acceptodds
Under review as a conference paper at ICLR 2027

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

Abstract

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: , 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.

Primary area:
foundation or frontier models, including LLMs
Keywords:
LLM-driven optimization, Adaptive optimization, Optimization frameworks.

Reviews

Among:
Shape:
1234this paper, Bottom 25%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.75 over 4 reviews, above 25% of the others.
Reviewer_7ARmOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes OptiCom, a unified framework for LLM-driven optimization that groups common search mechanisms into several roles. An LLM Controller picks a combination of mechanisms at each step based on the current state, and a slower Strategy Adapter updates the selection preferences and prompts over time. Experiments on a wide range of optimization tasks show that OptiCom achieves the best overall ranking against many existing methods.

Critical strengths
  1. The paper is well-written and easy to follow.
  2. The framework unifies many existing LLM-driven optimization methods, and the experiments cover a wide range of tasks.
Critical weaknesses
  1. Unfair Comparison. Most baselines are re-implemented by the authors and the compute is not matched across methods, so it is hard to tell where the gains come from.
  2. Unclear Contribution of the Controller. Most of the gain seems to come from the Strategy Adapter, while the per-step Controller that the paper emphasizes adds little on its own.
Questions for authors
  1. How does OptiCom compare with the baselines under the same token budget?
  2. How does the Controller compare with a simple bandit or random selection when the Adapter is kept?
  3. Missing references. https://arxiv.org/abs/2506.15196, https://arxiv.org/abs/2509.19349 and https://arxiv.org/abs/2309.16797 should be discussed.
Reviewer_FzQYOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents OptiCom, a unified framework that breaks LLM optimizers into six parts: artifact, query, operator, evaluation, memory and strategy. In general, a fast Controller chooses structured Action Packages; a slower Adapter updates search guidance from accumulated feedback and a performance-difference decomposition method motivates coordinating available opportunities and selection quality. Empirically, OptiCom gets the best average rank among 14 configurations on 32 benchmarks.

Critical strengths
  1. This paper proposes a unified AQOEMS representation that maps and aligns diverse LLM-driven optimizers to one shared configuration space, making their choices directly comparable.
  2. The evaluation covers 32 benchmark groups and five backbones. Within the system, the Strategy Adapter gives a clear gain over multiple baselines.
Critical weaknesses
  1. The experiments and framework design appear only loosely connected to the theory. The paper also does not empirically validate the proposed theoretical analysis.In addition, the reference policy π0π0 is not specifically mentioned in the experiments.
  2. The compute cost is not controlled, and there is no comparison of OptiCom with baselines in cost level. Also the "iteration" definition across different baselines is unclear, there can be multiple LLM requests within one “iteration” in OptiCom, you should at least report the cost of baselines.
  3. Theorem 1 and 2 can apply to any policy, not only OptiCom, which cannot explain why OptiCom outperforms the baselines. Neither theorem is closely tied to the paper’s specific claims about OptiCom.
Questions for authors
  1. Some of the baselines are modified and executed within your framework, are these baselines equivalent to their original implementations? would they reproduce the original results?
  2. Prompt design of Opticom, wall-clock time, token usage and LLM call counts can also be reported for baselines.
  3. How are untried actions distinguished from unsuccessful actions, and how are stochastic outcomes, state dependence, and delayed benefits handled? How to calculate the q-value across Action Packages? Is there evidence on representative states that the Adapter improves subsequent action selection?
  4. What is the reason for the stagnation in Figure 4? Does this point to a limitation of OptiCom itself? Also the caption states that “causal effects and the impossibility of further improvement are not established,” whereas Section 6.4 claims the trajectory “directly validates OptiCom’s core premise.”, which seems confusing.
  5. What is the purpose of proposing two theorems in the paper, they do not seem closely related to Opticom and proved nothing specific.
Reviewer_c7jGOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies LLM-driven iterative optimization, where an LLM repeatedly proposes candidates that are executed and scored (e.g., mathematical constructions, GPU kernels, prompts). Motivated by a diagnostic study showing that the usefulness of a search mechanism depends on the current search state (e.g., execution failures vs. objective stagnation), the authors propose (1) AQOEMS, a six-dimensional functional taxonomy (Artifact, Query, Operator, Evaluation, Memory, Strategy) into which existing LLM optimizers are mapped; (2) a decomposition of terminal improvement into cumulative decision opportunities minus selection losses; and (3) OptiCom, which uses an LLM Optimization Controller to compose a structured “Action Package” (evidence scope, operator, generation breadth, evaluation effort, retention) at every iteration, plus a slower Strategy Adapter that updates selection preferences and templates from trajectory feedback. Experiments on 32 benchmark groups from SkyDiscover compare OptiCom with 13 method-inspired profiles under a 30-iteration cap with Doubao-Seed-2.0-pro (average Max-score rank 1.72, best in 23 groups), and with official EvoX and AdaEvolve implementations on seven benchmarks, where OptiCom is best on all seven.

Critical strengths
  1. A well-motivated controller that conditions on failure type, not only on score progress. The diagnostic study shows that the best mechanism differs between execution-failure and stagnation states. OptiCom's controller reads recent feedback directly and can therefore respond differently to these cases, which improvement-statistics-based adaptation cannot distinguish. This is the most convincing idea in the paper.

  2. Broad evaluation with consistent gains. The method is evaluated on 32 benchmark groups across mathematical, systems, reasoning and prompt-optimization tasks with five runs each, and it outperforms official EvoX and AdaEvolve implementations on all seven benchmarks in Table 2, with sizeable margins on some (e.g., HotpotQA 0.513 vs. 0.409 Max, QN 0.900 vs. 0.850). The Adapter's gain also holds across five backbones (Table 4).

Critical weaknesses
  1. The core contribution is not clearly separated from AdaEvolve, and the comparison with it is not fair. AdaEvolve already adapts exploration intensity per island, allocates budget across islands with a UCB bandit, and, when search stagnates, calls a separate LLM that analyzes the problem, the evaluator and failed attempts to propose new tactics. The last of these is already state-conditioned LLM guidance. The main text describes AdaEvolve only as using "improvement statistics (M) to adapt exploration and allocation (S)"; its meta-guidance level is listed only in the appendix mapping (Table 6), and the paper never contrasts it with OptiCom's controller. The difference therefore reduces to invoking the LLM controller at every iteration and emitting a structured package rather than a tactic. The empirical comparison does not isolate that difference. AdaEvolve was designed and reported with a 100-iteration budget, but here it is capped at 30 iterations, where its stagnation-triggered mechanisms may barely activate; official AdaEvolve reaches only 0.769 on Heilbronn versus 0.961 for OptiCom. In addition, the "inspired" and "official" AdaEvolve rows are identical on Heilbronn, LLM-SQL and GPU without explanation. Finally, all comparisons are aligned on iterations, which the paper itself acknowledges "does not equalize candidate counts, API calls, tokens, or elapsed time." Token usage is reported only for OptiCom (about 865K tokens per Heilbronn run, up to about 50K per iteration), so it is unclear how much of the gain comes from the method rather than from additional LLM computation.

  2. The taxonomy and theory add little rigor. AQOEMS is explicitly "a functional taxonomy rather than a strict architectural blueprint." The six dimensions are hand-chosen, every existing optimizer can be mapped into them after the fact, and the paper gives neither a justification for this particular decomposition nor a testable prediction derived from it. No ablation shows that structuring the controller's decisions along these dimensions helps compared with an unstructured LLM controller. Likewise, the theoretical analysis is, by the authors' own description, a specialization of the existing performance-difference identity, and the Strategy Adapter does not estimate the quantities in the decomposition. The theory therefore neither derives nor constrains the method, yet it is presented as a main contribution. The paper is best judged as an empirical systems contribution, which places the full burden on the fairness of the comparisons in Weakness 1.

Questions for authors

What does OptiCom do that AdaEvolve’s meta-guidance level (an LLM that analyzes failed attempts and proposes tactics on stagnation) cannot? Please compare against official AdaEvolve under its original budget (100 iterations) or, more importantly, under a matched token budget.

Please report total tokens and LLM calls per run for every method in Table 2, and plot score against cumulative tokens. Does OptiCom’s advantage persist when costs are matched?

Suggestion: An ablation in which the LLM controller decides freely, without the AQOEMS-structured Action Package, would show whether the taxonomy contributes beyond being a descriptive lens.

Reviewer_A3nkOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper presents OptiCom, a framework that brings different LLM-driven optimizers (TextGrad, AdaEvolve, DSPy etc.) under a shared configuration space consisting of artifact, query, operator, evaluation, memory, and strategy. The paper argues that current methods each fix one search recipe for the entire run, whereas OptiCom makes the recipe adaptive with a fast LLM Controller that composes a structured Action Package and a slower Strategy Adapter improves selection preferences, operator weights, templates based on the accumulated trajectory feedback. Across 32 benchmark groups, it obtains the best rank than other methods.

Critical strengths

The motivation study is well designed and tests the paper’s premise directly.

The paper runs extensive evaluations across 32 benchmark groups, multiple runs each, with multiple backbones.

Critical weaknesses

The baselines are re-implementations as fixed configurations of the authors system, and are all not shown to be equivalent to their official implementations. Compute is not matched across the evaluations. The methods use a fixed iteration count, while OptiCom makes more LLM calls per iteration for the Controller and Adapter, and token usage is not reported for all methods.

Questions for authors

Please refer to weaknesses:

How does OptiCom compare against the baselines under a matched token budget. Please extend the official-implementation comparison to other baselines with an available implementation, and report results from the official implementation with the OptiCom implementation

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.