PaperlayerView exact revision on arXiv

Source-linked AI summary

Reasoning BO: Enhancing Bayesian Optimization with Long-Context Reasoning Power of LLMs

Zhuo Yang, Daolang Wang, Lingli Ge, Beilun Wang, Tianfan Fu, Yuqiang Li

arXiv:2505.12833v2cs.AI

TL;DR

Bayesian optimization can struggle with local optima and interpretable insight, and current LLM-BO approaches can hallucinate suggestions or fail to retain discoveries. Reasoning BO combines confidence-filtered LLM reasoning with dynamic knowledge management and reinforcement learning. It reports stronger performance across diverse optimization tasks, including higher Direct Arylation final yield than Vanilla BO.

  • Problem

    Bayesian optimization faces local-optimum and interpretability challenges, while LLM-BO approaches may hallucinate suggestions or fail to incorporate experimental discoveries into later cycles.

    Source: §1 · p. 1

  • Method

    Reasoning BO integrates LLM-generated hypotheses and confidence filtering into BO, dynamically manages domain knowledge, and fine-tunes lightweight models with reinforcement learning.

    Source: §1 · p. 2

  • Results

    Reasoning BO outperforms Vanilla BO across diverse optimization tasks; in Direct Arylation, final yield was 23.3% higher (94.39% vs 76.60%).

    Source: §1 · p. 2

  • Takeaways & Limitations

    The framework combines LLM reasoning, knowledge graphs, vector databases, and multi-agent architecture in a customizable system for scientific optimization.

    Source: §5 · p. 9

Abstract

from arXiv · show

Many real-world scientific and industrial applications require the optimization of expensive black-box functions. Bayesian Optimization (BO) provides an effective framework for such problems. However, traditional BO methods are prone to get trapped in local optima and often lack interpretable insights. To address this issue, this paper designs Reasoning BO, a novel framework that leverages reasoning models to guide the sampling process in BO while incorporating multi-agent systems and knowledge graphs for online knowledge accumulation. By integrating the reasoning and contextual understanding capabilities of Large Language Models (LLMs), we can provide strong guidance to enhance the BO process. As the optimization progresses, Reasoning BO provides real-time sampling recommendations along with critical insights grounded in plausible scientific theories, aiding in the discovery of superior solutions within the search space. We systematically evaluate our approach across 10 diverse tasks encompassing synthetic mathematical functions and complex real-world applications. The framework demonstrates its capability to progressively refine sampling strategies through real-time insights and hypothesis evolution, effectively identifying higher-performing regions of the search space for focused exploration. This process highlights the powerful reasoning and context-learning abilities of LLMs in optimization scenarios. For example, in the Direct Arylation task, our method increased the yield to 60.7%, whereas traditional BO achieved only a 25.2% yield. Furthermore, our investigation reveals that smaller LLMs, when fine-tuned through reinforcement learning, can attain comparable performance to their larger counterparts.

1 INTRODUCTION

Bayesian optimization can struggle with local optima and lacks interpretable scientific insights, while existing LLM–BO approaches can hallucinate suggestions and fail to retain experimental discoveries. Reasoning BO combines confidence-filtered reasoning, dynamic knowledge management, and lightweight-model reinforcement learning; its Direct Arylation benchmark reports higher final yield than Vanilla BO.

1 INTRODUCTION

  • Bayesian optimization faces local-optimum, initialization-sensitivity, and interpretability challenges, while LLM–BO approaches can hallucinate suggestions and fail to incorporate experimental discoveries into later cycles.These limitations motivate memory-augmented closed-loop optimization.Source: §1 · p. 1
  • Reasoning BO embeds an LLM in the BO loop to generate hypotheses and confidence scores, filters candidates for plausibility, updates knowledge stores, and fine-tunes lightweight models with reinforcement learning.The knowledge system combines structured domain rules in knowledge graphs with unstructured literature in vector databases.Source: §1 · p. 2
  • The paper’s contributions include a reasoning model for evolving hypotheses and confidence-based filtering, plus a full-cycle system that dynamically updates domain expertise during optimization.The system also supports customizable knowledge storage and open agent interfaces.Source: §1 · p. 2
  • 23.3% higher final yield (94.39% vs 76.60%) was achieved by Reasoning BO compared with Vanilla BO in the Direct Arylation benchmark.The reported initial performance was also 44.6% higher (66.08% vs 21.62%).Source: §1 · p. 2

2 RELATED WORK

The related work describes Bayesian optimization’s surrogate-and-acquisition framework and practical challenges, then surveys language-guided optimization and reinforcement-learning approaches to reasoning. It positions Reasoning BO against work on candidate sampling, knowledge transfer, and persistent use of reasoning insights.

2.1 BAYESIAN OPTIMIZATION

  • Bayesian optimization iteratively refines probabilistic surrogate models and uses acquisition functions to balance exploration and exploitation, but scaling costs and suboptimal local optima remain practical challenges.Expected Improvement is noted as a popular acquisition strategy.Source: §2.1 · p. 2

2.2 LANGUAGE-GUIDED BAYESIAN OPTIMIZATION

  • Language-guided BO work includes LLM-assisted candidate sampling and knowledge transfer, while cited approaches also use meta-learning initialization and textual chemical procedures as optimization features.Examples include SLLMBO, HypBO, BORA, MI-SMBO, and BOChemian.Source: §2.2 · pp. 2–3
  • Existing LLM-BO approaches face risks from hallucinations, manual prompt dependence, weak knowledge persistence, and possible benchmark data leakage; Reasoning BO adds evolving knowledge graphs, agents, and retrieval.The proposed retrieval-augmented generation mechanisms are intended to improve knowledge reuse and correct factual errors.Source: §2.2 · p. 3

2.3 REINFORCEMENT LEARNING-ENHANCED REASONING IN LLMS

  • Reinforcement-learning research uses policy optimization, including GRPO, to produce more structured reasoning; Figure 1 contrasts keyword-retrieved Info with CoT-derived Insights.The caption defines Info as database information retrieved by question keywords and Insights as critical insights extracted through CoT analysis.Source: §2.3 · p. 3, Figure 1 · p. 3

3 METHODOLOGY

Reasoning BO addresses the limited use of domain knowledge and persistent reasoning in optimization by coupling BO with LLM analysis and a continuously updated knowledge system. Its workflow combines BO-proposed candidates, confidence-ranked hypotheses, multi-agent knowledge storage, and post-training for smaller models.

3.1 MOTIVATION

  • Classical BO has difficulty using implicit domain knowledge and retaining CoT insights; Reasoning BO combines reasoning-enhanced BO, multi-agent knowledge management, and model post-training.The framework extracts structured notes from reasoning trails for later retrieval and delivers optimization results with interpretable insights.Source: §3.1 · pp. 3–4

3.2 REASONING-ENHANCED BAYESIAN OPTIMIZATION

  • The LLM initializes the search from an Experiment Compass, then analyzes BO-proposed candidates, generates hypotheses, assigns confidence scores, and selects the highest-confidence candidates for evaluation.Each round updates the knowledge base and surrogate model; BO proposes five candidates and the method selects three, while numerical optimization remains with BO.Source: §3.2 · p. 4

3.3 MULTI AGENTS SYSTEM FOR DYNAMIC KNOWLEDGE PRECIPITATION

  • Verifier, Formatter, and Notes agents extract, verify, structure, and store knowledge as triples in a knowledge graph and semantic information in a vector database.During optimization, keyword queries retrieve graph relationships first and contextual information from the vector database afterward.Source: §3.3 · p. 5

3.4 POST TRAINING STRATEGIES

  • Small-model post-training combines domain-specific LoRA supervised fine-tuning with GRPO on GSM8K; the authors report transfer to scientific BO guidance and retrieve graph and vector knowledge during reasoning.The knowledge-integration phase queries the knowledge graph and then the vector space at each reasoning step.Source: §3.4 · pp. 5–6

4 EXPERIMENTS

The experiments assess Reasoning BO on real-world and synthetic optimization benchmarks, then examine extended tasks and component ablations.

4.1 EXPERIMENTAL SETUP

  • The setup spans three real-world and three synthetic benchmarks, uses three candidates per round for up to 30 experiments, and reports averages over 10 fixed-seed trials.The real-world tasks include two chemical reactions and Lunar Lander; synthetic tasks include Ackley, Rosenbrock, and Hartmann.Source: §4.1 · p. 6

4.2 RESULTS

  • Reasoning BO robustly and efficiently outperforms baselines across real-world chemical optimization tasks and synthetic mathematical functions.Figure 3 plots the best observed objective value, with average trajectories and standard-error shading.Source: Figure 3 · p. 7, §4.2 · p. 7

4.3 EXTENDED EXPERIMENTAL ANALYSIS

  • Reasoning BO improves CPA-Catalyzed Thiol–Imine by 0.20 kcal/mol at initialization and 0.67 kcal/mol during convergence, and Buchwald-Hartwig yield by 10.37% (95.06% vs. 84.68%).The comparison is against vanilla BO.Source: §4.3 · p. 7, Figure 4 · p. 8
  • On Lunar Lander (12D), Reasoning BO gains 71.02 in initial reward over vanilla BO; on Ackley (15D), its difference is -2.34 initially and -1.46 at 90 observations.The reported high-dimensional trajectories are shown in Figure 5.Source: §4.3 · p. 7, Figure 5 · p. 8

4.4 ABLATION STUDIES

  • RBO initialization yields 44.6% higher initial Direct Arylation performance than Vanilla BO (66.08% vs 21.62%), while RBO optimization outperforms VBO across initialization methods.RBO initialization followed by VBO optimization can surpass full VBO, but is less efficient than full RBO.Source: Figure 6 · p. 8, §4.4 · p. 8
  • Structured notes improve chemical benchmarks, but constrain exploration and slow optimization on the Hartmann function, where prior knowledge is limited.Suzuki coupling has 2% better initial performance with notes; Direct Arylation improves during mid-to-late optimization.Source: Figure 7 · p. 9, §4.4 · p. 9
  • RL-enhanced models improve on baseline approaches, with RL-14b performing best across multiple benchmarks; distilled variants remain competitive with lower computational requirements.The model comparisons include larger models to assess scalability.Source: Figure 8 · p. 9, §4.4 · p. 9

5 CONCLUSIONS

The conclusion presents Reasoning BO as an integrated system for scientific discovery and outlines directions for further development.

5 CONCLUSIONS

  • Reasoning BO integrates knowledge graphs, multi-agent systems, and reinforcement learning with Bayesian optimization to support scientific discovery.Its knowledge graphs and vector databases support dynamic fusion of domain knowledge into optimization, while its multi-agent architecture is customizable.Source: §5 · p. 9
  • Future work targets context-length constraints, more efficient knowledge extraction through distillation, and synergies between reinforcement learning and generated data.Source: §5 · p. 9

A EXTENDED RESULTS

The API-model comparison finds strong, comparable chemical-task results for qwq-plus and DeepSeek, while their Hartmann trajectories differ.

A.1 IMPACT OF DIFFERENT API MODELS

  • On chemical datasets, qwq-plus and DeepSeek substantially outperform baseline BO with comparable results; on Hartmann, qwq-plus is more consistent while DeepSeek explores less later.The distilled model has a slight advantage over baseline BO, and the comparison covers Suzuki coupling, Direct Arylation, and 6D Hartmann.Source: Figure 9 · p. 16, §A.1 · p. 16

A.2 ERROR BAR ANALYSIS

The error-bar analysis covers chemical and high-dimensional benchmarks, reporting mean and standard deviation across multiple trials.

A.2 ERROR BAR ANALYSIS

  • For CPA-Catalyzed Thiol-Imine and Buchwald-Hartwig amination, Figure 10 presents trial means and standard deviations to compare framework consistency and robustness against baselines.Source: Figure 10 · p. 16
  • For Lunar Lander (12D) and Ackley (15D), Figure 11 presents trial means and standard deviations to assess stability and effectiveness on high-dimensional problems.Source: Figure 11 · p. 17

B.1 LOW-RANK ADAPTATION (LORA) IMPLEMENTATION

This material describes LoRA’s low-rank parameter updates, the experimental setup, and the structured input used to guide Bayesian optimization. It then outlines LLM-generated initialization hypotheses for exploring the reaction space.

B.1 LOW-RANK ADAPTATION (LORA) IMPLEMENTATION

  • LoRA freezes pretrained weights and adds trainable low-rank updates to attention layers, reducing trainable parameters while adapting to domain-specific optimization tasks.The update is expressed as ΔW = BA and h = W0x + BAx.Source: §B.1 · p. 17

B.2 BASELINE CONFIGURATION DETAILS

  • The CMA-ES baseline used population size 5, six iterations, σ0 = 0.1, three Sobol initial points, UCB with β = 0.1, and 30 trials per run across 10 repeats.Variables were bounded to [0, 1] in the normalized search space.Source: §B.2 · pp. 17–18

C.1 1. EXPERIMENT COMPASS INPUT

  • The Experiment Compass specifies Buchwald-Hartwig reaction components, categorical parameter bounds, and the objective of maximizing isolated amine yield, with one choice per category per run.The categories are ligand, additive, base, and aryl halide.Source: §C.1 · pp. 18–19

C.2 2. OVERVIEW GENERATION

  • The overview frames optimization as surrogate-guided exploration of four categorical variables, with domain-informed ligand–additive, base–substrate, and halide suggestions for plateaus.The one-choice-per-category constraint is intended to avoid overparameterization and maintain interpretability.Source: §C.2 · p. 20

C.3 3. CANDIDATE INITIALIZATION

  • For initialization, the LLM proposes diverse, chemically justified candidates from the defined parameter space, and those initial experiments and data inform later suggestions.The recommendations seek contrasting chemical properties to cover the space and identify high-yield regions.Source: §C.3 · pp. 21–23

4. Optimization Loop

The optimization loop feeds experimental history and optimizer candidates to an LLM, which revises hypotheses and recommends new settings for evaluation. The examples report high-yield parameter combinations and use observed outcomes to shape subsequent proposals.

4. Optimization Loop

  • Each loop gives the LLM historical experimental data, prior insights, new optimizer candidates, and task instructions; its recommendations are evaluated and fed back into later rounds.The example prompt asks the model to identify trends, revise hypotheses, and propose candidates based on the available data.Source: §4 · pp. 23, 32, 33
  • The Buchwald-Hartwig example reports that reduced ligand bulkiness, a polar additive, tertiary amine base, and methoxy-substituted aryl halide significantly boost yield.The accompanying account contrasts an underperforming bulky-ligand trial with success attributed to a moderate ligand and improved additive solubility.Source: §4 · p. 24
  • After the reported Buchwald-Hartwig results, the LLM proposes testing cyclohexyl-substituted ligands and iodide halides as untested avenues.The recommendations retain previously promising components in some proposed combinations.Source: §4 · p. 25
  • 97.32 objective mean value was recorded for the sulfonyl electrophile with boronic-acid nucleophile, diphenylphosphino-iron ligand, CsF, and DMF.The table also reports 8.67 for the chloride electrophile under the listed matched conditions.Source: §4 · pp. 31–32

C.4 5. SUMMARY AND CONCLUSION

The optimization reached a 99.6% maximum yield within 30 trials, with the conclusion linking performance to component synergies and evolving hypotheses. The section also describes four reaction benchmarks with distinct optimization challenges.

C.4 5. SUMMARY AND CONCLUSION

  • 99.6% maximum yield, a 658% improvement over the initial trial, was achieved for Buchwald-Hartwig amination within 30 trials.The conclusion reports that systematic exploration of substituent effects and catalyst-component synergies led to near-optimal yields.Source: §C.4 · p. 35
  • Hypotheses shifted from bulky ligands and electron-deficient additives toward iodide nitrogen aryl halides as trials revealed stronger-performing combinations.The reported optimization journey notes that confidence in iodide halides rose after a 92.6% trial and that a later trial reached 99.6%.Source: §C.4 · pp. 34–36
  • The top combination pairs non-bulky Ligand3, Additive2, Base3, and nitrogen-substituted iodide Aryl Halide14; bulky ligands disrupted coordination.The conclusion identifies substituent placement and additive-electron interactions as decisive, and reports that the top configuration repeatedly achieved yields above 99%.Source: §C.4 · pp. 34, 36
  • The authors recommend replicating the top configuration and exploring bromine-based nitrogen aryl halides as potentially scalable alternatives to iodide.The stated rationale for exploring bromine is cost; the conclusion also recommends further validation to confirm reproducibility.Source: §C.4 · p. 36

D.1 SUZUKI REACTION (5D)

  • The Suzuki benchmark tests optimization across categorical chemical choices with strong dependencies, including combinations that can be incompatible or cause catalyst poisoning.The benchmark is framed as a test of identifying synergistic effects while avoiding chemical incompatibility.Source: §D.1 · p. 38

D.2 DIRECT ARYLATION REACTION (5D)

  • Direct Arylation optimization uses sensitive ligand, base, solvent, concentration, and temperature choices, with subtle condition changes able to alter selectivity or decomposition.The benchmark is presented as a realistic design problem where minimizing reaction trials matters because experiments carry material and time costs.Source: §D.2 · pp. 38–39

D.3 BUCHWALD-HARTWIG REACTION (4D)

  • The Buchwald-Hartwig benchmark evaluates a four-dimensional categorical space where multimodality, catalyst deactivation, and component synergies complicate optimization.The task concerns a palladium-catalyzed C–N coupling reaction used in pharmaceutical synthesis and materials science.Source: §D.3 · pp. 39–40

D.4 CHIRAL PHOSPHORIC ACID–CATALYZED THIOL–IMINE ADDITION REACTIONS (3D)

  • The CPA-catalyzed thiol–imine benchmark maximizes ∆∆G for enantioselectivity in a nonlinear, multimodal categorical space with catalyst–substrate synergies.Higher ∆∆G indicates stronger stereoselective induction, and the benchmark tests whether algorithms can identify synergistic combinations while avoiding failed reactions.Source: §D.4 · pp. 40–41

D.5 LUNAR LANDER TASK (12D)

This appendix documents the optimization benchmarks, baselines, objects, and prompts used in Reasoning BO, including formats and schemas for guiding iterative hypotheses and extracting reasoning and experiment notes. It illustrates the schema workflow with Buchwald–Hartwig amination as a constrained optimization case study.

D.5 LUNAR LANDER TASK (12D)

  • The Lunar Lander task maximizes average landing reward across 50 terrain variations in a 12D hybrid space, testing delayed-reward planning and precise control against fuel efficiency.Successful landing requires resting within the target zone; only one action may be active at a time, and state variables must remain within physical bounds.Source: §D.5 · pp. 41–42

D.6 LEVY FUNCTION (5D); D.7 HARTMANN FUNCTION (6D)

  • The 5D Levy benchmark has a global minimum of 0 at [1, 1, 1, 1, 1] amid rugged local minima, while 6D Hartmann has an approximate global minimum of −3.32237 among about 10^6 local minima.Levy’s search bounds are [−10, 10]; Hartmann’s are (0, 1), with nonlinear coupling and barriers of ∆f ≥0.5. The passages cite Laguna & Mart´ı (2005) and Picheny et al. (2012), respectively.Source: §D.6 · p. 42, §D.7 · p. 42

D.8 ACKLEY FUNCTION (2D)

  • The 2D Ackley function has a global minimum of 0 at [0, 0], with a flat outer region and narrow optimum that challenge global exploration and precise local search.Its search bounds are [−32.768, 32.768], and its landscape contains numerous shallow local minima.Source: §D.8 · pp. 42–43

D.9 ROSENBROCK FUNCTION (3D)

  • The 3D Rosenbrock function has a global minimum of 0 at [1, 1, 1], and its long, narrow valley and oscillating gradient directions challenge convergence.The search bounds are [−5, 10]; the passage cites Zimmermann (1979).Source: §D.9 · p. 43

E BASELINES; E.2 CMA-ES

  • The baselines emphasize acquisition-function variants, particularly LogEI, and include CMA-ES as a stochastic, derivative-free baseline for nonlinear continuous optimization.CMA-ES adapts the mean and covariance of its search distribution and uses evolution-path tracking.Source: E BASELINES · p. 43, §E.2 · p. 44

E.1 LOGEI FAMILY; E.1.1 ANALYTIC LOGEI; E.1.2 MONTE CARLO PARALLEL LOGEI

  • LogEI uses log-space transformations to improve numerical stability over standard Expected Improvement; its analytic form uses piecewise decomposition, while Monte Carlo LogEI supports parallel batches.The passages cite Ament et al. (2023) for LogEI and describe Monte Carlo draws and temperature parameters in the batch approximation.Source: §E.1 · p. 44, §E.1.1 · p. 44, §E.1.2 · p. 44

F.2 OVERVIEW; F.4 OPTIMIZATION; F.4.1 INITIALIZATION PROMPTS

  • The Overview object conveys objectives and constraints for initial setup, while optimization proceeds through initialization and iterative refinement guided by LLM-generated hypotheses.Initialization uses the Overview and Experiment Compass, requests diverse, justified hypotheses, and prohibits reliance on memorized outcomes or prior experimental data.Source: §F.2 · p. 47, §F.4 · p. 50, §F.4.1 · pp. 50–51

F.3 INSIGHTS OBJECT

  • Reasoning BO’s Insights Object updates hypothesis confidence from accumulated evidence to refine sampling; its Suzuki example highlights concentration and temperature as promising variables.The example reports a highest-yielding trial at 0.057 mol/L and 90°C, while proposing lower concentrations and ambient temperatures as potentially beneficial.Source: §F.3 · p. 49

F.4.1 INITIALIZATION PROMPTS

  • Initialization prompts require parameter names and values to match predefined options exactly; deviations are automatically rejected.Source: §F.4.1 · p. 52

F.4.2 OPTIMIZATION LOOP PROMPTS

  • The optimization-loop prompt asks the LLM to use current experiment data and prior insights to revise hypotheses and propose 3–5 hypotheses, each with at least one constraint-satisfying point.The agent receives the Experiment Compass and latest Insights Object; it may discard low-performing hypotheses and propose new ones.Source: §F.4.2 · pp. 52–53

F.6 EXTRACT NOTES PROMPTS

  • The prompts extract structured notes from both the Experiment Compass and reasoning data.Source: §F.6 · p. 59

F.7 CUSTOM NOTES SCHEMA DESIGN

  • The framework uses customizable note schemas to structure extracted knowledge for downstream domains, including domain-specific formats such as chemistry SMILES expressions.ReasoningNotesResponse captures findings, parameter relationships, and optimization principles; CompassNotesResponse covers experimental descriptions and scientific background.Source: §F.7 · pp. 60–61

F.1 EXPERIMENT COMPASS OBJECT

The Experiment Compass standardizes natural-language experiment descriptions, variables, constraints, and optimization targets. After experiments, Reasoning BO synthesizes accumulated insights and data into interim conclusions.

F.1 EXPERIMENT COMPASS OBJECT

  • The Experiment Compass captures experiment context, constraints, parameter definitions, and optimization targets in a standardized natural-language template.Its descriptions provide prior knowledge for LLM-generated overviews and initial points, while standardized variable names support insights across experiment rounds.Source: §F.1 · pp. 45–46
  • The Suzuki example specifies compatible reaction conditions and seeks to maximize cross-coupled product yield across electrophile, nucleophile, ligand, base, and solvent choices.The template identifies yield as the target and requires chemically compatible reagents and conditions that maintain intermediate stability.Source: §F.1 · pp. 46–47

F.5 EXPERIMENT SUMMARY AND CONCLUSION

  • The Experiment Summary synthesizes accumulated Insight History and experimental data into interim conclusions about hypothesis evolution, evidence, confidence shifts, and key findings.The process ends with a table summarizing changes in confidence in the hypotheses.Source: §F.5 · p. 54

Experiment Summary

The Suzuki experiment summary reports data-supported findings about reaction conditions and tracks how confidence in hypotheses changed. The accompanying material specifies how to extract scientific notes and discloses the manuscript’s limited editorial use of an LLM.

Extracting Notes from Experiment Compass

  • The summary reports how hypotheses evolved and were supported or refuted, confidence shifts, and conclusions, using experimental data rather than prior assumptions.The extraction instructions also request theoretical background, variable properties and relationships, and other relevant scientific notes.Source: Experiment Summary · p. 54, Extracting Notes from Experiment Compass · p. 59

Experiment Summary

  • Triflate electrophiles outperformed halides by 3-10x; DMF/NMP reduced yields by >50% versus dioxane, while chloride/bromide systems stayed below 45%.Repeated validation increased confidence in the triflate/trifluoroborate/K3PO4/ferrocene combination.Source: Experiment Summary · pp. 54–55
  • 74.65% was the peak yield with trifluoroborate and K3PO4, outperforming boronic acids/pinacol esters (<44.08%); ferrocene ligands beat bulky phosphines (74.65% vs. 15.15-30.48%).The reported optimal system combines triflate, trifluoroborate, K3PO4, a ferrocene ligand, and dioxane.Source: Experiment Summary · p. 55

G USAGE OF LARGE LANGUAGE MODELS IN THIS MANUSCRIPT

H LIMITATIONS AND FUTURE WORK

The discussion identifies context-growth and instruction-following risks, and notes that the framework currently handles only single-objective optimization.

H.1 CONTEXT WINDOW CONSTRAINTS

  • Long optimization histories may approach context limits and raise computational costs, although current experiments have not reached that boundary.More complex or prolonged campaigns could incur higher costs as context expands.Source: §H.1 · p. 62

H.2 INSTRUCTION-FOLLOWING CAPABILITY DEPENDENCIES

  • 3B–7B models frequently fail format compliance, and task-specific adaptation may weaken generalization and instruction following.Valid outputs must match the Insight format, satisfy constraints, and omit extraneous markers.Source: §H.2 · pp. 62–63

H.3 MULTI-OBJECTIVE OPTIMIZATION POTENTIAL

  • The framework currently supports single-objective optimization, leaving trade-offs such as yield and cost for future multi-objective work.The paper identifies modeling these relationships and explaining optimization decisions as future directions.Source: §H.3 · p. 63
Loading 2505.12833v2…