acceptodds
Under review as a conference paper at ICLR 2027

Collaborative Principle Evolution via Evidence Transfer for Scientific Discovery

Abstract

Large Language Model (LLM)-based agents promise to automate scientific discovery, yet exploring the vast hypothesis space remains costly. Existing principle-evolution methods accelerate this loop, but operate sequentially, which caps exploration breadth and wastes wall-clock time on challenging problems. To address this, we formulate collaborative scientific discovery as evidence transfer between parallel principle-evolution branches. We present CoEvolve, which realizes this transfer through a coordination core over parallel branches. By integrating value-of-information-gated routing and context-discounted likelihood injection, CoEvolve enables branches to collaborate through shared measurements while keeping their principle posteriors separate. Across six scientific-discovery tasks under a matched evaluation budget, CoEvolve attains a mean solution quality of 66.5% versus 57.0% for single-branch principle evolution, with a mean wall-clock speedup on the GPT-5.6-Terra backbone; on five auto-research tasks delegated to an autonomous research harness, it is the only arm whose mean stays above the published SOTA anchor on every task. These results establish when evidence sharing accelerates parallel discovery and when transfer safeguards are necessary to limit negative or inert transfers.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
large language models, scientific discovery, principle evolution, collaboration

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

This paper proposes an improved Agentic AI method to scientific discovery, introducing and leveraging the idea of evidence transfer, which shares information across parallel branches. It benchmarks the approach on AI for scientific discovery benchmarks and shows modest gains.

Critical strengths
  1. Extensive benchmarking is conducted.
Critical weaknesses
  1. The paper is very difficult to read and as such is not ready for publication.
  2. The key proposed concept 'evidence transfer' does not appear to be defined.
Questions for authors
  1. Address the weaknesses
Reviewer_JC6uOpenReview ↗
Rating
3
Confidence
3(+1)
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The work introduces CoEvolve, a multi agent system that coordinates parallel PiEvo search branches through selective evidence transfer while maintaining separate principal posteriors. The method helps branches learn from one another’s experimental results by selecting useful evidence, reducing the influence of less reliable evidence, and avoiding redundant information. Experiments cover six surrogate-scored optimization tasks with two LLM backbones and an additional five autonomous research tasks. Results show better solutions on average and shorter search times, generally at the cost of higher token usage. Ablation studies support the benefits of evidence sharing.

Critical strengths
  1. The paper provides a clear theoretical establishment of how sharing results can help parallel search and under what assumptions harmful transfers can be controlled. This gives a useful rationale for the method’s design.
  2. The experiments cover multiple scientific tasks, two LLM backbones, and autonomous research workflows. Comparisons with independent and non-sharing branches help establish the value of evidence sharing, while component ablations support key design choices. These controls make the reported improvements more convincing on the tested benchmarks.
Critical weaknesses
  1. The six main tasks rely on surrogate scores without independent validation of the final designs. Despite safeguards against reward hacking, it remains unclear whether higher scores reflect real biological or physical improvements rather than predictor biases.
  2. The main benchmarks lack comparisons with domain-specific optimization methods. Evaluating these methods under the same oracle, constraints, and query budget would help readers understand the performance gap between agent frameworks and specialized approaches.
Questions for authors
  1. Could you clarify the configurations used for Codex and Claude Code, including reasoning-effort levels, token budgets, and whether subagents were enabled? How were these settings controlled across the compared systems?
  2. Why was DeepSeek-v4-Flash-0731 chosen for the harness experiments, while the main benchmarks use Gemma and GPT? What motivated using DeepSeek with Codex and Claude Code rather than their native GPT and Claude models? Are results available for these native configurations or for DeepSeek Harness using the same DeepSeek backbone? These comparisons would help clarify the effects of the model and harness choices.
Reviewer_LdWNOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes CoEvolve, a multi-agent collaborative optimization framework for scientific discovery. Building upon the existing PiEvo framework, CoEvolve extends sequential scientific exploration by a single agent to parallel exploration by multiple agents. It also introduces a carefully designed information-sharing mechanism that enables collaborative evolution across different branches.

Critical strengths
  1. The paper presents a comprehensive and well-motivated mechanism for information transfer across branches. It employs Value of Information VoI-based scoring and routing to identify experimental evidence valuable to target branches, and injects the selected evidence through discounted likelihood updates. This design effectively addresses challenges such as evidence redundancy, differences in experimental contexts, and potential negative transfer.

  2. Unlike conventional multi-agent approaches that primarily rely on sharing contextual information, CoEvolve enables collaboration through probabilistic updates to each branch's principle posterior. This mechanism provides a more explicit mathematical formulation for information exchange, supported by theoretical analysis of its effectiveness, and offers a new perspective on multi-agent scientific discovery.

  3. The paper provides extensive experiments that demonstrate the effectiveness of the proposed framework. In particular, the ablation studies clearly illustrate the contributions of individual components, providing empirical support for the overall design.

Critical weaknesses
  1. Although the ablation studies demonstrate the importance of evidence discounting, the paper does not establish whether the proposed adaptive weighting mechanism is necessary or superior to simpler fixed discount factors. A comparison with fixed discount weights would help clarify the contribution of the adaptive discounting design.
Questions for authors

see weakness

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.