acceptodds
Under review as a conference paper at ICLR 2027

Optimizing the Optimizer: Language Models Discover Faster Molecular Relaxation Algorithms

Abstract

Geometry optimization is a major cost in many quantum-chemical workflows: each optimization step requires one force evaluation, and at the density-functional level that evaluation dominates the wall time. Research in this area has produced a broad range of optimization methods, and we ask whether a language model can improve on the best of them through autoresearch. An agent rewrites the optimizer itself to minimize force-call counts, restrained by two admission gates that reject premature stopping and improvements that do not generalize to unseen molecules. Starting from Sella, the fastest open-source optimizer available, the search produces AutoSella, a family of two optimizers. Both of them deliver consistent force-call reductions relative to Sella across held-out molecular benchmarks and potentials not used during the search. Most notably, at the r2SCAN-3c DFT level, the best variant requires only —% of Sella's force calls while achieving the same energy reduction, even though agent used no DFT gradients.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
autoresearch, automated algorithm discovery, LLM-assisted research, AI for science, molecular geometry optimization, computational chemistry, quantum chemistry, optimization algorithm

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_gnVgOpenReview ↗
Rating
2(+1)
Confidence
3
Informativeness
1(−1)
Presentation
2
Soundness
1(−1)
Read review
Paper summary

The paper introdcues AutoSella, an LLM-based framework that modifies the molecular geometry and reduce the number of force evaluations. The paper develop two variants and test them on four benchmarks. The results show that the approach reduces computational costs while maintaining accuracy.

Critical strengths

Generalization evaluation: the experiments cover four benchmarks and provide some evidence of generalization beyond the training setting.

Critical weaknesses

Limited technical contribution: the mathematical formulations and explanations provide little insight into why the proposal works.

Questions for authors

Can the author explain why the discovered modifications imrpove Sella?

How robust are the proposed framework?

Can the author provide more detailed experiments to explain the improvement?

1 comment

Response to Reviewer gnVg

Weakness 1 — Limited technical insight into the modifications

Answer. We added Sections 4.3–4.4 to describe the changes to Sella, trace their relationship to existing optimizers, and report the effect of removing each component or group of related changes. Appendices G–I give pseudocode and mathematical formulations, while Appendix J gives ablations results.

Question 1 — Why do the discovered modifications improve Sella?

Answer. We have added sections 4.3, 4.4 and Appendices G, H, I to the manuscript.

Question 2 — How robust is the search?

Answer. We have conducted three additional independent 66-hour searches with the same GPT-6.1-Sol Max configuration and one search with Claude Opus 5.5 Max. All runs produced optimizers that saved force calls on validation and exceeded Sella’s mean energy recovery. Please see the newly added Appendix M and Table 10 for details.

Question 3 — What further experiments explain the improvement?

Answer. We performed the search using GFN2-xTB and tested transfer to two unseen potentials, GFN-FF and r2SCAN-3c DFT. We have also tested generalization to new chemical domains: amino acid–ligand pairs and solvated drug-like molecules. AutoSella-F used 31.4% and 59.8% fewer DFT force calls per molecule than Sella, respectively, with greater mean energy recovery on both. These experiments demonstrate generalizability of the optimizer.

Reviewer_hYe1OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents AutoSella, an autoresearch-based approach to improving the Sella molecular geometry optimizer. It uses LLM agents to iteratively modify the optimizer under validity and generalization constraints. The resulting optimizers reduce force evaluations across several molecular benchmarks, with improvements transferring to unseen potentials, including DFT.

Critical strengths
  • The paper addresses an important and practical problem in computational chemistry, with encouraging improvements in optimization efficiency.

  • The extensive evaluation across molecular systems and potential functions provides convincing evidence of the practical effectiveness of the resulting optimizers.

Critical weaknesses

The main weakness is the limited evaluation of the autoresearch algorithm itself. Algorithm 1 largely follows the existing autoresearch paradigm, with additional validity and generalization gates. While the resulting optimizers perform well, it is unclear whether the proposed research procedure offers advantages over existing autoresearch or automated program-search methods.

Questions for authors
  • How does Algorithm 1 compare with existing autoresearch-like approaches?

  • Could the authors provide ablations of the key design choices in Algorithm 1?

3 comments

Weakness 1 — Limited evaluation of the autoresearch procedure

Answer. We added experiments on the procedure itself. Appendix L explains key design choices and provides ablations for heuristics in our autoresearch loop, while Appendix M measures variation across independent searches and compares with OpenEvolve. The general comment summarizes both sets of experiments.

Question 1 — How does Algorithm 1 compare with existing autoresearch approaches?

Answer. Algorithm 1 is an adaptation of the autoresearch loop for our selected task. We explain its design choices in Appendix L. To compare with other autoresearch workflows, we used OpenEvolve (see the general comment, Table 10). It searches a population of programs and selects candidates using training scores alone. After 500 iterations, its best training candidate saved 1.97% of validation force calls and fell below Sella’s mean energy recovery, while consuming approximately 4 times more API credits. We assume this inferior performance is due to the lack of diagnostic tools and agentic capabilities.

Question 2 — What do ablations of the key design choices show?

Answer. We have added Appendix L. Also please refer to general comment and Table 9.

Thanks the authors the detailed responses, which help solves some of my concerns. After careful evaluation, I would like to maintain my score.

Thank you for taking the time to consider our responses. Could you please let us know which concerns remain and what changes to the manuscript or additional evidence would lead you to consider a higher score?

Reviewer_Eaq7OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper applies an autoresearch workflow to modify Sella, producing two molecular geometry optimizers, AutoSella-F and AutoSella-A. It evaluates the resulting optimizers against Sella and other established molecular optimization methods, and uses component ablations to assess the modifications. It also investigates whether improvements developed using inexpensive GFN2-xTB evaluations transfer to more expensive DFT calculations. The experiments primarily evaluate the resulting optimizers instead of the autoresearch procedure that produces them.

Critical strengths

The transfer from inexpensive evaluations to DFT is encouraging, but does not resolve the concerns below.

Critical weaknesses
  1. Limited methodological novelty and insufficient evidence of repeatability. The framework largely consists of an existing coding agent iteratively modifying code using evaluation feedback. The paper does not clearly establish what new AI methodology or generalizable insight this introduces. Moreover, each main agent configuration is represented by a single discovery run; using different agents does not substitute for independent repetitions under the same configuration. Consequently, the reliability and variability of the discovery process remain unclear.

  2. The evaluation does not establish the framework-level contribution. The baselines compare the resulting AutoSella implementations with existing molecular optimizers, while the ablations remove components from those implementations. These experiments establish the utility of the resulting artifacts, but do not isolate the contribution of the discovery framework. There are no matched-budget comparisons against alternative autoresearch approaches, such as appropriately adapted AIScientist or OpenEvolve, or a simpler coding-agent feedback loop. Together with the insufficiently established algorithmic novelty of AutoSella itself, this leaves the contribution to AI methodology unclear.

Questions for authors
  1. What is the central contribution: a better molecular optimizer, or a reliable method for discovering optimizers? For the former, which mechanisms are substantively new relative to existing molecular optimization techniques? For the latter, what distinguishes the proposed procedure from existing autoresearch workflows?

  2. What evidence separates the contribution of your framework from the capabilities of Fable and GPT Astra themselves? In particular, do comparisons using the same underlying models and matched development budgets, together with independent repetitions, support an advantage over a simpler feedback-driven coding-agent workflow?

1 comment

Weakness 1 — Methodological novelty and repeatability of discovery

Our primary aim is to find a better molecular optimizer, and building a task-specific harness was a necessary part of this search. We agree that the previous version left this implicit and focused mostly on the search framework. We have rewritten the contribution statement accordingly (see the general comment).

We see two generalizable insights. First, molecular geometry optimization underlies most routine quantum-chemistry calculations and has been refined for decades. Even so, a coding agent with a task-adapted harness substantially improved it. Second, the agent could evaluate only GFN2-xTB during the search, yet the improvements transfer to a force field (GFN-FF) and to DFT (r2SCAN-3c), which are fundamentally different methods. The discovered changes are therefore not heuristics tied to one level of theory. Together, these results suggest that our harness, molecular datasets and GFN2-xTB-based evaluation pipeline can serve as a baseline setup for improving other molecular algorithms. We release all code and data (Section 6).

Our methodological contribution is a harness that steers the agent away from improvements that do not generalize (Appendix L), together with an interpretation of the discovered algorithms (Sections 4.3, 4.4, and Appendices G–J).

Because our goal was the best optimizer rather than a reproducible pipeline, the main Fable and Astra configurations were each run once. To assess repeatability, we ran three independent 66-hour searches with GPT-6.1-Sol Max under the same configuration (general comment, Appendix M, Table 10).

Weakness 2 — Optimizer comparisons do not isolate the contribution of the search procedure

As noted in our answer to Weakness 1, framework-level contribution is not the main focus of this work. Nevertheless, our task required substantial adaptation of the autoresearch loop and a dedicated harness. The new Appendix L describes these design choices and explains why each is needed. To isolate their contribution, we started from a simple coding-agent feedback loop and added the components one at a time. We reran the search with GPT-6.1-Sol Max for each version, using the same starting algo and 24 active research hours (Appendix L.3, Table 9). We also conducted OpenEvolve run from the same starting code: its selected optimizer saved only 1.97% of validation force calls and fell below Sella’s mean energy recovery (Appendix M, Table 10). The general comment summarizes both experiments. Finally, regarding the algorithmic novelty of AutoSella itself, Sections 4.3–4.4 relate each change to existing optimizers and highlight the new elements.

Question 1 — What is the central contribution, and what is new relative to existing methods?

The central contribution is the optimizer. Sections 4.3 and 4.4 describe all changes to Sella, relate them to existing optimizers and highlight the new elements. Appendices G–I give pseudocode and derivations, and Appendix J reports component ablations.

Question 2 — What separates the effect of the workflow from the capabilities of Fable and Astra?

We ran these comparisons with GPT-6.1-Sol Max, keeping the model, starting algorithm and search duration fixed. First, the validity gate prevents the agent from exploiting the convergence criteria. Without it, agents found shortcuts, such as re-evaluating forces at unchanged coordinates, that save force calls but leave structures underoptimized (Appendix L.2). Without the generalization gate, the autoresearch loop finds optimizers that recover less energy than Sella on validation; only the gated versions recover as much energy as Sella on both training and validation (Appendix L.3, Table 9). Second, our agent can inspect per-molecule results and run diagnostic experiments. OpenEvolve, run with the same model (GPT6.1-Sol Max) and starting code, selected candidates by the aggregate training score alone, with no per-molecule data or tools. Its selected optimizer saved only 1.97% of validation force calls and fell below Sella’s mean energy recovery (Appendix M, Table 10). The general comment summarizes both experiments.

Comments on the paper

2 comments
AuthorsOpenReview ↗

General Comment

All three reviewers raised closely related questions about why AutoSella improves on Sella, which parts of the autoresearch loop matter, and how repeatable the gains are. We thank the reviewers for these suggestions and address them in a single general response.

We have updated the manuscript and rewritten our contribution statement. Our primary contribution is AutoSella-F, a state-of-the-art molecular geometry optimizer for force-call efficiency. Across four held-out GFN2-xTB benchmarks, AutoSella-F had the lowest mean relative force-call cost compared with 27 established optimizers. The savings also transferred to r2SCAN-3c DFT, where it used 22.8–59.8% fewer force calls per molecule than Sella and achieved greater mean energy recovery on all four benchmarks (Tables 1 and 2). We present the search procedure as an adaptation of existing autoresearch loop and evaluate its design choices separately. The headings of new or substantially reworked sections are marked in red.

Why do the optimizers use fewer force evaluations? (Sections 4.3 and 4.4)

In Sections 4.3 and 4.4, we describe in detail all the changes that AutoSella-F and AutoSella-A introduce to Sella. We explain how these changes relate to techniques used in existing optimizers, highlight novel ideas and design choices, and report how validation force-call cost changes when individual components or groups of related changes are disabled.

What do the autoresearch gates contribute? (Appendix L, Tables 8 and 9)

We document and justify all major heuristics in our autoresearch loop adaptation. We first explain why the validity gate, isolated container, and web content filter exist (Appendix L.1, L.2). We then compare five versions of the loop using the same starting code, GPT-6.1-Sol at maximum reasoning effort, and 24 active hours per configuration (Appendix L.3, Table 9).

Table 9. Training and validation performance of the last optimizer accepted by each harness version.

Harness versionTraining Cper-molC_{\text{per-mol}} (%)Training rˉ\bar rValidation Cper-molC_{\text{per-mol}} (%)Validation rˉ\bar r
00-simple72.820.9980772.370.99010
01-validity74.071.0004976.520.99464
02-generalization-cost79.961.0006279.530.99744
03-generalization-energy81.851.0001784.251.00046
04-diagnostics-backlog82.461.0005582.981.00315

The 00-simple loop saved the most calls, but its optimizer recovered less energy than Sella on average. The 01-validity version improves energy recovery on train but still recovers less energy than Sella on validation. The 02-generalization-cost and 03-generalization-energy further improve on validation energy recovery, while 04-diagnostics-backlog (the version our main results were obtained with) improves agent's debugging abilities.

How repeatable is the search, and how does it compare with OpenEvolve? (Appendix M, Table 10)

We ran three independent searches with GPT-6.1-Sol Max under the same configuration. We also report one Claude Opus 5.5 Max run (we will provide 2 additional Opus-5.5 runs in camera ready version) and a 500-iteration OpenEvolve with GPT6.1-Sol Max reasoning effort search from the same Sella starting code.

RunTraining Cper-molC_{\text{per-mol}} (%)Training rˉ\bar rValidation Cper-molC_{\text{per-mol}} (%)Validation rˉ\bar rHoursUSD/h
GPT-6.1-Sol R173.531.00143774.081.00387766.005.86
GPT-6.1-Sol R283.301.00233585.801.00898566.005.37
GPT-6.1-Sol R381.481.00428882.561.00752266.006.39
Opus 5.5 R166.191.00259766.421.00641456.1916.96
OpenEvolve96.991.00000298.030.99882459.1624.28

All three GPT-6.1-Sol searches reduced validation force calls while exceeding Sella’s mean energy recovery. Savings ranged from 14.20% to 25.92%, with a median of 17.44%. OpenEvolve’s selected optimizer saved 1.97% of validation force calls and fell below the energy-recovery threshold. OpenEvolve selected candidates using training scores alone. Our loop used validation feedback during selection and gave the agent tools to inspect failures. Additionally, we provide per hour cost in API credits (OpenRouter credits for OpenEvolve) and observe that OpenEvolve costs approximately 4 times more than our autoresearch loop while delivering substantially less imrpovements.

Thank you again for your thoughtful reviews. With the discussion period closing today, we’d appreciate hearing whether the revisions summarized in our General Comment address your concerns or whether any points still need clarification.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.