acceptodds
Under review as a conference paper at ICLR 2027

Learning to Design SR Solvers: Harness Optimization for Symbolic Regression

Abstract

Symbolic regression (SR) seeks explicit equations that explain observed data, and recent methods increasingly employ large language models (LLMs) and agents for equation discovery. However, these systems typically rely on manually designed workflows and prompts. We argue that the SR agent harness should itself be an optimization target: its design shapes search priors and how numerical feedback guides subsequent equation refinement. We therefore propose HarnessSR, a framework for jointly optimizing prompts and workflows to construct reusable SR agents. During both harness optimization and agent execution, HarnessSR supports either PySR or an LLM-based equation generator through tool-specific interfaces, allowing the framework to optimize search guidance for different equation-discovery mechanisms. HarnessSR first optimizes role prompts through equation-discovery feedback in basic workflows, then jointly searches prompt assignments and workflows with brief adaptation to align the two. Finally, it refines role instructions through reflection on residual patterns, candidate equation structures, and search outcomes, improving how these signals guide subsequent equation search. On LSR-Synth in LLM-SRBench, HarnessSR achieves a mean ID of 61.33% across three backbones, a 45.0% relative improvement over the next-best baseline.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
symbolic regression

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_GnKDOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

HarnessSR optimizes reusable symbolic-regression agent harnesses while keeping language-model weights and the underlying SR algorithms fixed. It first optimizes role instructions and demonstrations, then searches workflow configurations and prompt assignments with brief adaptation, and finally refines instructions using SR-specific diagnostics of execution traces. The main configuration uses LLM roles to guide PySR. On 129 LSR-Synth tasks, the paper reports a mean ID Acc0.01 of 61.33% across three backbones, a 45.0% relative improvement over the next-best baseline, together with shorter per-task solving times. The appendices also evaluate OOD inputs, symbolic accuracy, and alternative SR backends.

Critical strengths

The method improves numerical accuracy across all three evaluated backbones. For Qwen3.6, ID Acc0.01 rises from 46.51% for SR-Scientist to 65.37%, while OOD Acc0.01 rises from 44.19% to 61.24% (Table 3 and Appendix F). These are meaningful gains, and the OOD results provide useful evidence that the benefit extends beyond the observed input range.

Critical weaknesses
  1. Limited evidence of methodological novelty. The additional contribution over existing prompt and workflow optimization methods, such as MASS and GEPA, remains unclear. Direct comparisons under comparable budgets are needed.

  2. Unmatched search budgets. HarnessSR uses 13 PySR calls per task, while the independent PySR baseline uses only one. Matched-budget comparisons and controlled ablations are needed to separate the benefit of optimized guidance from additional search effort.

  3. Insufficient reproducibility and statistical details. The optimization score Q, diagnostic/reflection models and prompts, and repetition protocol need clearer specification. Variability across runs should also be reported.

  4. Figure readability. Text in Figure 1 and several other figures and tables is too small, making them difficult to read. Please increase font sizes and improve legibility.

Questions for authors

See the weaknesses above.

4 comments
AuthorsOpenReview ↗

Author Response to Reviewer GnKD (1/3)

We thank the reviewer for the constructive feedback and for recognizing the consistent accuracy gains across backbones and on OOD inputs. Below, we address your concerns about methodological contributions, budget-matched comparisons, reproducibility, and figure readability point by point.


📝 W1 Methodological Novelty and Budget-Matched Comparisons

We thank the reviewer for this valuable feedback. We explain HarnessSR's four methodological contributions from two perspectives and present additional experiments to evaluate their effectiveness.

1. Contributions to general harness optimization. MASS, GEPA, and HarnessSR all optimize the harness outside the LLM, but use different mechanisms. MASS optimizes prompts and workflows separately, while GEPA uses reflection to optimize prompts only. We consider prompts and workflows strongly coupled, so optimizing either separately does not guarantee an optimal combination. Therefore, HarnessSR proposes:

(Innovation 1) Joint optimization of role prompts and workflows. We define a joint search space for role prompts and common workflows and perform an initial joint optimization of both.
(Innovation 2) Brief prompt adaptation further reduces the mismatch between locally optimized prompts and newly selected workflows.

2. Contributions specific to SR. MASS and GEPA were not designed for SR and do not include SR-specific mechanisms. Beyond defining roles and workflow topologies suited to SR, HarnessSR introduces two further designs:

(Innovation 3) Bidirectional adaptation between SR search guidance and backend feedback. Dedicated interfaces for PySR and an LLM-based equation generator translate structural hypotheses and search plans into executable backend guidance, and feed candidate equations, residuals, and search results back into harness optimization.
(Innovation 4) Instruction optimization driven by SR diagnostics. We build on and adapt GEPA's trajectory reflection mechanism by adding structured SR diagnostics and targeted role optimization. These diagnostics examine evidence-to-action alignment, structural exploration, constraint justification, and expression quality, identify the roles needing improvement, and guide revisions to their instructions. This turns solving experience into search guidance that can be reused across tasks.

Compare HarnessSR with MASS-SR and GEPA-SR. To directly evaluate the combined effect of these designs, we added the comparisons in Table R3.1. We adapted MASS and GEPA to SR using the same backend interfaces introduced in Innovation 3, and all three methods use Qwen3.6-35B-A3B as the backbone. Because the methods differ substantially, it is difficult to match exactly the number of solver executions used for optimization feedback. We therefore compare them under optimization budgets that are as close as possible. Both MASS-SR and GEPA-SR use the SR role set defined in our paper. Since GEPA-SR optimizes prompts only, it uses HarnessSR's final workflow shown in Figure 7 of the submitted manuscript.

Under comparable budgets of complete solver executions, HarnessSR exceeds MASS-SR and GEPA-SR in ID [email protected] by 16.53 and 12.92 percentage points, in OOD [email protected] by 14.47 and 13.18 percentage points, and in SA by 2.85 and 4.14 percentage points, respectively. HarnessSR also uses fewer model requests, input tokens, and output tokens than MASS-SR. Compared with GEPA-SR, it uses more model requests and input tokens but fewer output tokens. Overall, HarnessSR achieves better equation recovery under comparable solver-execution budgets, demonstrating the effectiveness of its optimization designs for SR.

We will include these comparisons in the revised manuscript and clarify HarnessSR's methodological contributions.

Table R3.1. Performance comparison of HarnessSR, MASS-SR, and GEPA-SR. MASS-SR and GEPA-SR are partial adaptations of MASS and GEPA to SR.

MethodOptimization budget (solver executions during training)OOD [email protected] (%) ↑SA (%) ↑Model requestsInput tokensOutput tokens
HarnessSR4,72861.2410.3454,701172,675,18716,804,235
MASS-SR4,80046.777.4959,112175,218,69020,877,578
GEPA-SR4,77748.066.2038,408139,788,69624,015,332
AuthorsOpenReview ↗

Author Response to Reviewer GnKD (2/3)

📝 W2 Comparison Under Matched Search Budgets (PySR & LLM generator)

We thank the reviewer for this constructive suggestion. As backend usage is optimized within the harness, we initially overlooked matching backend search budgets. We therefore added controlled experiments to isolate the effect of search effort.

1. PySR backend. Table R3.2 compares HarnessSR with two budget-matched controls: (1) PySR×13, which performs 13 independent searches of 125 iterations each and pools their candidates; and (2) PySR Long, which performs one continuous search of 1,625 iterations. Both match HarnessSR's total PySR iteration budget. Increasing PySR search effort yields only modest gains, and HarnessSR still maintains a substantial advantage under the same PySR budget.

2. LLM equation-generator backend. We conducted analogous comparisons with two budget-matched controls (Table R3.3). HarnessSR makes 13 backend calls, each with two rounds of eight generation requests. LLM Generator×13 performs 13 independent runs with the same per-run configuration and pools candidates, while LLM Generator Long performs 26 continuous rounds, each conditioned on the best equation found so far. HarnessSR outperforms LLM Generator×13 by 16.02 and 16.80 percentage points on ID and OOD, respectively.

Overall, Tables R3.2 and R3.3 demonstrate that HarnessSR's gains persist under matched backend search budgets, highlighting the benefits of agent-guided search coordination rather than additional search effort.

Table R3.2. Comparison under the same PySR search budget (LSR-Synth, 129 tasks).

MethodPySR backend calls × iterations per callID [email protected] (%) ↑OOD [email protected] (%) ↑SA (%) ↑
PySR1*12532.8231.785.17
PySR×13 (independent restarts with candidate pooling)13 × 125 iterations40.8337.217.24
PySR Long (continuous search)1 × 1,625 iterations37.9834.636.72
HarnessSR / Qwen3.6-35B-A3B13 × 125 iterations65.3761.2410.34
HarnessSR / Qwen3-Coder-30B-A3B-Instruct13 × 125 iterations57.3655.0412.66
HarnessSR / GPT-OSS-20B13 × 125 iterations61.2459.6911.63

Table R3.3. Comparison under the same LLM equation-generator search budget (LSR-Synth, 129 tasks). LLM backbone: Qwen3.6-35B-A3B.

MethodLLM backend calls × rounds per callID [email protected] (%)OOD [email protected] (%)SA (%) ↑
LLM Generator×1313 × 243.4141.347.75
LLM Generator Long1 × 2642.1237.475.43
HarnessSR13 × 259.4358.1413.44
AuthorsOpenReview ↗

Author Response to Reviewer GnKD (3/3)

📝 W3 Clarification on Reproducibility and Statistical Details

We thank the reviewer for this valuable suggestion and apologize for the lack of clarity. We clarify the optimization objective, the diagnosis and reflection mechanisms, and the repetition protocol below, and provide additional statistical results. We will explain these details more clearly in the revised manuscript.

1. Optimization objective QQ. For each development task qq, we use the negative average of the final equation's normalized mean squared error (NMSE) on the held-out ID and OOD data as the quality score:

ℓq=NMSEq,ID+NMSEq,OOD2,Q(S(q),q)=−ℓq\ell_{q} = \frac{NMSE_{q,ID}+NMSE_{q,OOD}}{2}, \quad Q(S(q),q)=-\ell_{q}

All three optimization stages use this quality objective to evaluate and optimize the harness.

2. Diagnosis and reflection. We apologize for omitting further details about the reflection model from the submitted manuscript. In HarnessSR, diagnosis and reflection use the same LLM as the solver backbone, rather than a different model. The diagnostic prompt instructs the model to perform four SR-specific checks on failed execution traces, identify problems, provide supporting evidence, and specify the affected roles. Based on this diagnostic evidence, the reflection prompt asks the model to analyze and revise role instructions through four steps: Observed, Supported By, Instruction at Fault, and Change. The diagnostic and reflection prompts have been added to Appendices G.3 and G.4 of the revised manuscript, respectively.

3. Repetition protocol. We apologize for not clarifying that the main experiments in the original submission were conducted with three random seeds. We have added the mean ± standard deviation (SD) of the main results in Appendix J (Table 17) of the revised manuscript.

4. Additional experiments on performance variability. We added experiments to further assess variability in harness optimization and inference.

(1) Variability in harness optimization. For each LLM backbone, we ran the full harness optimization procedure independently three times on the same training data. Each resulting harness was then evaluated on the same test set with three random seeds. Tables R3.4 and R3.5 report the mean ± SD. Comparing the three final harnesses, we found that two used workflow (1, 3, 3, 0, 1), with three Diagnoser–Repairer rounds and no cross-branch discussion. The third, which performed slightly worse, used workflow (1, 3, 2, 1, 1), with two Diagnoser–Repairer rounds and cross-branch discussion enabled. Although the resulting harnesses are not identical, their final performance differs only modestly, indicating that optimization randomness had a limited effect on performance in these runs.
(2) Variability in inference with a fixed harness. Tables R3.4 and R3.5 report the mean ± SD over three seeds for each final harness, showing relatively stable performance when the harness is held fixed. The main experimental results in the updated manuscript also report mean ± SD over three seeds.

Table R3.4. Performance of three independently optimized harnesses, each evaluated with three random seeds. LLM backbone: Qwen3.6-35B-A3B.

HarnessID [email protected] (%) ↑OOD [email protected] (%) ↑SA (%) ↑
165.37 ± 1.9561.24 ± 2.0510.34 ± 0.90
262.27 ± 2.2460.47 ± 1.559.82 ± 1.18
364.34 ± 1.3460.21 ± 2.4911.11 ± 0.45
Across three optimization runs63.99 ± 1.5860.64 ± 0.5410.42 ± 0.65

Table R3.5. Performance of three independently optimized harnesses, each evaluated with three random seeds. LLM backbone: GPT-OSS-20B.

HarnessID [email protected] (%) ↑OOD [email protected] (%) ↑SA (%) ↑
161.24 ± 1.5559.69 ± 2.3311.63 ± 0.78
258.91 ± 0.7857.11 ± 1.1810.59 ± 0.45
362.02 ± 3.3858.66 ± 2.7211.89 ± 0.45
Across three optimization runs60.72 ± 1.6258.49 ± 1.3011.37 ± 0.69

📝 W4 Figure and Table Readability

We apologize for the difficulty in reading the figures and tables. We will increase the font sizes in the revised manuscript to improve overall readability.

Dear Reviewer,

Thank you again for your thoughtful review. We have provided detailed responses to your comments and added experiments to address your concerns. As the discussion deadline is approaching, we would greatly appreciate any further feedback you may have, particularly on whether our responses and new results address your main concerns. We would be happy to clarify any remaining questions.

Thank you for your time and consideration.

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

This paper proposes HarnessSR, a framework that improves symbolic regression performance by automating the prompts and workflows of large language model agents. Unlike methods that directly optimize symbolic regression search algorithms or language model parameters, HarnessSR treats the agent’s role instructions, examples, and workflow structure as optimization objects.

Specifically, the method consists of three stages: first, it optimizes prompts for different roles in a simple workflow; second, it uses a Tree-structured Parzen Estimator (TPE) to jointly search workflow configurations and prompt assignments, and adapts candidate configurations; finally, it uses residuals, candidate expression structures, and search results from symbolic regression execution traces to generate diagnostic feedback and further improve role instructions.

Experiments are conducted on the LSR-Synth dataset of LLM-SRBench, covering 129 scientific symbolic regression tasks. Results show that HarnessSR outperforms the main baseline methods on three different base language models and significantly reduces the number of inference calls relative to other LLM agent methods.

Critical strengths

Existing LLM-based symbolic regression methods usually rely on manually designed prompts and workflows. This paper treats these components as optimization objects, providing a valuable research direction for improving the performance of symbolic regression agents. HarnessSR does not need to modify the underlying symbolic regression algorithm or the weights of the base language model, and can be combined with PySR, LLM-based expression generators, and hybrid configurations.

Critical weaknesses
  1. Lack of controlled experiments under the same search budget, so it cannot fully prove that the performance gain comes from improved search guidance strategy. PySR is an evolutionary search-based symbolic regression algorithm. Because the expression space is complex and contains many local suboptimal solutions, different random initializations and search trajectories may produce different candidate expressions. Even without changing the underlying search strategy, running independently multiple times and selecting the best candidate may improve the quality of the final expression. However, the final optimized workflow in this paper contains 13 PySR backend calls, while the standalone PySR baseline performs only one search. Therefore, the current performance improvement may partly come from additional search and candidate expression selection, and cannot fully prove that the optimized prompts and workflow effectively improve the search ability of symbolic regression. The authors are suggested to add a controlled experiment under the same search budget, that is, independently run PySR 13 times and aggregate candidate expressions for selection, and compare the numerical prediction accuracy and symbolic equivalence accuracy of the two, so as to clarify the actual contribution of the search guidance strategy.

  2. The relationship between the algorithm optimization objective and the improvement in structure recovery rate is unclear. Section 3.1 of this paper takes maximizing the average final expression quality on training tasks as the optimization objective, but does not clearly define the specific quality score. In addition, improving expression quality is not equivalent to improving the structural recovery rate of the true expression; increasing the number of searches or expanding the search range may also improve the final expression quality. Why can the inconsistency between the optimization objective and the actual objective improve the recovery rate? Appendix B shows that HarnessSR’s symbolic equivalence accuracy (SA) increases from 5.17% for standalone PySR to 10.34%, but the paper lacks an analysis of the reason for this improvement. Appendix E.1 adopts a strategy of preferring simple expressions when fitting errors are close. Therefore, is the improvement in structure recovery rate partly due to a preference for simple expressions, rather than more effective structural diagnosis and search correction? When the true target expression is relatively complex, will this strategy limit its recovery ability? In addition, the improvement in recovery rate may also come from multiple searches and candidate expression selection, and does not necessarily mean that the search ability has been substantially improved. Therefore, the authors are suggested to clarify the definition of the expression quality score, and further explain why the method can improve symbolic equivalence accuracy when the optimization objective does not directly target the structural recovery rate of the true expression. They should also analyze whether this improvement comes from the search guidance mechanism, the preference for expression simplicity, or the gains brought by additional search, so as to clarify the relationship between the algorithm optimization objective and the improvement in structure recovery rate.

Questions for authors

See Weaknesses Could the authors clarify the essential boundary between HarnessSR and A-SR regarding runtime adaptation and cross-run evolution, and between HarnessSR and Meta-Harness regarding whether its prompt/workflow search is merely a different search space from code search? Finally, relative to harness–model co-evolution, is the advantage of the fixed-weight route fairly quantified?

7 comments
AuthorsOpenReview ↗

Author Response to Reviewer oogc 1/3

We thank the reviewer for the constructive feedback and for recognizing the value of optimizing SR agent prompts and workflows while keeping model weights and underlying search algorithms unchanged. Below, we address your concerns about budget-matched comparisons, symbolic recovery, and distinctions from related work point by point.


📝 W1 Comparison Under Matched Search Budgets (PySR & LLM generator)

We thank the reviewer for this constructive suggestion. As backend usage is optimized within the harness, we initially overlooked matching backend search budgets. We therefore added controlled experiments to isolate the effect of search effort.

1. PySR backend. Table R2.1 compares HarnessSR with two budget-matched controls: (1) PySR×13, which performs 13 independent searches of 125 iterations each and pools their candidates; and (2) PySR Long, which performs one continuous search of 1,625 iterations. Both match HarnessSR's total PySR iteration budget. Increasing PySR search effort yields only modest gains, and HarnessSR still maintains a substantial advantage under the same PySR budget.

2. LLM equation-generator backend. We conducted analogous comparisons with two budget-matched controls (Table R2.2). HarnessSR makes 13 backend calls, each with two rounds of eight generation requests. LLM Generator×13 performs 13 independent runs with the same per-run configuration and pools candidates, while LLM Generator Long performs 26 continuous rounds, each conditioned on the best equation found so far. HarnessSR outperforms LLM Generator×13 by 16.02 and 16.80 percentage points on ID and OOD, respectively.

Overall, Tables R2.1 and R2.2 demonstrate that HarnessSR's gains persist under matched backend search budgets, highlighting the benefits of agent-guided search coordination rather than additional search effort.

Table R2.1. Comparison under the same PySR search budget (LSR-Synth, 129 tasks).

MethodPySR backend calls × iterations per callID [email protected] (%) ↑OOD [email protected] (%) ↑SA (%) ↑
PySR1*12532.8231.785.17
PySR×13 (independent restarts with candidate pooling)13 × 125 iterations40.8337.217.24
PySR Long (continuous search)1 × 1,625 iterations37.9834.636.72
HarnessSR / Qwen3.6-35B-A3B13 × 125 iterations65.3761.2410.34
HarnessSR / Qwen3-Coder-30B-A3B-Instruct13 × 125 iterations57.3655.0412.66
HarnessSR / GPT-OSS-20B13 × 125 iterations61.2459.6911.63

Table R2.2. Comparison under the same LLM equation-generator search budget (LSR-Synth, 129 tasks). LLM backbone: Qwen3.6-35B-A3B.

MethodLLM backend calls × rounds per callID [email protected] (%)OOD [email protected] (%)SA (%) ↑
LLM Generator×1313 × 243.4141.347.75
LLM Generator Long1 × 2642.1237.475.43
HarnessSR13 × 259.4358.1413.44
AuthorsOpenReview ↗

Author Response to Reviewer oogc 2/3

📝 W2 Clarification on the Optimization Objective Q, Symbolic Accuracy (SA), and the Selector

We thank the reviewer for this suggestion and clarify three points: (1) the quality score definition, (2) numerical optimization versus SA, and (3) the sources of SA improvement.

1. Definition of the quality score. We optimize the harness using numerical accuracy. For each development task qq, we define the loss as the average of the final equation's NMSE on the held-out ID and OOD data. Thus, maximizing the average task quality score JD(H)J_{\mathcal D}(\mathcal H) in Section 3.1 of the submitted manuscript is equivalent to minimizing the mean ID/OOD NMSE over development tasks:

ℓq=NMSEq,ID+NMSEq,OOD2,Q(S(q),q)=−ℓq\ell_{q} = \frac{NMSE_{q,ID}+NMSE_{q,OOD}}{2}, \quad Q(S(q),q)=-\ell_{q}

2. Numerical optimization and SA. We fully agree that improved numerical accuracy does not imply improved symbolic accuracy (SA). Most SR methods evaluate candidates using continuous numerical accuracy during search and optimization because it provides smoother feedback than discrete symbolic accuracy [1,2,3,4,5]. HarnessSR follows the same approach. This mismatch remains a challenge in SR, leaving considerable room for improvement in SA across existing methods. Nevertheless, HarnessSR outperforms the baselines on SA.

3. Reasons for the improvement in SA. We added SA results for the three optimization stages in Table R2.3. Both the joint optimization in Stage 2 and the reflection-guided optimization in Stage 3 improve SA. Although both stages still use a numerical objective, we believe that joint prompt–workflow optimization and reflection-guided refinement help adjust the search direction using information about variables, candidate expressions, and residuals. This increases the chances of discovering functional structures consistent with the data. Improvements in numerical generalization can therefore indirectly support symbolic recovery.

We also examined the effect of the simplicity constraint on SA. Since simplicity is an important criterion in SR, we introduced a mild simplicity constraint following prior work [3,5,6]. We agree with the reviewer's concern that this constraint could hurt SA. We therefore added experiments to examine its effects on SA and Acc at different levels of ground-truth equation complexity, as shown in Table R2.4. Each cell reports results with / without the simplicity constraint.

The results show small changes in Acc and particularly small changes in SA when the simplicity constraint is applied. On this dataset, the preference for simplicity mainly affects the final equations' accuracy–complexity trade-off, rather than being the main source of HarnessSR's structural recovery performance.

Table R2.3. Symbolic accuracy after the three optimization stages (Qwen3.6-35B-A3B, LSR-Synth).

StageID [email protected] (%) ↑SA (%) ↑ΔSA (pp) ↑
After Stage 127.912.58—
After Stage 243.416.72+4.13
After Stage 3 (Full)65.3710.34+3.62
ΔSA denotes the change in SA relative to the preceding stage, measured in percentage points.

Table R2.4. Results stratified by reference equation complexity.

Reference equation complexityID [email protected] (%) ↑OOD [email protected] (%) ↑SA (%) ↑Complexity ↓
Simple (≤19)79.59 / 80.9574.83 / 75.5114.97 / 15.6520.52 / 26.16
Medium (20–25)68.47 / 71.1761.26 / 62.167.21 / 8.1126.22 / 29.77
Complex (≥26)46.51 / 48.0645.74 / 48.847.75 / 8.5323.35 / 25.85
Each cell reports results with/without the simplicity constraint.
AuthorsOpenReview ↗

Author Response to Reviewer oogc 3/3

📝 Q1 The Boundary: HarnessSR vs. A-SR and HarnessSR vs. Meta-Harness

We thank the reviewer for this question.

1. HarnessSR vs. A-SR. The key differences lie in where adaptation occurs and what is learned for cross-task reuse. A-SR dynamically adjusts role utility, selection, and memory routing based on feedback from individual solves. Its A-SR-LoRA variant further distills trajectories into role-specific LoRA modules. In contrast, HarnessSR optimizes role prompts and workflows using feedback across training tasks, then freezes the resulting harness for new tasks without updating LLM weights. Importantly, a frozen harness can still adapt its search plans at runtime. For example, the Synthesizer can integrate candidates and results across branches to generate new plans. Thus, A-SR primarily focuses on runtime coordination and optional model distillation, while HarnessSR learns reusable role instructions and workflows through outer-loop optimization.

As an initial attempt at harness optimization for SR, HarnessSR uses relatively simple roles. If a role were to select tools or subroles autonomously, for example for data analysis or equation search, its prompt, degree of parallelism, number of repetitions, and other coordination settings could also be optimized within the harness. We hope HarnessSR can inform further work on harness optimization for SR.

2. HarnessSR vs. Meta-Harness. Both optimize the harness externally, but differ in their search spaces and domain-specific mechanisms. HarnessSR introduces four SR-oriented designs:
(1) Joint optimization of role prompts and workflows. We define a joint search space for SR role prompts and workflows, and perform an initial joint optimization using TPE.
(2) Brief prompt adaptation. We combine joint search with brief prompt adaptation to reduce the mismatch between locally optimized prompts and new workflows.
(3) Bidirectional adaptation between SR guidance and feedback. We design dedicated interfaces for PySR and an LLM-based equation generator. These interfaces translate structural hypotheses and search plans into executable backend guidance, and feed candidate equations, residuals, and search results back into harness optimization.
(4) SR diagnostic-driven instruction optimization. We examine evidence-to-action alignment, structural exploration, constraint justification, and expression quality, identify the relevant roles, and revise their prompts accordingly. This turns experience from solving training tasks into instructions that can be reused across tasks.


📝 Q2 Harness–model Co-evolution V.S. Fixed-weight Route

We thank the reviewer for this question. Under matched harness optimization budgets, harness–model co-evolution additionally requires model training and associated computational resources. One possible approach is to fine-tune the model on successful trajectories after harness optimization. However, limited computational resources and rebuttal time prevent a controlled comparison.

As an approximate reference, Table R2.5 compares HarnessSR (harness-only optimization) with SR-Scientist+RL (model-only optimization) in training data, execution counts, LLM requests, tokens, and hardware. Based on its published configuration and code, SR-Scientist+RL performs 15,360 single-iteration rollouts (60 steps × 32 tasks × 8 samples), compared with 4,728 complete solver executions for HarnessSR. Since actual LLM requests and token counts per RL rollout are unavailable, we estimate them from non-RL SR-Scientist inference traces using the same Qwen3-Coder backbone.

HarnessSR's optimization cost is much lower than the estimated model post-training cost of SR-Scientist+RL. Since harness–model co-evolution combines harness and model optimization, we therefore expect the fixed-weight approach to be considerably less costly than co-evolution.

Table R2.5. One-time optimization and post-training costs. LLM backbone: Qwen3-Coder-30B-A3B-Instruct.

MethodDevelopment tasksExecutions during optimization/post-trainingLLM requestsInput tokensOutput tokensOptimization hardware
HarnessSR (submitted version)1604,728 complete solver executions76,226214.8M23.6M2×H100
SR-Scientist+RL1,02415,360 single-iteration rollouts≈258,181 (estimated)≈1,402.5M (estimated)≈75.3M (estimated)32×H200

Estimation details: The measured inference traces average 16.81 LLM requests, 91,309 input tokens, and 4,900 output tokens per iteration. Multiplying these averages by 15,360 rollouts gives an estimated post-training consumption of approximately 258,000 LLM requests, 1.40 billion input tokens, and 75.3 million output tokens.

Reviewer_6izGOpenReview ↗
Rating
4(+1)
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

The paper treats the executable harness of a symbolic-regression agent—role instructions, demonstrations, and workflow topology—as the object of optimization, while keeping the LLM weights and the underlying equation-search algorithm fixed. HarnessSR optimizes the harness in three stages: (i) role-level prompt optimization with TPE in simple workflows; (ii) joint search over workflow parameters and prompt assignments with brief adaptation; and (iii) refinement of role instructions using an external diagnostic model that reads SR execution traces (residual patterns, candidate structures, validity, and search outcomes). The framework supports PySR and an LLM-based equation generator through tool-specific adapters. On LLM-SRBench LSR-Synth (129 tasks) with three backbones, it reports mean [email protected] of 61.33% (65.37/57.36/61.24 per backbone) and a 45.0% relative improvement over the next-best baseline.

Critical strengths
  1. Well-posed and timely problem. The paper correctly identifies that the agent harness—not only the base model or the numerical search algorithm—shapes how scientific knowledge and numerical feedback are converted into equation proposals. Treating the harness as a design variable is a useful reframing, and the distinction from weight post-training is clearly motivated.

  2. Strong empirical results on a standardized benchmark. The gains on LSR-Synth are substantial and consistent across three backbones. The stage ablation is persuasive about the contribution of each component: 5.43 → 27.91 → 43.41 → 65.37 [email protected]. The comparison with weight post-training under the same Qwen3-Coder base is a particularly useful control.

  3. Practical efficiency. The per-task deployment cost (3.66/4.03/3.29 minutes) is competitive with the baselines, and the 80→640 training-task scaling analysis suggests the harness improves with more optimization data. The tool-agnostic adapter design makes the approach reusable.

  4. Good ablation discipline. The paper separates workflow search, prompt assignment, and SR-guided reflection, and reports both accuracy and LLM-call cost. The comparison between direct trace revision and SR-diagnostic revision (Table 6) is informative.

  5. Clear exposition. The three-stage procedure, the workflow parameterization, and the execution pipeline are easy to follow, and the appendix documents the optimization budget and the per-stage cost.

Critical weaknesses
  1. Single benchmark family; limited evidence of generality. All main results are on LSR-Synth, a synthetic benchmark. The OOD and SA evaluations remain within the same benchmark suite, and the paper does not include a real scientific dataset or a case study of a discovered equation. Since the stated motivation is scientific discovery, this is a significant gap. At minimum, the authors should evaluate on a small set of real-world SRBench tasks or provide a qualitative case study.

  2. No variance or uncertainty estimates. The harness optimization is stochastic (TPE, LLM sampling, prompt selection), yet the paper reports single optimized harnesses and point estimates without standard deviations, confidence intervals, or repeated optimization seeds. Several claimed differences—especially the 2–3 point gaps between harness variants—could be within run-to-run noise. The authors should report mean ± std over at least three independent harness-optimization runs and paired tests on the 129 tasks.

  3. The one-time optimization cost is not included in the comparison. The paper reports 4,728 solver executions on 160 tasks for Stage 3 and 2×H100 for optimization, but the main table compares per-task deployment time only. An amortized comparison, or an iso-budget comparison in total LLM calls and tool calls (including harness optimization), is needed to substantiate the efficiency claim.

  4. The harness space is hand-designed. The optimization searches five discrete workflow variables, a fixed set of five roles, and prompt libraries; the roles, tool interfaces, and adapter semantics remain human-designed. The claim that the harness is “automatically optimized” is therefore true only within a restricted, hand-specified space. The paper should be explicit about this scope and discuss how sensitive the results are to the role/tool vocabulary.

  5. The reflection model is a potential confound. Stage 3 relies on an “external diagnostic LLM” (cited as Agrawal et al., 2026) whose identity, scale, and prompting are not specified in the main text. If this model differs from the solver backbone, the gains may reflect the diagnostic model’s capability rather than the harness optimization procedure. Please specify the model and provide a control where the same diagnostics are produced by a weaker/stronger model.

  6. OOD and SA gains are much smaller than ID gains. [email protected] drops from 65.37 to 61.24 on Qwen3.6, and SA remains around 10–13%. The paper’s strongest claims are about ID accuracy; the evidence that the optimized harness improves genuine scientific generalization is weaker. The SA results in particular need more discussion.

  7. The learned-SR landscape is under-discussed. The paper positions harness optimization against weight post-training, but does not discuss non-LLM learned search methods such as neural-guided dynamic symbolic networks (e.g., DySymNet, Li et al., ICML 2024). These are not direct competitors—they learn a proposal/search network rather than an agent harness—but they are a useful reference point for the design space between a fixed solver and an LLM agent. A short positioning paragraph would clarify whether harness optimization is intended as a complement or an alternative. A head-to-head comparison is optional; if added, it should use the same evaluation budget and adapter interface.

  8. Possible benchmark contamination is not discussed. The LLM backbones may have seen LLM-SRBench tasks or the underlying equations during pretraining. A contamination check (e.g., prompting the backbone to recall equations, or evaluating on newly held-out equations) would strengthen the claims.

Questions for authors
  1. How many independent harness-optimization runs were performed, and what are the mean ± std and paired-test results for the main comparisons?
  2. What is the total compute/LLM-call budget including the one-time harness optimization, and how does HarnessSR compare at equal total budget?
  3. Which model produces the Stage 3 diagnostics, and is it independent of the solver backbone? How much of the gain survives if the diagnostics come from a weaker model?
  4. Can you factorize the gains from prompt content, workflow topology, and reflection? The current stage ablation changes several factors at once.
  5. Do harnesses optimized on one backbone transfer to another backbone? This would be a strong test of whether the method learns reusable harness design principles rather than backbone-specific prompts.
  6. Can you add a real-world or real-scientific evaluation, or at least a case study of an equation discovered by the optimized harness?
  7. Did you check for benchmark contamination in the LLM backbones?
7 comments
AuthorsOpenReview ↗

Author Response to Reviewer 6izG 1/4

We thank the reviewer for the positive score, thorough and constructive review, and for recognizing the value of harness optimization, the consistent empirical gains, and the informative ablations. We address each of your concerns and questions in detail below.


📝 W1 / Q6 Additional Experiments on a New Benchmark and Real-World Case Studies

We provide clarifications and additional experiments in three parts.

1. Why LSR-Synth rather than SRBench. We chose LSR-Synth over SRBench due to the latter's known limitations in evaluating LLM-based methods [1]. LSR-Synth is widely used and designed to resist LLM memorization. We follow SR-Scientist's evaluation protocol on LSR-Synth to ensure fair comparison.

2. Additional Experiments on Anonymized LSR-Transform. We additionally evaluated HarnessSR on anonymized LSR-Transform. Originally designed to transform common physical models into less familiar mathematical representations [1], LSR-Transform remains susceptible to memorization [2, 3]. Following [4], we anonymized it to mitigate this risk. As shown in Table R1.1, HarnessSR achieves 73.87% [email protected] and 51.35% symbolic accuracy (SA), outperforming SR-Scientist (55.86%, 43.24%) and PySR×13 (46.55%, 42.04%). Following Reviewers oogc and GnKD, PySR×13 performs 13 independent searches with pooled candidates, matching the number of SR backend calls. These results show HarnessSR's effectiveness beyond LSR-Synth.

3. Additional Experiments on Two Real-World Scientific Problems. We further evaluated HarnessSR on two real-world datasets: CALiSol-23 electrolyte conductivity [5] and wind-turbine wake flow fields [6]. As shown in Table R1.2, HarnessSR achieves lower ID and OOD NMSE than others on both tasks. Since neither dataset provides ground-truth equations, symbolic accuracy cannot be assessed. These results further support HarnessSR's generalization to real-world scientific problems.

Note: All HarnessSR results in Tab. R1.1, R1.2 use the same final optimized harness shown in Fig. 7 of the submitted text, without repeating the three-stage optimization.

Table R1.1. Performance on LSR-Transform-Anon (LLM: Qwen3.6-35B-A3B).

Method[email protected] (%) ↑SA (%) ↑
PySR×1346.5542.04
SR-Scientist55.8643.24
HarnessSR73.8751.35

Table R1.2. Performance on two scientific datasets.

TaskMethodID NMSE ↓OOD NMSE ↓
LiPF6 electrolyteHarnessSR0.029010.4635
LiPF6 electrolytePySR×130.061861.917
LiPF6 electrolyteSR-Scientist0.038140.6852
Wind-turbine wakeHarnessSR0.010530.02159
Wind-turbine wakePySR×130.056430.8132
Wind-turbine wakeSR-Scientist0.046080.07332

📝 W2 / Q1 Optimization Variability and Statistical Uncertainty

We thank the reviewer for this valuable suggestion. We address these two sources of variability separately.

1. Variability across harness optimization runs. For each LLM backbone, we conducted three independent harness optimizations on the same training data, then froze each resulting harness and evaluated it on the same test set with three random seeds. Tables R1.3 and R1.4 report the mean ± standard deviation (SD). Taking Qwen3.6-35B-A3B as an example, two optimized harnesses adopted workflow (1, 3, 3, 0, 1), with three Diagnoser–Repairer rounds and no discussion, while the third adopted (1, 3, 2, 1, 1), with two rounds and discussion, achieving slightly lower performance. Despite these workflow differences, performance varies only modestly, suggesting limited sensitivity to optimization randomness. We will include additional results in the final revised manuscript.

2. Variability during inference with a fixed harness. We clarify that the main results in the original text already average over three random seeds and apologize for not stating this explicitly. Tables R1.3 and R1.4 report the mean ± SD across three seeds for each fixed harness, showing relatively stable performance across repeated solves. These statistics are also provided in Appendix J (Tab.17) of the revised manuscript.

Table R1.3. Performance of three independently optimized harnesses (LLM: Qwen3.6-35B-A3B).

HarnessID [email protected] (%)OOD [email protected](%)SA (%)
165.37 ± 1.9561.24 ± 2.0510.34 ± 0.90
262.27 ± 2.2460.47 ± 1.559.82 ± 1.18
364.34 ± 1.3460.21 ± 2.4911.11 ± 0.45
Across three optimization runs63.99 ± 1.5860.64 ± 0.5410.42 ± 0.65

Table R1.4. Performance of three independently optimized harnesses (LLM: GPT-OSS-20B).

HarnessID [email protected] (%)OOD [email protected] (%)SA (%)
161.24 ± 1.5559.69 ± 2.3311.63 ± 0.78
258.91 ± 0.7857.11 ± 1.1810.59 ± 0.45
362.02 ± 3.3858.66 ± 2.7211.89 ± 0.45
Across three optimization runs60.72 ± 1.6258.49 ± 1.3011.37 ± 0.69
AuthorsOpenReview ↗

Author Response to Reviewer 6izG 2/4

📝 W3 / Q2 One-Time Optimization Cost

We thank the reviewer for this suggestion and understand the concern about optimization cost.

(1) Inference-time comparison. We follow prior work [7, 8, 9] to compare different types of SR methods fairly. Our main table reports only each method's mean inference time, measured from the start of solving a new task to the output of its final expression. This includes inference-stage LLM calls and SR backend search, but excludes model training for E2E and NeSymReS, model post-training for SR-Scientist+RL, and harness optimization for HarnessSR. This comparison is common and fair because training and optimization are performed once, whereas inference can be repeated indefinitely. As the number of inference tasks tends to infinity, the one-time cost amortized over each task approaches zero.

(2) One-time optimization cost comparison. We also consider it meaningful to compare one-time optimization and training costs separately and have added this comparison. Table R1.5 compares HarnessSR and SR-Scientist+RL in terms of training data size, execution counts, LLM requests, input and output tokens, and optimization hardware. Based on the training configuration in the SR-Scientist+RL paper and public code, its post-training involves 60 training steps × 32 tasks per step × 8 samples per task = 15,360 single-iteration rollouts. HarnessSR used 4,728 complete solver executions during optimization, substantially fewer than SR-Scientist+RL. Since SR-Scientist+RL does not report the actual LLM requests and token consumption per rollout, we estimate these quantities using inference traces from SR-Scientist without RL training, with the same Qwen3-Coder backbone. The comparison in Table R1.5 shows that HarnessSR's one-time optimization is less costly than SR-Scientist+RL's post-training.

Estimation details: The measured inference traces average 16.81 LLM requests, 91,309 input tokens, and 4,900 output tokens per iteration. Multiplying these averages by 15,360 rollouts gives an estimated post-training consumption of approximately 258,000 LLM requests, 1.40 billion input tokens, and 75.3 million output tokens.

Table R1.5. One-time optimization and post-training costs. LLM backbone: Qwen3-Coder-30B-A3B-Instruct.

MethodDevelopment tasksExecutions during optimization/post-trainingLLM requestsInput tokensOutput tokensOptimization hardware
HarnessSR (submitted version)1604,728 complete solver executions76,226214.8M23.6M2×H100
SR-Scientist+RL1,02415,360 single-iteration rollouts≈258,181 (estimated)≈1,402.5M (estimated)≈75.3M (estimated)32×H200

📝 W4 Manually Designed Search Space and Sensitivity Analysis

We thank the reviewer for this valuable suggestion. We will clarify that our harness optimization is constrained by the search budget, the fixed set of candidate roles, tool interfaces, and adapters.

We designed this search space around the characteristics of SR tasks. To examine how sensitive the results are to its design, we added the experiments in Table R1.6. Given the limited time and computational resources available during rebuttal, we compared the original search space with three alternatives: removing Analyst, removing both Diagnoser and Repairer, and removing Synthesizer. All other settings were kept unchanged in each case.

The results show that the choice of harness search space has a substantial effect on performance. All three removals reduced ID/OOD [email protected] and SA to varying degrees. Removing Diagnoser and Repairer had the largest effect: ID [email protected] fell from 65.37% to 43.41%, and OOD [email protected] fell from 61.24% to 44.70%. This highlights the importance of diagnosis and iterative repair for search performance and suggests that optimizing the harness search space itself is a promising direction. Table 13 of the submitted test also examines sensitivity to the choice of SR backend, further demonstrating its adaptability to different SR tools.

Table R1.6. Sensitivity to the harness search space.

Roles removed from the search spaceID [email protected] (%)OOD [email protected] (%)SA (%)Change in ID [email protected] (percentage points)
None (HarnessSR)65.37 ± 1.9561.24 ± 2.0510.34 ± 0.90Reference
Analyst57.62 ± 3.1353.75 ± 1.188.53 ± 0.78−7.75
Diagnoser + Repairer43.41 ± 1.5544.70 ± 1.955.17 ± 0.45−21.96
Synthesizer56.07 ± 1.7950.90 ± 2.726.20 ± 0.00−9.30
AuthorsOpenReview ↗

Author Response to Reviewer 6izG 3/4

📝 W5 / Q3 Clarification on the Reflection and Diagnosis Models

We thank the reviewer for the suggestion and question, and apologize for omitting the details of the reflection model from the submitted manuscript. In all experiments, the entire optimization procedure, including reflection, uses the same LLM as the solver backbone. This ensures a fair comparison with the baselines.

We cite Agrawal et al. (2026) for their general reflection mechanism. We designed the reflection-based prompt revision procedure in Stage 3 specifically for SR tasks. We apply four types of checks to failed SR trajectories, use the findings to identify which roles need improvement, and revise their prompts accordingly. The complete Stage 3 diagnostic and reflection prompts have been added to Appendices G.3 and G.4 of the revised manuscript, respectively.


📝 W6 Interpretation of OOD and SA Performance

We thank the reviewer for this valuable feedback.

1. OOD performance. OOD evaluation is more challenging than ID evaluation, as it tests the selected equation on unseen distributions. Most methods therefore perform slightly worse on OOD data. Nevertheless, HarnessSR outperforms the baselines on OOD evaluation (Table 14 of the submitted manuscript).

2. SA performance. Symbolic accuracy (SA) is a stricter metric that directly assesses symbolic correctness. However, most SR methods, including HarnessSR, optimize numerical accuracy during search [2, 3, 8, 9, 10, 11], as it provides smoother feedback than discrete SA. This mismatch leaves room for SA improvement across existing methods. Nevertheless, HarnessSR outperforms the baselines on SA. We will further discuss these results in the revised manuscript.


📝 W7 Positioning Relative to Non-LLM Learned SR

We thank the reviewer for this suggestion and agree that we should discuss non-LLM learned search methods, including neural-guided symbolic regression methods such as DySymNet.

For a given task, DySymNet uses reinforcement learning to guide the search over symbolic network structures and fits the network parameters to find an equation that explains the observed data. HarnessSR instead uses feedback from multiple training tasks to optimize role prompts and workflows, then freezes the resulting harness for reuse on new tasks while keeping LLM weights fixed throughout.

Thus, DySymNet improves equation search within a task, whereas HarnessSR optimizes search guidance and the organization of the search process for reuse across tasks. We therefore view harness optimization as a complement to existing SR methods rather than a replacement. We will discuss DySymNet in the related work and clarify this distinction and positioning.


📝 W8 / Q7 Experimental Evidence Against Benchmark Contamination

We thank the reviewer for this valuable feedback. Benchmark contamination is an important concern, so we carefully chose LSR-Synth from LLM-SRBench, a commonly used dataset that is resistant to LLM memorization [1, 2, 3]. Our main results show that all LLM-based methods, including HarnessSR, still have room for improvement across the three backbones, particularly in SA. If data leakage were present, both Acc and SA would approach saturation after only a few iterations.

A controlled comparison to assess contamination: To provide further evidence and given the limited rebuttal time, we focused on Qwen3.6-35B-A3B, the newest and largest backbone in our main experiments. Table R1.7 reports SR-Scientist's convergence across datasets. We chose SR-Scientist because its LLM directly generates and refines equations using analysis and evaluation tools, making memorization effects of LLM easier to isolate than in HarnessSR, which additionally relies on PySR's evolutionary search.

On LSR-Transform, SR-Scientist converges rapidly, reaching 55.86% ID [email protected] after 25 calls and 72.97% after 100 calls, with a final SA of 78.38%, suggesting memorization of classical problems. In contrast, performance on LSR-Synth improves gradually and remains far from saturation, suggesting lower contamination risk. The other two backbones in our main text, Qwen3-Coder-30B-A3B-Instruct and GPT-OSS-20B, perform worse than Qwen3.6-35B-A3B, so their contamination risk is also low. We will add the corresponding control experiments.

Furthermore, anonymized LSR-Transform shows no clear signs of data leakage, supporting the value of our additional experiments addressing W1.

Table R1.7. Benchmark contamination check: convergence with Qwen3.6-35B-A3B across datasets.

DatasetID [email protected] (25 calls)ID [email protected] (100 calls)ID [email protected] (400 calls)ID [email protected] (final)Final SA (%)
LSR-Transform55.8672.9787.3989.1978.38
LSR-Transform-Anon28.8342.3453.1555.8643.24
LSR-Synth14.7329.4644.9646.518.79

Comments on the paper

2 comments

Dear Reviewers,

We are currently in the discussion phase. The authors have responded to the comments, and I would greatly appreciate it if you could take a moment to check the rebuttal and leave any follow-up comments or questions. If the response has addressed your concerns, please also consider updating your review accordingly.

Your feedback is very valuable to both the authors and the decision process. Thank you again for your help.

Best regards, AC

AuthorsOpenReview ↗

Summary of the rebuttal

Dear Area Chair, Senior Area Chairs, and Reviewers,

We sincerely thank you for your careful assessment and constructive feedback. Your suggestions have helped us strengthen the empirical evidence and clarify HarnessSR’s contributions and evaluation protocol. Below, we summarize the additional experiments and clarifications provided during the discussion.

Main ConcernReviewersAdditional Results and Clarifications
Reproducibility and presentation6izG, oogc, GnKDClarified the numerical objective and that diagnosis/reflection use the solver backbone. Added full prompts and three-seed statistics; table and figure readability have been improved.
Matched backend search budgetsoogc, GnKDUnder matched backend budgets, HarnessSR achieves 65.37% vs. 40.83% ID accuracy against PySR×13 and 59.43% vs. 43.41% against LLM Generator×13.
Optimization and inference variability6izG, GnKDThree independent optimizations, each evaluated with three inference seeds, yield ID accuracy of 63.99 ± 1.58% for Qwen3.6 and 60.72 ± 1.62% for GPT-OSS-20B.
Prompt/workflow contributions and symbolic recovery6izG, oogcRandomizing workflows or prompts reduces ID accuracy from 65.37% to 52.71% or 45.74%. SA improves from 2.58% → 6.72% → 10.34% across optimization stages.
One-time optimization cost6izGReported 4,728 solver executions over 160 development tasks, alongside LLM/token consumption and hardware, separately from deployment costs and estimated RL post-training costs.
Generality and scientific applications6izGWithout re-optimization, the harness achieves 73.87% accuracy / 51.35% SA on anonymized LSR-Transform and lower ID/OOD NMSE than both baselines on electrolyte and wake-flow tasks.
Cross-backbone harness transfer6izGFive of six transfers achieve comparable or better ID/OOD accuracy. Transferring the Qwen3.6 harness to Qwen3-Coder improves ID accuracy from 57.36% to 67.18%.
Potential benchmark contamination6izGAnonymization reduces rapid equation recovery on LSR-Transform. Convergence checks support using LSR-Synth and anonymized LSR-Transform.
Optimization scope and positioning6izG, oogcClarified the specified search space, cross-task harness reuse, distinctions from A-SR and Meta-Harness, and complementarity with learned SR methods.
Comparisons with and distinctions from related methodsGnKD, oogcUnder comparable solver-execution budgets, HarnessSR exceeds MASS-SR/GEPA-SR by 16.53/12.92 percentage points in ID accuracy. We clarified distinctions from A-SR and Meta-Harness, and discussed compatibility with and the comparison scope relative to harness–model co-evolution.

We appreciate Reviewer 6izG’s follow-up confirming that most concerns have been addressed and increasing the score from 3 to 4. We will consolidate the additional experiments and clarifications in the revised manuscript and remain available to address any further questions.

Thank you again for your time and consideration.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.