PaperlayerView exact revision on arXiv

Source-linked AI summary

AlphaResearch: Accelerating New Algorithm Discovery with Language Models

Zhaojian Yu, Kaiyue Feng, Yilun Zhao, Shilin He, Xiao-Ping Zhang, Arman Cohan

arXiv:2511.08522v2cs.CL

TL;DR

Open-ended algorithm discovery is difficult when the target solution is unknown, and repeated attempts can raise validation costs. AlphaResearch combines simulated peer review with program execution in an iterative idea-and-code search. It outperforms AlphaEvolve on two of eight problems and converges faster than ShinkaEvolve on three of four, with a best-of-known result on Packing Circles.

  • Problem

    Open-ended coding agents struggle to converge on the best solution in a few attempts when the optimal implementation is underspecified, while repeated sampling can increase validation cost.

    Source: §1 · p. 2

  • Method

    AlphaResearch iteratively generates research ideas, filters them with a simulated peer-review reward model, verifies programs by execution, and optimizes proposals using both rewards.

    Source: §1 · p. 2, §3.1 · p. 4, §3.2 · p. 4

  • Results

    AlphaResearch outperforms AlphaEvolve on 2/8 problems and converges faster than ShinkaEvolve on three of four; Packing Circles reaches best-of-known scores of 2.636 for n=26 and 2.939 for n=32.

    Source: §1 · p. 3, §4.3 · p. 7, §5 · p. 9

  • Takeaways & Limitations

    The results support combining interleaved idea generation and program generation with a peer-review environment for autonomous algorithm discovery.

    Source: §5 · p. 9

    AlphaResearch-RM is trained exclusively on ICLR reviews, which may make it less reliable outside core machine-learning topics.

    Source: A Limitations · p. 13

Abstract

from arXiv · show

LLMs have made significant progress in complex but easy-to-verify problems, yet they still struggle with discovering the unknown. In this paper, we present \textbf{AlphaResearch}, an autonomous research agent designed to discover new algorithms on open-ended problems by iteratively running the following steps: (1) propose new ideas (2) program to verify (3) optimize the research proposals. To synergize the feasibility and innovation of the discovery process, we construct a novel dual environment by combining the execution-based verifiable reward and reward from simulated real-world peer review environment in AlphaResearch. We construct \textbf{\dataset}, a set of questions that includes an eight open-ended algorithmic problems competition to benchmark AlphaResearch. Experimental results show that AlphaResearch achieves stronger discovery performance than other agentic discovery systems on six open-ended problems. Notably, the algorithm discovered by AlphaResearch on the \emph{``packing circles''} problem achieves the best-of-known performance, surpassing the results of human researchers and strong baselines from recent work (e.g., AlphaEvolve). Additionally, we conduct a comprehensive analysis of the benefits and remaining challenges of autonomous research agent, providing valuable insights for future research.

1 Introduction

Open-ended algorithm discovery remains difficult because optimal implementations are underspecified, while brute-force search and imperfect novelty judgments can waste validation effort. AlphaResearch combines simulated peer review with executable verification and is evaluated on eight open-ended problems.

1 Introduction

  • Open-ended coding agents struggle to find the best solution in few attempts when the target implementation is underspecified.Prior coding-agent scaffolds can reach expert performance on end-to-end problems with verifiable targets.Source: §1 · p. 2
  • AlphaEvolve (Novikov et al., 2025) scales attempts through repeated sampling, while ShinkaEvolve (Lange et al., 2025) filters edits but relies on imperfect novelty judgments.Embedding thresholds may miss meaningful changes or subtle redundancy, and LLM novelty judgments may diverge from expert assessments.Source: §1 · p. 2
  • AlphaResearch iteratively generates research ideas, filters them with simulated peer review, verifies programs by execution, and optimizes proposals using both rewards.The authors describe this dual environment as reducing ineffective attempts and program executions.Source: §1 · pp. 2–3
  • AlphaResearch scores 2.939 on Packing Circles (n=32), surpassing AlphaEvolve, and converges faster than ShinkaEvolve on three of four problems.The benchmark contains eight open-ended research problems.Source: §1 · pp. 2–3

2 Related Work

Related work addresses idea generation and code generation as complementary parts of autonomous research. AlphaResearch combines program execution with a reward model trained on peer-reviewed research records.

LLMs for New Ideas

  • Prior idea-generation approaches refine novelty or compare ideas with experts, but lack a verification method or rely on automatic execution or LLM evaluation.The cited approaches include iterative refinement (Wang et al., 2024; Baek et al., 2024), expert comparison, and execution-based studies (Huang et al., 2024; Tian et al., 2024; Li et al., 2024; Lu et al., 2024; Aygün et al., 2025).Source: §2 · p. 3
  • AlphaResearch combines program execution with a reward model trained on real-world peer-reviewed research records.This is presented as a direction for combining idea verification with research feedback.Source: §2 · p. 3

LLMs for Code Generation

  • Code-generation models and benchmarks target natural-language-to-code synthesis, while real-world agentic coding tasks include SWE-Bench, PaperBench, and MLE-Bench.The passage cites SWE-Agent and OpenHands as frameworks that contributed to agentic LLM development.Source: §2 · p. 3

3 AlphaResearch

AlphaResearch iteratively samples prior research states, generates ideas and programs, and evaluates programs through peer review and executable verification. The loop retains the best result and stops at the round limit or when it exceeds the human benchmark.

3.1 Overview

  • The overview initializes an idea, program, and execution rating, then updates the best record as rounds produce higher ratings.The loop ends when the best result exceeds the best-of-human score or the maximum number of rounds is reached.Source: §3.1 · pp. 3–4

3.2 Actions

  • Each action samples a prior trajectory step, generates an idea, rejects it if its reward-model score is negative, then generates and executes a program.The resulting idea, program, and evaluation are added to the trajectory; the best result is returned after the maximum rounds.Source: §3.2 · p. 4

3.3.1 Reward from Real-world Research Records

  • AlphaResearch-RM-7B is fine-tuned on ICLR abstracts paired with average review ratings; it reaches 72% binary accuracy on unseen 2025 ideas.The binary threshold treats overall ratings above 5.5 as positive; the training and test periods are disjoint.Source: §3.3.1 · p. 5, Table 2 · p. 5

3.3.2 Reward from Program-based Execution

  • The executor checks whether programs satisfy problem constraints and measures their performance; stored results can inform later idea generation.The authors say verifiable execution simplifies the action space and enhances discovery efficiency.Source: §3.3.2 · p. 5

4 Experiments

Across eight open-ended problems, AlphaResearch achieved a 2/8 win rate against human researchers, with its Packing Circles result exceeding the cited human and AlphaEvolve values. Comparisons and ablations examine discovery performance, peer-review filtering, and the iterative process.

4.1 Setup

  • The agent used o4-mini, while AlphaResearch-RM-7B was produced by supervised fine-tuning Qwen-2.5-7B-Instruct on collected ICLR records; OpenEvolve and ShinkaEvolve served as baselines because AlphaEvolve's codebase was unavailable.The paper reports a learning rate of 1e-5, a global batch size of 128, and two training epochs for the reward model.Source: §4.1 · p. 6

4.2 Evaluation

  • AlphaResearchComp covers eight program-based research tasks across geometry, number theory, harmonic analysis, and combinatorial optimization, each with an automatically evaluable objective and a best-known human result.The stated human values are conjectured best-known values, not proven optima.Source: §4.2 · pp. 6–7

4.3 Main Results

  • AlphaResearch won on 2/8 problems; on Packing Circles it scored 2.636 (n=26) and 2.939 (n=32), exceeding the cited human and AlphaEvolve values.For n=26, human researchers scored 2.634 and AlphaEvolve 2.635; for n=32, they scored 2.936 and 2.937, respectively.Source: §4.3 · p. 7

4.4 Comparison with OpenEvolve and ShinkaEvolve

  • Within 1,000 iterations, AlphaResearch surpassed OpenEvolve on the first four problems and ShinkaEvolve on three of four.The reported advantage over ShinkaEvolve tended to grow as iterations increased.Source: Figure 3 · p. 7, §4.4 · p. 8

4.5 Ablations and Analysis

  • On Third Autocorrelation Inequality, GPT-5 reached strong performance faster than o4-mini early, but their later performance was comparable.The comparison ran for 200 steps.Source: §4.5 · p. 8, Figure 4 · p. 8
  • Compared with the baseline without RM during discovery, AlphaResearch-RM-7B filtered 151 ideas below threshold, correctly rejecting 108 execution failures and erroneously rejecting 43 viable ideas, for 71.5% accuracy (108/151).Source: §4.5 · p. 8
  • About 30%–40% of proposed ideas fell below the reward-model threshold; execution success was 28.9% on Packing Circles and 51.7% on Third Autocorrelation Inequality, while rewards rose in both examples.Execution outcomes varied by problem, but the reported reward trends increased in both examples.Source: §4.5 · pp. 8–9, Figure 6 · p. 9

4.6 Case Study

  • In the Packing Circles case study, each iteration samples prior idea, program, and execution metadata, then uses a positively scored idea to guide a program patch before verification.The process retains the best program and idea as its final solution.Source: §4.6 · p. 9

5 Conclusion

AlphaResearch combines research-idea generation with program-based verification in a dual environment for algorithm discovery. On eight open-ended problems, it outperformed AlphaEvolve on two and showed better discovery ability than prior evolutionary agents.

5 Conclusion

  • AlphaResearch combines new-idea generation with program-based verification in a dual environment and outperformed AlphaEvolve on 2/8 problems.The paper also reports better discovery ability than previous state-of-the-art evolutionary agents.Source: §5 · p. 9

A Limitations

The limitations concern the reward model's training data and target, as well as the benchmark's scope.

A Limitations

  • AlphaResearch-RM-7B is trained on ICLR reviews, which may make it less reliable outside core machine-learning topics; review scores may not reflect long-term value, feasibility, or practical utility, and the benchmark scope is narrow.The paper notes that peer review assesses perceived novelty, technical soundness, and clarity at submission, while most evaluations focus on standard machine-learning tasks and benchmarks.Source: A Limitations · p. 13

C Other Details

The additional analyses examine reward trends, generated-idea ratings, and comparisons with evolutionary agents. They report rising execution-based rewards in two examples and stronger idea scores and discovery performance in the stated comparisons.

Reward trends

  • For Packing Circles and Third Autocorrelation Inequality, execution-based rewards initially grew rapidly and then more slowly plateaued.The paper reports consistent improvement throughout discovery on 6/8 problems.Source: C Other Details · p. 13, Figure 7 · p. 13

Execution-only agent against AlphaResearch

  • Ideas generated by AlphaResearch generally received higher novelty scores than ideas from the execution-only agent.The comparison used AlphaResearch-RM-7B to score the full distributions of 1,000 ideas from each agent without filtering.Source: C Other Details · pp. 13–14, Figure 8 · p. 14

Agent comparison

  • On Packing Circles (n=26), AlphaResearch performed better than OpenEvolve and slightly better than ShinkaEvolve during the first 500 steps.The comparison used program-based rewards for OpenEvolve and ShinkaEvolve, and both program-based and peer-review rewards for AlphaResearch.Source: Figure 9 · p. 14

D Examples of Packing Circles

This section illustrates AlphaResearch’s Packing Circles constructions, including a 32-circle comparison and an example formatted task.

D Examples of Packing Circles

Loading 2511.08522v2…