caption specifies three situations per state and ten trials per mechanism and situation.Source: Table 9 · p. 21, §B.3 · p. 22

B.4 MECHANISMS, MEMORY GATING, AND COMPLETE SITUATION-LEVEL RESULTS

The appendix reports state- and situation-level mechanism comparisons, including a conditional memory comparison and a joint query/operator intervention. It also specifies budgeted-process assumptions that distinguish terminal utility from optimization-time feedback and account for decision costs.

B.4 MECHANISMS, MEMORY GATING, AND COMPLETE SITUATION-LEVEL RESULTS

  • The memory comparison is conditional: episodic, structured, and adaptive memory are paired with Q4 lookup, unlike the strict one-coordinate intervention; the figure averages three situations per state with ten trials each.The Q4 row itself is evaluated under M1, and the figure’s memory panel uses the gate-open comparison.Source: §B.4 · p. 23, Figure 5 · p. 23
  • In s4-2, bandit adaptation reached 9/10 versus 3/10 for the fixed reference, while staged evaluation reached 8/10; in s3-2, meta-evolution and recombination each reached 6/10 versus 2/10 for repair.In the s1-2 construction-timeout situation, environment probing and repair reached 7/10 and 6/10, respectively.Source: §B.4 · p. 23

C.1 BUDGETED PROCESS AND NOTATION

  • The budgeted process uses a common transition and accounts for decision, query, generation, evaluation, memory, and strategy costs; its terminal utility is task-specific and bounded in [0, 1].The stated numerical bound is not asserted for every raw experimental metric.Source: §C.1 · pp. 24–25

B.5 JOINT QUERY AND OPERATOR COMPOSITION

  • 83.3% success for Q+O exceeded 76.7% for either single change and 46.7% for the reference; successful-trial mean iterations were 2.1, 2.9, 3.6, and 5.2, respectively.The joint intervention’s observed gain was approximately 6.7 percentage points over either single change, but the study does not establish general superadditivity.Source: §B.5 · p. 24, Figure 6 · p. 25

C.2 PROOF OF THE CUMULATIVE OPPORTUNITY–LOSS DECOMPOSITION

The proof expresses expected terminal improvement as accumulated decision opportunities minus selection losses, then derives a sufficient improvement bound under a conditional selection assumption. The accompanying remarks clarify that the guarantee concerns expected quality at a shared budget, not every decision or realized trajectory.

C.2 PROOF OF THE CUMULATIVE OPPORTUNITY–LOSS DECOMPOSITION

  • The proof telescopes adjacent hybrid policies, replacing one reference-policy decision at a time while accounting for changes through continuation values.Stopping must count as a decision because it can forgo a reference continuation value.Source: §C.2 · pp. 26–27
  • The finite-horizon identity equates policy improvement with the accumulated per-decision opportunity minus selection loss.It also equals the expected terminal reference value under the policy minus its initial reference value.Source: §C.2 · pp. 26–27

C.3 SELECTION-LOSS BOUND AND GLOBAL IMPROVEMENT

  • The selection-loss bound gives a sufficient condition for improvement, but its conditional selection premise is not established for every adaptive policy.The exact decomposition also allows negative local advantages and does not guarantee monotonic separation of realized trajectories.Source: §C.3 · p. 28

C.4 THEORETICAL REMARKS AND APPLICABILITY

  • At a fixed history, adding actions without changing existing actions or costs cannot lower the best feasible value, but unchanged selection leaves the opportunity-loss difference unchanged.Component diversity therefore needs selection that exploits added choices; unfavorable decisions may be offset by positive contributions elsewhere.Source: §C.4 · p. 28, §C.3 · p. 28
  • When selection failures remain frequent, refining already near-optimal choices alone leaves the loss bound close to the failure probability.Diagnostics, accumulated experience, and state-dependent effort are motivated, but extra querying or reflection consumes budget and changes action values.Source: §C.4 · p. 28
  • The theory is a design guide rather than a runtime estimator: the Controller and Adapter do not estimate its analytical quantities, and empirical scores are unnormalized.The analysis assumes normalized bounded utility.Source: §C.4 · p. 29

D FRAMEWORK SPECIFICATION

OptiCom coordinates registered optimization mechanisms through state-conditioned action choices and slower strategy updates, while preserving task contracts and resource limits. Its evaluation reports the strongest average ranks across 32 groups, using shared task criteria and a common iteration limit.

D.1 OPTICOM ALGORITHM AND EXECUTION ORDER

  • Each loop summarizes the current state, selects a feasible composition, executes it within the remaining budget, updates candidates and memory, and may revise strategy guidance.The algorithm returns the best valid candidate, or the task-defined failure outcome when none is valid.Source: §D.1 · p. 29, D FRAMEWORK SPECIFICATION · p. 29

D.2 COMPONENT INTERFACES AND ACTION PACKAGES

  • The AQOEMS dimensions are functional responsibilities, not isolated modules; components can share implementations or be omitted, and strategy coordinates resource allocation.The task contract and registry define permitted interfaces, inputs, outputs, and execution limits.Source: §D.2 · pp. 29–30, Table 13 · p. 30

D.3 CONTROLLER CONTEXT AND ACTION VALIDATION

  • The Controller uses task, candidate, diagnostic, progress, resource, experience, registry, and strategy context to choose a feasible Action Package.Runtime validation checks package structure, registered components, preconditions, and updated-budget limits; bounded corrections are charged.Source: §D.3 · p. 30

D.4 STRATEGY ADAPTATION RULES

  • The Strategy Adapter reviews accumulated outcomes and may revise preferences, operator weights, and templates when configured history-based triggers are met.Updates are not evidence of improvement by themselves, and the architecture does not assume selection error or adaptation cost decreases monotonically.Source: §D.4 · p. 31

D.5 ARCHIVE, MEMORY, AND CONTEXT CONSTRUCTION

  • Archive and experience policies retain candidate lineage, evaluation status, and scoped lessons, while query policy determines which records enter later contexts.Provisional candidates do not displace a verified incumbent on score alone.Source: §D.5 · p. 31

D.6 EXECUTION RELIABILITY AND RESOURCE ACCOUNTING

  • Execution records distinguish failure types and actual resource use, while reported configurations must identify implementation details to support alignment claims.The appendix’s functional mapping criteria are not themselves a completed implementation-equivalence audit.Source: §D.6 · p. 32

E.1 BENCHMARKS AND EVALUATION PIPELINES

  • All compared methods use shared task-specific evaluation criteria, and the main experiments use five runs with a 30-iteration budget per task.A common iteration limit does not imply equal candidate counts, LLM calls, tokens, elapsed time, or cost; ranking uses five-run Max and Mean summaries.Source: §E.1 · p. 32, §E.2 · p. 33, §E.3 · p. 34

E.3 AGGREGATION AND REPORTING

  • 1.72 (Max) and 2.08 (Mean) are OptiCom’s average ranks across 32 groups and 14 configurations, with groups equally weighted.Ranks use within-group score order; ties at retained precision receive the mean of occupied ranks.Source: §E.3 · p. 34

E.4 ABLATION CONFIGURATIONS

Across tested backbones, enabling the Strategy Adapter improved mean scores, while observed variability was lower; trajectory and ablation analyses show that progress, remaining budget, and evaluation conditions matter. The appendix also reports the scope and limits of the benchmark comparisons.

E.4 ABLATION CONFIGURATIONS

  • Core ablations used five runs and a 30-iteration cap; frequency and feedback studies varied Adapter schedules and information access, but the cap did not equalize token, API-call, or monetary budgets.The frequency study compared event-triggered, no, five-iteration, and every-iteration adaptation; the feedback study compared basic and diagnostic feedback with the same evaluator and scoring criterion.Source: §E.4 · pp. 34–35
  • Ablations included fixed, random, no-Adapter, operator-only, and no-experience-memory variants, plus fixed-versus-adaptive comparisons from three distinct initial compositions.The initial compositions were local revision, diagnosis and repair, and exploration and recombination; fixed and adaptive comparisons held initial candidates and evaluation protocol constant.Source: §E.4 · pp. 34–35

F.1 COMPLETE MAX-SCORE RESULTS

  • Table 15 reports 448 retained Max scores across 32 benchmark groups and 14 configurations, sufficient to reproduce ranks at retained precision but not a substitute for run logs or candidate artifacts.Scores should be compared within tasks, not averaged across heterogeneous raw scales.Source: §F.1 · p. 35

F.2 OPTIMIZATION TRAJECTORIES

  • OptiCom’s early advantage on Erdos Min Overlap reached 0.997 at iteration 16, while on Circle Packing Rect it overtook earlier strong baselines to reach 0.998 at iteration 9.On Circle Packing, its 0.968 at iteration 3 was later surpassed by AlphaEvolve, showing early progress need not determine final quality.Source: §F.2 · p. 36
  • After several near-stagnant iterations, OptiCom improved Heilbronn Triangle to 0.961 at iteration 13 and Matmul to 0.842 at iteration 11, matching the strongest final Matmul score shown.The score histories do not identify which mechanism produced each improvement.Source: §F.2 · p. 36
  • Refinement horizons varied: final improvements on the two Minimizing Max Min Dist tasks occurred at iterations 26 and 13, while Sums Diffs Finite Sets improved incrementally through iteration 20.A late final improvement may follow much earlier attainment of useful quality, so iteration of last improvement alone does not measure efficiency.Source: §F.2 · p. 36
  • OptiCom did not lead every task: EvoX exceeded it on both autocorrelation tasks, and several methods scored higher on Hexagon Packing 12, where OptiCom reached approximately 0.764.The trajectories cannot distinguish insufficient alternatives from inaccurate selection or inadequate feedback.Source: §F.2 · p. 36

F.3 CROSS-BACKBONE ROBUSTNESS

  • Adapter gains were positive in every tested backbone–task pair, ranging from 23.1% to 33.7% on Heilbronn Triangle and 11.1% to 18.1% on HotpotQA, with lower observed standard deviations in all ten comparisons.Gain ordering varied by task, and the five-run summaries do not establish statistical significance or a universal stability guarantee.Source: §F.3 · pp. 39–40

F.4 ADDITIONAL ABLATION DETAILS

Additional ablations compare adaptation schedules, feedback access, and initial compositions. They report similar observed means for event-triggered and every-iteration adaptation, benefits from richer feedback and adaptive optimization, and qualifications on cost and interpretation.

F.4 ADDITIONAL ABLATION DETAILS

  • Event-triggered adaptation scored 0.953 mean versus 0.954 for every-iteration adaptation, with about 65.5% fewer Adapter calls but only about 1.37% lower total token consumption.Every-iteration adaptation had the smallest observed standard deviation, so less frequent adaptation was not uniformly preferable.Source: §F.4 · p. 41, Table 18 · p. 41
  • Rich feedback improved mean scores by 0.056 with the Adapter and 0.034 without it, while both configurations had lower standard deviations with diagnostic feedback.Because downstream decisions also received richer information, this comparison does not isolate immediate Controller selection.Source: §F.4 · p. 41, Table 19 · p. 42
  • Adaptive optimization improved mean scores from all three initial compositions, with absolute gains of 0.141, 0.130, and 0.094 for local revision, diagnosis and repair, and exploration and recombination.It also improved over the strongest fixed configuration, whose mean was 0.857.Source: §F.4 · p. 42, Table 20 · p. 42
  • The mean-score range across initializations fell from 0.044 for fixed compositions to 0.005 for adaptive runs, but only the first iteration was constrained in adaptive variants.This addresses sensitivity to initialization, not recovery from a prolonged unsuitable strategy.Source: §F.4 · p. 42, Table 20 · p. 42

G LIMITATIONS AND BROADER IMPACT

The reported evidence has limits in benchmark coverage, records, computation matching, and evaluation reliability.

G LIMITATIONS AND BROADER IMPACT

  • Evidence is bounded by two official-code comparisons on seven tasks, a one-benchmark motivation study, unavailable per-run records for the overall matrix, unmatched computation, and harness-specific evaluation limits.The experiments do not establish long-horizon scaling, monotonic adaptation, causal effects of individual action changes, or satisfaction of the near-optimal selection assumption.Source: G LIMITATIONS AND BROADER IMPACT · p. 42

Related papers

Loading 2609.37221v2…