{"id":"ab413529-fb99-4323-a9f8-252b00723f57","arxiv_id":"2505.04354","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"An evolutionary loop of foundation-model agents could automate the full optimization pipeline, but the paper's evidence only covers two isolated components.","lead":"This paper argues that optimization problems, from formulating the model to tuning the algorithm, can be handled by an automated pipeline in which large language models and evolutionary search work together. It supports this position with two small case studies, one in cloud scheduling and one in tuning the ADMM optimization method.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Headline claim spans all four subspaces, but both case studies fix problem and formulation spaces; no evidence tests autonomous navigation of the problem space or formulation space.","rationale":"The reader's weakest_assumption focuses on whether foundation-model proposals are good enough for evolutionary selection to improve (§3.1.2 assumptions vs. §6 counterevidence). I agree that is a real risk, but the more precise and more load-bearing gap in the paper's own evidence is the mismatch between the headline claim over O = P ⊗ F ⊗ A ⊗ H and the two case studies, which explicitly reduce P and F to singletons (§4.1, §4.2). That mismatch is stated by the authors themselves and is therefore not a manufactured objection. The claim that the evolutionary loop can bootstrap in A and H is partially supported by the two case studies (suggestive, though lacking error bars and code), and the related-work survey shows separate successes in NL-to-formulation and in LLM+evolution algorithm design, but no integrated navigation of P and F. The limitations in §5 and the alternative-view evidence in §6 (LLM planning failures) further undermine the extrapolation from A/H to P/F, because formulation choice lacks the crisp, simulation-based reward that makes the A/H case studies work. Thus the verdict should remain CONDITIONAL: plausible and worth testing, but the full claim is not yet evidenced. I do not move to REJECT because the paper is a position paper, the evidence it does present is supportive, and the missing test is clearly specified and feasible.","tokens_in":24933,"tokens_out":2013,"duration_ms":18131,"concrete_test":"Run a single end-to-end experiment in which the workflow receives only the natural-language problem description (no pre-supplied formulation or algorithm family) for a nontrivial optimization benchmark (e.g., the NL4Opt competition problems or a selected LibADMM application), with the evolutionary loop allowed to vary formulations and problem interpretations across generations. Report the fraction of generations that produce a feasible, correctly parsed optimization model (verified by an independent solver, not by the LLM's self-report), the best quality achieved relative to the human-formulated baseline, and whether the evolutionary trajectory improves over random restart of foundation-model proposals. If this fraction is near zero or does not improve with generations, the claim that the workflow navigates P and F is unsupported.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The paper's strongest claim is that an evolutionary agentic workflow can autonomously navigate the full optimization space O = P ⊗ F ⊗ A ⊗ H, including the problem space P (natural-language/multimodal problem descriptions) and the formulation space F. Yet the two case studies deliberately shrink the tested space to singletons before evolution: §4.1 says 'we fix the problem formulation and hyperparameter settings to isolate the algorithm design space,' and §4.2 says 'we consider a fixed problem formulation and algorithm design... spaces for problem formulation and algorithm design are singletons.' Thus the empirical support covers only A (new scheduling heuristic) and H∩A (adaptive ADMM penalty schedule). The related-work section (§7, §3.6) cites separate lines of work on NL-to-formulation and on automated algorithm design, but none of it runs the proposed integrated evolutionary loop across P or F. The central conjecture therefore rests on an extrapolation: that the evolutionary mechanism that works when a human has already selected P and F will continue to work when the foundation agent must jointly propose and explore P and F. This is load-bearing because if P and F are not actually navigable (e.g., because problem interpretation is ambiguous and formulation errors are not detectable by the reward signal), the paradigm does not deliver end-to-end automation and instead reduces to AutoML/AAC over A and H. Section 5 itself concedes 'foundation agents lack built-in mechanisms for theoretical verification,' and Section 6 cites evidence that LLMs still fail at multi-step reasoning and planning. These limitations interact with the P/F gap: the first case study's reward is simulation-based scheduling length and the second's is iteration count, both well-defined; no such grounded reward exists for choosing a formulation among equally 'feasible' alternatives. So the weakest link is the untested joint navigation of P and F, not merely candidate quality in A/H.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper is a position paper arguing that optimization problem solving can be automated through \"evolutionary agentic workflows\" that combine foundation models (FMs) with evolutionary search. The authors define an optimization space O as the product of problem space P, formulation space F, algorithm space A, and hyperparameter space H, and propose that an agentic workflow can navigate all four subspaces. Two case studies are presented: cloud VM scheduling, where the workflow evolves a heuristic that outperforms BestFit, and ADMM step-size tuning, where evolved adaptive update rules reduce iteration counts on LibADMM benchmarks. The paper also includes a design for the agent architecture, a review of related work, and a discussion of limitations and alternative views.","tokens_in":25228,"tokens_out":5887,"duration_ms":55685,"significance":"If the central claim were established, the framework would be important: it would replace human expert effort in optimization modeling and algorithm design with automated, adaptive search, potentially accelerating deployment of advanced optimization methods in industry. The paper makes a useful conceptual contribution by unifying problem formulation, algorithm design, and hyperparameter tuning in a single space O and by proposing an explicit agent architecture (memory, reasoning, world modeling, action). It also provides two concrete case studies with transfer tests, which go beyond pure speculation, and it candidly discusses the current lack of theoretical verification and high inference cost. However, the empirical evidence is preliminary and narrow: both case studies fix P and F to singletons, and the reported improvements lack statistical rigor and reproducibility details. The strength of the position therefore rests more on the plausibility of the component-wise literature cited in §3.6 than on the presented demonstrations.","major_comments":[{"comment":"The central claim, stated in the abstract and conclusion, is that the workflow autonomously navigates the full optimization space O = P⊗F⊗A⊗H. However, both case studies explicitly reduce the problem and formulation spaces to singletons: §4.1 says \"we fix the problem formulation and hyperparameter settings to isolate the algorithm design space,\" and §4.2 says \"we consider a fixed problem formulation and algorithm design... spaces for problem formulation and algorithm design are singletons.\" Thus the presented evidence supports navigation only over the algorithm space A (VM scheduling) and over adaptive rules in H (ADMM). This is load-bearing because if P and F cannot be navigated, the paradigm reduces to AutoML/AAC over A and H, and the end-to-end claim fails. The paper should either provide evidence for P and F navigation or explicitly reframe the claim as a conjecture for those subspaces, supported by the separate component-wise literature in §3.6.","section":"§2 (Eq. 1) and §4.1/§4.2"},{"comment":"The ADMM case study reports single-point iteration counts without standard deviations, statistical tests, or details of multiple runs. Since the performance comparison is a central piece of evidence, the lack of repeated trials leaves open the possibility that the results are due to overfitting or to idiosyncrasies of the specific problems. Additionally, the paper does not state what convergence properties, if any, are preserved by the evolved update rules; §5 acknowledges that \"foundation agents lack built-in mechanisms for theoretical verification,\" but the presented demonstration would be considerably stronger if it included at least an empirical convergence check and a comparison of run-to-run variability.","section":"§4.2, Tables 2 and 3"},{"comment":"The reasoning module described in §3.1.2—CoT data construction, domain-focused SFT, and rule-based RL—is presented as a component of the framework, but it is not clear whether this pipeline was actually implemented in the case studies. §4.1 mentions \"300 training epochs\" and the use of DeepSeek-Coder, but does not specify which of the proposed training steps were performed. This conflation of proposed design and implemented system makes it difficult to assess what the experiments actually demonstrate. The authors should clearly state which components of the architecture (memory, reasoning, world modeling, action) were used in each case study and how the reasoning module was instantiated.","section":"§3.1.2 and §4.1"},{"comment":"The description of the VM scheduling experiment omits critical details of the evolutionary search: population size, number of generations, mutation/crossover operators, LLM sampling temperature, selection mechanism, and the exact fitness evaluation procedure. Without these, the reader cannot assess whether the reported improvement over BestFit is robust or a result of particular search settings. The paper should report these settings, along with multiple random seeds and standard deviations, to make the empirical claims reproducible.","section":"§4.1"}],"minor_comments":[{"comment":"The caption contains typographical errors: \"perforcement\" should be \"performance,\" and \"traing\" should be \"training.\"","section":"Figure 4 caption"},{"comment":"The caption says \"Iterative number\" and \"tolenence\"; these should be \"Number of iterations\" and \"tolerance.\"","section":"Table 2 caption"},{"comment":"The word \"summerized\" should be \"summarized.\"","section":"§2"},{"comment":"The phrase \"atlasing the current solution approach\" is unclear; presumably \"adjusting\" or \"updating\" was intended.","section":"§3.1.4"},{"comment":"The section opens with \"there remain two key challenges\" but then lists three (theoretical verification, inference cost, and scalability); please renumber or rephrase.","section":"§5"},{"comment":"The definitions of primal and dual residuals appear swapped relative to the standard ADMM convention: p_k is defined as ρ AᵀB(z^k−z^{k−1}) (usually the dual residual) and d_k as Ax^k+Bz^k−c (usually the primal residual). Please clarify the notation to avoid confusion with the literature cited.","section":"§4.2, Eq. (7)"}],"recommendation":"major_revision","confidential_remarks":"The paper is a position paper with a broad claim and two preliminary case studies. The main concern is the mismatch between the ambitious claim of navigating the full optimization space and the narrow experiments that fix P and F to singletons. This is fixable by reframing the claim as a conjecture and strengthening the empirical rigor, but as written the manuscript overstates its evidence. The paper's candid discussion of limitations is a strength, and the related-work survey is useful. I recommend major revision with the expectation that the authors either add evidence for P/F navigation or explicitly limit the scope of the claim."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Let me give you the short version: this is a position paper with a clean framework and two real case studies, but the evidence only covers two of the four subspaces it claims to automate. The headline claim—that an evolutionary agentic workflow can autonomously navigate the full space O = P⊗F⊗A⊗H—is a conjecture, not a demonstrated result.\n\nWhat's genuinely useful: the product-space decomposition is a good mental model for where human experts sit in the loop. The two case studies are actual attempts to do evolutionary agentic search: one evolves a VM scheduling heuristic against BestFit, the other evolves an adaptive ADMM step-size rule against an expert-designed scheme. Both include transfer tests—to different VM cluster sizes and to held-out LibADMM problems—which is more than most position papers bother to do. The limitations section is honest, and the alternative-views section cites the recent LLM-reasoning critiques rather than ignoring them.\n\nWhere it falls short: in both case studies, the problem space and formulation space are deliberately fixed to singletons (Section 4.1, Section 4.2). The ADMM study fixes the algorithm too; only the hyperparameter policy evolves. So the empirical support covers A and H, not P and F. That matters because the paper's distinctiveness over AutoML/AAC lies precisely in navigating formulation and problem spaces. The authors acknowledge this gap implicitly, and Section 5 admits there is no built-in verification mechanism, but the central claim is still worded far more strongly than the data support. There are also no error bars, no code release, and the tables are single-point comparisons, so the magnitude of the gains is hard to judge.\n\nThis is a fair package for what it is: a position paper with preliminary, suggestive evidence. I would not cite it as a source for empirical claims, but the framework is useful for framing future work. It deserves peer review—a serious referee could push the authors to either narrow the claims or test P/F navigation, which is the actual bottleneck.\n\nRecommendation: accept for review, with the expectation of substantial revision.","headline":"A coherent position paper with a clean product-space framing and two honest-but-partial case studies; the headline claim about navigating P and F outstrips the evidence, which only covers A and H.","tokens_in":25816,"tokens_out":3026,"would_cite":false,"duration_ms":29928,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Optimization problem-solving can run end-to-end through evolutionary AI agents.","keywords":["evolutionary agentic workflow","foundation models","optimization space","automatic algorithm design","hyperparameter adaptation","ADMM","cloud resource scheduling","AutoML"],"falsifier":"Run the identical evolutionary loop with the foundation-model proposer replaced by random mutation of existing candidate code and parameters, keeping the evaluation budget and selection pressure the same; if random search matches the agentic version on the VM-scheduling and ADMM benchmarks, the claim that model knowledge drives the improvement would be refuted. A second check: if more than roughly 90 percent of the model's proposals are rejected as infeasible by the reasoning module across a broad sample of problems, the evolutionary bootstrap cannot work regardless of selection.","tokens_in":24751,"feed_emoji":"🤖","tokens_out":11396,"duration_ms":98118,"temperature":0.7,"pith_summary":"This position paper argues that the main bottleneck in optimization practice is human expertise, and that an evolutionary agentic workflow -- foundation models proposing candidate problems, formulations, algorithms, and hyperparameters, with evolutionary search selecting and refining them -- can remove that bottleneck. It formalizes the search target as the optimization space $O := \\mathcal{P} \\otimes \\mathcal{F} \\otimes \\mathcal{A} \\otimes \\mathcal{H}$, the product of problem, formulation, algorithm, and hyperparameter spaces, so that solving a problem becomes searching for a good point in that space. Two case studies support the position: an evolved virtual-machine scheduling heuristic beats a best-fit baseline on training and test scenarios, and an evolved adaptive ADMM step-size rule reaches the same accuracy in far fewer iterations than an expert-designed scheme, including when transferred to new problems. A sympathetic reader would care because, if the position is right, optimization would no longer depend on a small pool of specialists and academic advances could reach industry without a long expert translation step. The paper itself notes that the workflow still lacks built-in theoretical verification and currently has a high inference cost.","feed_headline":"Evolutionary AI agents can run optimization end to end","feed_subtitle":"A workflow with foundation models plus evolutionary search beats expert heuristics in two real test cases.","key_machinery":"The central object is the optimization space $O := \\mathcal{P} \\otimes \\mathcal{F} \\otimes \\mathcal{A} \\otimes \\mathcal{H}$, where each subspace may be a singleton or an infinite-dimensional functional space, and 'solving an optimization problem' is rewritten as 'searching for a point in $O$'. The argument is carried by a closed feedback loop: foundation agents (built from memory, reasoning, world-modeling, and action modules) generate candidate points; the reasoning module, trained on chain-of-thought data through supervised fine-tuning and rule-based reinforcement learning, filters and validates them; and evolutionary search maintains distributed island populations, clusters solutions to preserve diversity, and uses knowledge from memory to guide the next generation. This machinery is what converts creative, possibly hallucinated proposals into a disciplined search that improves over generations.","core_discovery":"The central claim is that an evolutionary agentic workflow can autonomously navigate the entire optimization space $O := \\mathcal{P} \\otimes \\mathcal{F} \\otimes \\mathcal{A} \\otimes \\mathcal{H}$, not merely tune hyperparameters or select algorithms inside a fixed model. Foundation agents provide broad-domain knowledge, context-sensitive parsing, and creative generation; a reasoning module filters candidates for mathematical and operational validity; and evolutionary algorithms manage population diversity, memory of successful patterns, and selection pressure. The case studies hold the formulation and hyperparameter spaces to singletons in order to isolate single subspaces: the cloud-scheduling study evolves the algorithm space, while the ADMM study evolves the hyperparameter space. In both, the evolved artifact outperforms the expert baseline and generalizes to unseen instances. The authors intend these results as evidence for the general mechanism, which they argue can search over problem descriptions and formulations as well.","pith_inferences":["A testable extension the paper leaves implicit is evolving the formulation space itself: starting from one natural-language description, the loop should move among LP, network-flow, assignment, and set-cover models rather than holding the formulation fixed as a singleton.","If the position is right, the bottleneck shifts from scarce experts to compute and verification: the workflow's industrial viability will be governed by inference cost and by automated proof or validity checking of generated proposals, both of which the paper lists as open problems.","The evolutionary framing suggests the approach may tolerate frequent hallucinations: as long as the reasoning filter passes a small but steady share of viable candidates, selection pressure could compensate, so performance should degrade gradually rather than collapse as proposal quality worsens -- a direct empirical prediction.","The authors do not address the no-free-lunch direction: the claim is most plausible for structured families of industrial problems where foundation models encode transferable knowledge, and least plausible for entirely novel problem structures where random search is the benchmark."],"forward_implications":["A natural-language optimization problem could be carried from formulation to tuned solver with minimal human intervention, lowering the specialist barrier the paper identifies.","The workflow goes beyond AutoML: because the formulation space is part of $O$, the system can reformulate objectives and constraints, not merely pick among fixed algorithm and hyperparameter configurations.","Evolved artifacts can transfer: the VM scheduling algorithm beats the baseline on sizes it was not trained on, and the ADMM step-size rules trained on two problems accelerate other problems in the same library.","Memory, retrieval, fine-tuning, and continual learning could let the workflow remain effective as industrial problem specifications shift over time.","If the position is correct, the relevant comparison class broadens: AutoML, learning-to-optimize, meta-optimization, and algorithm configuration become components or special cases of a single search over $O$."],"supporting_citations":[{"why":"supplies the evolutionary search mechanisms that give the workflow selection pressure and population-based exploration.","marker":"(Yu & Gen, 2010)"},{"why":"surveys combining evolutionary computation with large language models, grounding the proposed generalized evolutionary framework.","marker":"(Wu et al., 2024)"},{"why":"establishes the precedent that large language models combined with evolutionary program search can discover new heuristics.","marker":"(Romera-Paredes et al., 2024)"},{"why":"demonstrates agentic evolutionary heuristic design, supporting the algorithm-design component of the workflow.","marker":"(Liu et al., 2024a)"},{"why":"defines ADMM and its convergence properties, the substrate for the hyperparameter-tuning case study.","marker":"(Boyd et al., 2011)"},{"why":"provides the expert-designed self-adaptive penalty-parameter baseline that the evolved ADMM rule must beat.","marker":"(He et al., 2000)"},{"why":"establishes the line of work in which language models convert problem descriptions into optimization formulations.","marker":"(Ramamonjison et al., 2023)"},{"why":"documents limitations of language models in multi-step reasoning, motivating the filtering reasoning module.","marker":"(Mirzadeh et al., 2024)"}],"fun_headline_variants":["Evolutionary agents auto-pilot optimization end to end","AI agents evolve to solve optimization without experts","From expert tuning to self-evolving optimization agents","Evolutionary agentic workflows beat expert heuristics","Autonomous optimization: foundation models meet evolution"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that foundation models, after filtering by the reasoning module, generate candidates -- formulations, algorithms, and hyperparameters -- that are good enough, often enough, for evolutionary selection to improve upon, despite documented failures of such models at multi-step mathematical reasoning.","fun_headline_variants_meta":{"raw":{"variants":["Evolutionary agents auto-pilot optimization end to end","AI agents evolve to solve optimization without experts","From expert tuning to self-evolving optimization agents","Evolutionary agentic workflows beat expert heuristics","Autonomous optimization: foundation models meet evolution"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000133,"raw_usage":{"total_tokens":1077,"prompt_tokens":829,"completion_tokens":248,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":445,"completion_tokens_details":{"reasoning_tokens":179}},"tokens_in":445,"tokens_out":248,"duration_ms":2838,"temperature":1.0,"reasoning_tokens":179,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T23:30:40.243387+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the identical evolutionary loop with the foundation-model proposer replaced by random mutation of existing candidate code and parameters, keeping the evaluation budget and selection pressure the same; if random search matches the agentic version on the VM-scheduling and ADMM benchmarks, the claim that model knowledge drives the improvement would be refuted. A second check: if more than roughly 90 percent of the model's proposals are rejected as infeasible by the reasoning module across a broad sample of problems, the evolutionary bootstrap cannot work regardless of selection.","supporting_citations":[],"review_version":1}