acceptodds
Under review as a conference paper at ICLR 2027

Loop Reinforcement Learning for Scientific Discovery

Abstract

Large language models (LLMs) have demonstrated strong search capabilities when given well-defined problems and quantitative verifiers. Advances in both model capabilities and orchestration harnesses have accelerated progress in scientific discovery. However, LLMs tend to generate similar candidates, concentrating exploration within narrow regions of the search space and risking premature convergence that limits the breadth of discovery. Recent approaches seek to broaden exploration by orchestrating LLMs with evolutionary algorithms, using test-time scaling to maintain diversity and expand search coverage. Despite their effectiveness, these approaches rely on manually designed search rules, heuristics, or agentic workflows that may constrain exploration. To address this limitation, we propose Loop Reinforcement Learning (Loop RL) for scientific discovery. We first introduce a simple loop harness consisting only of sequential iteration, parallel execution, and fully asynchronous shared memory for exchanging information across parallel trajectories. Beyond these basic mechanisms, the harness leaves exploration strategies to the model, providing a minimal baseline with ample flexibility to explore and learn. We then develop Loop RL to optimize the discovery performance of this loop through credit assignment across iterations and parallel trajectories. Across scientific discovery tasks, Loop RL achieves faster search speed and higher-quality solutions, attaining state-of-the-art results on several tasks.

Primary area:
foundation or frontier models, including LLMs
Keywords:
loop discovery, scientific discovery, loop reinforcement learning

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_ZRRLOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper introduces two new mechanisms to improve scientific discovery of LLM: Simple Loop and LoopRL. Simple loop builds a collaborative working environment for multiple agents with the same policy. These agents share the same memory, but can propose different improvements. Different agents proceed independently. Approaches like reflection and knowledge consolidation are removed to grant agents more freedom for decision. LoopRL is a reinforcement learning algorithm with a carefully designed reward function. This reward function encourages the improvement relative to the current best, and also gives credit to the experience which insights the improvement. The paper claims that these two mechanisms can improve the diversity of exploration, and thus make the quality of discoveries higher.

Critical strengths
  1. Details. This paper provides a lot of details of how the proposed approaches work on different problems, and the observations on how the improvements are generated through the collaboration between multiple agents.
  2. Width of benchmarks. This work tests the performance of the proposed approaches on a wide range of test cases.
Critical weaknesses
  1. Clarity. The reviewer finds the paper really hard to follow and understand. For example, in section 4.2, the authors analyze how the improvements are generated by the collaboration between agents in a very lengthy manner. However, this section is named results, and is supposed to show some quantitative results.
  2. Lack of comparisons with other approaches. Despite a lengthy discussion of the improvements introduced by multiple agents, this work has no mention of what other approaches can achieve. Although in lines 251-252 the authors promise to compare with CORAL, later this comparison is missing. Without a fair benchmark with other approaches, one cannot assess the performance of the proposed methods.
  3. Contribution of reinforcement learning is not discussed. There are two approaches introduced: Simple Loop and LoopRL. Simple Loop introduces a collaborative environment, and seems to work without reinforcement learning. In the "results" section, it seems the contribution from multiple agents is sound, but RL is not essential. An ablation study with just Simple Loop and without LoopRL would help figure out how important the RL is in the proposed combination of methods.
Questions for authors

See critical weaknesses section. Overall, this work is not clearly presented, and not well-organized. The contributions cannot be assessed based on its current format.

2 comments
AuthorsOpenReview ↗

Response to Reviewer ZRRL [1/n]

We sincerely thank you for your careful and thorough review. We appreciate your recognition of our work, particularly the detailed explanations and the breadth of our benchmark evaluation. Below, we address your concerns point by point.


Weakness 1: Clarity and quantitative results

Thank you for pointing this out. Following the presentation style of AlphaEvolve [1], ShinkaEvolve [2], and CORAL [3], we provided detailed accounts of the search process and its progression. Section 4.2 was intended to illustrate how communication among agents can substantially improve discovery and to analyze how specific communication patterns affect discovery outcomes.

To complement this analysis with quantitative evidence, we provide results across all ten benchmarks below. Under the same resource budget, SimpleLoop outperforms CORAL on every task, while LoopRL further improves performance by enabling the model to learn search strategies specific to each problem.

TaskMetricCORALSimpleLoopLoopRL
Second autocorrelationAutocorrelation lower bound C₂ ↑0.944800.948330.95210
Third autocorrelationReference ratio ↑0.9987000.9991051.000250
Erdős minimum overlapReference ratio ↑1.0000801.0001251.000180
Heilbronn convex / 13Reference ratio ↑0.9998001.0000001.000010
Hexagon packing / 11Reference ratio ↑1.0011001.0014201.001650
Signal processingComposite score ↑0.63250.64180.747
Circle packing / 26Sum of radii ↑2.6354002.6358002.635983
Hadamard / 29Determinant ratio ↑0.9320000.9345000.935673
Bio / BPG0ID NMSE ↓2.00 × 10⁻¹³1.60 × 10⁻¹³1.30 × 10⁻¹³
Materials / MatSci0ID NMSE ↓1.00 × 10⁻¹⁴8.00 × 10⁻¹⁵6.40 × 10⁻¹⁵

[1] AlphaEvolve: A coding agent for scientific and algorithmic discovery

[2] ShinkaEvolve: Towards Open-Ended And Sample-Efficient Program Evolution

[3] CORAL: Towards Autonomous Multi-Agent Evolution for Open-Ended Discovery


Weakness 2: Comparison with CORAL

In addition to the comparisons with CORAL presented in our response to Weakness 1, we provide further results on the two challenging benchmarks used for stress testing in CORAL: Anthropic Kernel Engineering and Frontier-CS Polyominoes Packing. All methods are evaluated under the same benchmark settings, computational resources, and wall-clock time limits.

MethodModelAnthropic Kernel Engineering ↓Frontier-CS Polyominoes Packing ↑
CORALClaude Opus 4.6110389.4%
SimpleLoopClaude Opus 4.687295.4%
CORALQwen 3.8 27B117387.3%
SimpleLoopQwen 3.8 27B107293.2%
LoopRLQwen 3.8 27B104294.9%

With the same underlying model, SimpleLoop achieves better final performance and faster search than CORAL. With the smaller open-source model, Qwen 3.8 27B, SimpleLoop also clearly outperforms CORAL, while LoopRL further improves performance.

AuthorsOpenReview ↗

Response to Reviewer ZRRL [2/n]

Weakness 3: Contribution of reinforcement learning

Thank you for your question. Regarding the impact of reinforcement learning on scientific discovery, Section 4.3 discusses the effectiveness of training, showing that LoopRL improves communication across branches, cooperation strategies, and search patterns.

In Section 4.4 and Figure 5, we analyze how different credit assignment strategies affect RL training. We find that LoopRL’s credit assignment strategy outperforms that used in conventional multi-turn RL, achieving faster search, measured by the number of iterations, and higher final performance under the same experimental settings.

Additionally, our responses to Weaknesses 1 and 2 provide detailed comparisons between SimpleLoop and LoopRL on ten common discovery tasks and two challenging stress-test tasks. LoopRL consistently outperforms SimpleLoop, quantitatively demonstrating the effectiveness of RL training in enabling the model to autonomously learn search strategies for scientific discovery.


We sincerely appreciate your time in reviewing our paper and your valuable suggestions for improving it. We hope that our detailed responses address your concerns. If you have any remaining concerns, we will promptly provide further details. Thank you.

Reviewer_h5wZOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The paper studies how RL can improve LLM-based discovery while reducing reliance on manually designed search heuristics that may constrain exploration and introduce biais. The authors train agents to refine solutions over mm sequential attempts in each of nn parallel loops, exchanging experience through shared memory. Their methods rewards not only the immediate improvement but also later improvements deeper in the loop.

Critical strengths
  • The idea is interesting and addresses an important challenge in LLM-based evolutionary search.
  • Another merit is : the flexible shared memory; agents can share the results, failed attempts, and useful approaches, then choose what information to use for their current search attempt.
  • The framework is relatively simple, and the reward and return definitions are explicit.
  • Method seems to improve the final score the two dimensional comparison was conducted.
Critical weaknesses
  • Figure 2 was quite understandable but not quite formative
  • The method is not compared with fun search alpha evolve type of frameworks
  • The experimentations is quite scarce for example : no diversity in solutions or search strategies was measured.
Questions for authors
  • Add experiments mentioned in the weaknesses.
2 comments
AuthorsOpenReview ↗

Response to Reviewer h5wZ [1/n]

We sincerely thank you for your recognition of our work and for taking the time to provide constructive suggestions. Below, we address your concerns point by point.


Weakness 1: Informativeness of Figure 2

Thank you for this feedback. Figure 2 is intended to visualize communication and collaboration across branches during the SimpleLoop search process and their relationship with the performance of individual branches. To make the figure more informative, we have revised it to explicitly represent communication between branches using directed arrows. These arrows clarify the direction of information flow and make the communication patterns and associated performance changes more explicit.


Weakness 2: Comparison with evolutionary discovery frameworks

Thank you for raising this concern. To compare with AlphaEvolve-style evolutionary discovery frameworks, we directly evaluate against CORAL [1], a state-of-the-art framework for autonomous multi-agent evolution available before our ICLR submission.

Under the same benchmark settings and resource budgets, we compare CORAL, SimpleLoop, and LoopRL across ten scientific discovery tasks:

TaskMetricCORALSimpleLoopLoopRL
Second autocorrelationAutocorrelation lower bound C₂ ↑0.944800.948330.95210
Third autocorrelationReference ratio ↑0.9987000.9991051.000250
Erdős minimum overlapReference ratio ↑1.0000801.0001251.000180
Heilbronn convex / 13Reference ratio ↑0.9998001.0000001.000010
Hexagon packing / 11Reference ratio ↑1.0011001.0014201.001650
Signal processingComposite score ↑0.63250.64180.747
Circle packing / 26Sum of radii ↑2.6354002.6358002.635983
Hadamard / 29Determinant ratio ↑0.9320000.9345000.935673
Bio / BPG0ID NMSE ↓2.00 × 10⁻¹³1.60 × 10⁻¹³1.30 × 10⁻¹³
Materials / MatSci0ID NMSE ↓1.00 × 10⁻¹⁴8.00 × 10⁻¹⁵6.40 × 10⁻¹⁵

SimpleLoop outperforms CORAL on every task, while LoopRL further improves performance.

We additionally evaluate on the two challenging benchmarks used for stress testing in CORAL: Anthropic Kernel Engineering and Frontier-CS Polyominoes Packing. These comparisons use the same benchmark settings, computational resources, and wall-clock time limits.

MethodModelAnthropic Kernel Engineering ↓Frontier-CS Polyominoes Packing ↑
CORALClaude Opus 4.6110389.4%
SimpleLoopClaude Opus 4.687295.4%
CORALQwen 3.8 27B117387.3%
SimpleLoopQwen 3.8 27B107293.2%
LoopRLQwen 3.8 27B104294.9%

With the same underlying model, SimpleLoop achieves better final performance than CORAL on both tasks. With Qwen 3.8 27B, LoopRL provides further improvements over SimpleLoop.

AuthorsOpenReview ↗

Response to Reviewer h5wZ [2/n]

Weakness 3: Diversity of solutions and search strategies

Thank you for this valuable suggestion. We have added quantitative analyses of both solution diversity and search strategy diversity, together with an experiment evaluating explicit diversity control.

Solution diversity. We measure solution diversity as the number of unique solutions divided by the total number of solutions. We use GPT-6.1 Sol with xhigh reasoning effort to analyze individual solutions and identify duplicates.

Search strategy diversity. We examine the search strategy used at each iteration and measure diversity as the number of unique strategies divided by the total number of iterations.

MethodSolution diversity ↑Search strategy diversity ↑
CORAL0.370.42
SimpleLoop0.790.47
LoopRL0.870.50

These results show that SimpleLoop increases both solution diversity and search strategy diversity relative to CORAL, with LoopRL providing further improvements.

Effect of a diversity gate. We additionally introduce a diversity gate, implemented by an agent, to explicitly control solution diversity during search. Whenever a candidate is submitted, the gate checks whether it duplicates any previously accepted candidate. Duplicate candidates are rejected, and the proposing agent is asked to submit a distinct candidate; otherwise, the gate accepts the candidate.

With this diversity gate, the search reaches the same best score using 79% of the tokens required without the gate, which indicates that explicit diversity control can improve token efficiency.


Question 1: Additional experiments

Thank you for this suggestion. We have added the benchmark comparisons, quantitative diversity analyses, and diversity gate experiment described above to the revised manuscript.

We sincerely appreciate your time and constructive feedback. We hope that these additional results address your concerns and would be happy to provide further details if needed.

[1] CORAL: Towards Autonomous Multi-Agent Evolution for Open-Ended Discovery

Reviewer_MA4GOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper introduces Simple Loop, a minimal framework for LLM-based discovery that provides sequential refinement, parallel exploration, and a shared memory as primitives in contrast to predefined heuristics (as in previous work), which may constrain search. The authors then present an RL recipe called LoopRL to provide a method to train Simple Loop across multiple, parallely-executing agents. They evaluate their approach on 10 discovery problems in maths construction, signal processing, and symbolic regression, using a Qwen3.8-27B model over 4 repeat runs.

Critical strengths

The approach presented a method for LLM-based discovery that accounts for the practical deployment scenario of using several agents operating asynchronously, which while not novel to this paper, has been implemented with modifications that seem to provide benefit (e.g., async read/write timestamps, discounted future rewards across loops, improvement over previous best to compute rewards). The minimal simple RL loop also helps to reduce complexity, focusing on the question of whether RL training in such a set up can in fact work, versus the more typical single-turn setups that most previous work has explored. In my view, the authors' main goal, which is well taken, is to show that the recipe is suitable for scaling to realize gains similar to those seen in other settings where scaling has helped.

Critical weaknesses

The quantitative evidence is provided only for signal processing; other domains are only discussed as qualitative comparisons. This makes analyzing the strength of the contributions quite difficult. Separately, it wasn't clear what the motivation is for the choice in eq. 4 for credit assignment to earlier iterations even from unrelated improvements (please see related point in questions).

Questions for authors
  1. Can you provide the quantitative results across all ten benchmarks, not just signal processing?
  2. Why does eq. 4 assign credit to subsequent improvements even when there is no evidence of information reuse? Isn't this a heuristic?
  3. In Fig. 5(b), what does the std. look like across the independent runs?
  4. Can you provide quantitative results for diversity improvement of Simple Loop/Loop RL over existing methods?
  5. How does the proposed paradigm compare to ICL-only discovery loops such as AutoDiscovery [Agarwal et al., 2025], AIScientist [Lu et al., 2024], etc.?
  6. Could you provide some discussion about the time and computational expense of the proposed methodology? RL in search/discovery can require several samples/rollouts before sufficient gains are observed, which has been the focus of many papers as the bottleneck to resolve. Does this method address that at all?
  7. More broadly, it would help to provide some discussion of when the proposed "heuristic-free" loop may be helpful and when it may not. In particular, it is reasonable to imagine that in several discovery domains, domain-specific intuition, even in terms of the search algorithm, may be helpful to incorporate into the loop, in which case an "uninformed" loop may in fact be inefficient in terms of number of rollouts needed to learn an appropriate strategy. Would you agree with this?
2 comments
AuthorsOpenReview ↗

Response to Reviewer MA4G [1/n]

We sincerely thank you for recognizing the strengths of our work and for your insightful questions. Below, we respond to your questions one by one.


Weakness & Q1: Quantitative results across all ten benchmarks

Thank you for pointing this out. We provide quantitative evidence from the other domains in the following table and have updated the paper accordingly.

TaskMetricCORAL [1]SimpleLoopLoopRL
Second autocorrelationAutocorrelation lower bound C₂ ↑0.944800.948330.95210
Third autocorrelationReference ratio ↑0.9987000.9991051.000250
Erdős minimum overlapReference ratio ↑1.0000801.0001251.000180
Heilbronn convex / 13Reference ratio ↑0.9998001.0000001.000010
Hexagon packing / 11Reference ratio ↑1.0011001.0014201.001650
Signal processingComposite score ↑0.63250.64180.747
Circle packing / 26Sum of radii ↑2.6354002.6358002.635983
Hadamard / 29Determinant ratio ↑0.9320000.9345000.935673
Bio / BPG0ID NMSE ↓2.00 × 10⁻¹³1.60 × 10⁻¹³1.30 × 10⁻¹³
Materials / MatSci0ID NMSE ↓1.00 × 10⁻¹⁴8.00 × 10⁻¹⁵6.40 × 10⁻¹⁵

Under the same resource budget, SimpleLoop outperforms CORAL on every task, while LoopRL further improves performance by enabling the model to learn search strategies specific to each problem.

During the rebuttal period, we also conducted additional experiments on the two challenging benchmarks used for stress testing in CORAL [1]: Anthropic Kernel Engineering and Frontier-CS Polyominoes Packing. All methods are evaluated under the same benchmark settings, computational resources, and wall-clock time limits.

MethodModelAnthropic Kernel Engineering ↓Frontier-CS Polyominoes Packing ↑
CORALClaude Opus 4.6110389.4%
SimpleLoopClaude Opus 4.687295.4%
CORALQwen 3.8 27B117387.3%
SimpleLoopQwen 3.8 27B107293.2%
LoopRLQwen 3.8 27B104294.9%

Q2: Motivation for credit assignment in Equation (4)

Thank you for this detailed question. Our intuition is that an ideal scientific discovery loop should achieve better performance as the number of iterations increases. Therefore, in a global, fully asynchronous setting, we directly reward an iteration for improving upon previous iterations. Here, previous iterations are defined as those whose WRITE has completed before the current iteration's READ, since their information is already accessible to the current iteration.

We do not check whether a later iteration directly references an earlier iteration. Instead, we define dependencies between iterations according to information availability. As long as a later iteration achieves an improvement given the contextual information currently accessible to it, we assign a reward. Whether the model actually reads this information or context is left to the model itself to learn.

In our experiments, we observed that the model initially lacked the habit of reading previous information, leading to non-positive iteration improvements and thus zero rewards, which become negative advantages after global advantage normalization. As training progresses, the model's behavior changes, and it learns to write diverse retrieval functions to access information from earlier iterations.


Q3: Interpretation of Figure 5(b)

Figure 5(b) compares two experimental settings using different credit assignment strategies: within-loop credit assignment and LoopRL. Each curve shows the best composite score achieved up to each completed iteration under its respective setting. The standard deviations across independent runs are 0.017 for within-loop credit assignment and 0.009 for LoopRL, both smaller than the performance gap of 0.054. We will clarify this point in the paper.

AuthorsOpenReview ↗

Response to Reviewer MA4G [2/n]

Q4: Quantitative results on diversity

We provide quantitative results on the diversity improvements achieved by SimpleLoop and LoopRL as follows.

For solution diversity, we use the number of unique solutions divided by the total number of solutions as the metric. We use GPT-6.1 Sol with xhigh reasoning effort to analyze the solutions individually.

For search strategy diversity, we examine the strategies used across different iterations and use the number of unique strategies divided by the total number of iterations as the metric.

MethodSolution diversity ↑Search strategy diversity ↑
CORAL0.370.42
SimpleLoop0.790.47
LoopRL0.870.50

These results demonstrate improvements in both solution diversity and search strategy diversity.


Q5: Comparison with AutoDiscovery and The AI Scientist

We compare SimpleLoop with AutoDiscovery [2] and The AI Scientist [3] below and will cite and discuss these works in the related work section.

  1. AutoDiscovery [2] uses a tree-search structure guided by Bayesian surprise. SimpleLoop uses a mesh structure guided by performance scores at individual nodes, with fully asynchronous communication, which contains multiple tree structures spanning different branches as substructures.

  2. The AI Scientist [3] is a framework covering the entire automated research workflow, from topic selection to manuscript writing. SimpleLoop focuses on optimizing solutions to existing tasks through exploration of the search space. The former covers the full research workflow, whereas SimpleLoop focuses on optimizing given problems.


Q6: Time and computational expense

Thank you for this question about resource requirements. We consider time and computational expense in terms of the total time spent using the same GPUs, measured by wall-clock time. For example, when using the same GPUs, we account for the sum of the time spent on inference and training.

For online RL, the main bottleneck is wasted rollouts, since training takes much less time than rollout generation. For example, when rollouts are generated only for training rather than discovery, this computation is wasted. In LoopRL, each sampled trajectory from every branch is used for both training and discovery, and its experience is retained for subsequent discovery, avoiding this waste.

During rollout, i.e., during discovery, we also launch fully asynchronous training to hide the idle time associated with checkpoint updates behind ongoing rollout, allowing fuller utilization of GPU resources and time.


Q7: When a heuristic-free loop is useful

Thank you for this question about the heuristic-free loop. We agree that domain-specific search requires introducing domain-specific heuristics into the loop in advance to avoid unproductive exploration when the model's parametric knowledge does not cover the necessary information.

We will discuss this in the paper. Our experimental observations suggest the following:

  1. When the required domain knowledge is context-dependent and not fully covered by the model's parametric knowledge, a heuristic-guided loop is necessary. Heuristics can incorporate implicit constraints identified through human experience and prevent the generation of invalid solutions. In other words, providing heuristics that humans have already discovered, rather than requiring the agent to rediscover them, can reduce wasted rollouts.

  2. When the relevant domain knowledge is already present in the model's parameters, a heuristic-free loop can be used for search. Fewer heuristics allow a broader search space, enabling the discovery of novel solutions that differ from previous ones. Meanwhile, an overly long prompt in pθ(y;prompt)p_\theta(y;\mathrm{prompt}) can restrict the diversity of solutions sampled by the model.


References

[1] Qu et al. (2026). CORAL: Towards Autonomous Multi-Agent Evolution for Open-Ended Discovery.

[2] Agarwal et al. (2025). AutoDiscovery: Open-ended Scientific Discovery via Bayesian Surprise.

[3] Lu et al. (2024). The AI Scientist: Towards Fully Automated Open-Ended Scientific Discovery.

Comments on the paper

2 comments
AuthorsOpenReview ↗

General Response

We sincerely thank all reviewers for their thorough and constructive feedback.


We are encouraged that the reviewers recognized several key strengths of our work, which we summarize as follows:

  1. Importance and practical relevance. The work addresses an important challenge in LLM-based scientific discovery and studies RL beyond conventional multi-turn settings in an asynchronous multi-agent environment. (MA4G, h5wZ)

  2. Simplicity of the framework. The framework combines explicit reward and return definitions with flexible shared memory, allowing agents to exchange results, failed attempts, and useful approaches while choosing how to use this information. (MA4G, h5wZ)

  3. Detailed analysis. The work covers a broad range of benchmarks and provides detailed descriptions of the discovery process, including observations of how collaboration between agents contributes to improvements. (ZRRL)


In our rebuttal, we have provided the following additional results and clarifications to address the concerns raised:

  1. Quantitative evaluation across domains. We provided comparisons across all ten scientific discovery tasks under the same resource budgets. SimpleLoop outperforms CORAL on every task, while LoopRL consistently provides further improvements. (MA4G, ZRRL)

  2. Challenging benchmarks. We additionally evaluated Anthropic Kernel Engineering and Frontier-CS Polyominoes Packing under the same benchmark settings, computational resources, and wall-clock time limits. (h5wZ, ZRRL)

  3. Contribution of RL Training We provided direct comparisons between SimpleLoop and LoopRL to quantify the contribution of RL training. We also explained how READ/WRITE timestamps define information availability and how the model learns whether and how to retrieve and use earlier experience. (ZRRL, MA4G)

  4. Diversity. We quantified both solution diversity and search strategy diversity, and an explicit diversity gate also reaches the same best score using only 79% of the original token consumption. (MA4G, h5wZ)

  5. Presentation. We revised Figure 2 to show communication between branches using directed arrows, making the information flow and associated performance changes more explicit. We also clarified the role of the qualitative analyses alongside the additional quantitative comparisons. (h5wZ, ZRRL)

  6. Related work. We clarified the differences from ICL-only discovery systems, including AutoDiscovery and The AI Scientist, and expanded the discussion of computational expense, rollout reuse, and asynchronous training. We also discussed when domain-specific heuristics can reduce unproductive exploration and when a minimally constrained loop can provide greater freedom to discover alternative solutions. (MA4G)

We believe these additional results and detailed clarifications strengthen the empirical support for our work and make its contributions clearer. We hope the above information is helpful for the review process. Thank you!

AuthorsOpenReview ↗

General Response

In addition to the results reported in our paper, we conducted further experiments during the rebuttal period:

  1. Frontier-CS2: Across the full task suite, our approach achieves higher overall performance than GPT Astra 6 and Claude Fable 5.1.

  2. Kernel optimization: Our approach achieves 6.42×–37.16× speedups over PyTorch eager execution on VGG16 (full model), LayerNorm, ConvDiv, and ConvMax.

Beyond fixed benchmarks, we also evaluated practical applications in the life sciences:

  1. Protein–ligand docking: On the Interformer task, our approach improves docking accuracy from 63.6% to 69.7%, with a 2.72× end-to-end inference speedup.

  2. Protein structure prediction: Our approach trained a model end to end, reaching the Pareto frontier for efficiency and predictive performance.

These results show that our approach can be used for real-world scientific discovery.

We also evaluated mathematical proof formalization. Our framework formalized the upper bound on the number of maximal zoom-out function pairs from Near Optimal Alphabet-Soundness Tradeoff PCPs, a STOC 2024 Best Paper. It completed the formalization within three hours and passed Lean 4 verification, while Astra and Fable failed on this task in our evaluation. We will publicly release the formalization.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.