PaperlayerView exact revision on arXiv

Source-linked AI summary

Experimental Experience Modeling for Autonomous Research

Wenda Wei, Yingchen Zhang, Ruqing Zhang, Jiafeng Guo, Daiting Shi, Xueqi Cheng

arXiv:2609.39392v1cs.AI

TL;DR

Autonomous research agents face the costly choice of whether to run more experiments, while current agents lack a systematic way to reuse experience for those decisions. EEM distills and retrieves experimental experience and acquires missing evidence through targeted low-cost pilots. On ARC-Bench, it improves research performance while reducing token consumption.

  • Problem

    Experimentation is a major computational cost, and current agents lack a systematic way to reuse prior experimental outcomes to decide whether further investment is warranted.

    Source: §1 · pp. 1–2

  • Method

    EEM distills experimental records into reusable experience, retrieves it to guide investment decisions, and runs a targeted low-cost pilot when historical experience is insufficient.

    Source: §1 · p. 2

  • Results

    EEM improves the overall score by 13.1% over AutoResearchClaw on 25 ARC-Bench topics while reducing average total token consumption by 11.4%.

    Source: §1 · p. 2

  • Takeaways & Limitations

    On ARC-Bench, the results support reusing accumulated experience and selectively acquiring additional evidence to improve research performance while reducing token consumption.

    Source: §5 · p. 9

    The experience library is initialized separately for each research task; systematic cross-task transfer and its risk of negative transfer are left to future work.

    Source: §2.4 · p. 5

Abstract

from arXiv · show

Autonomous research agents can generate hypotheses and conduct experiments, but experimentation remains a major source of computational cost. A fundamental challenge is deciding which experiments are worth running, particularly when prior evidence is insufficient to resolve uncertainty. Yet current research agents lack a systematic way to leverage experimental experience when making such decisions. We introduce Experimental Experience Modeling (EEM), a framework for making informed experimental decisions by acquiring, reusing, and accumulating experimental experience. EEM extracts decision-relevant records from earlier experimental trajectories, distills them into reusable experience, and organizes them in an experience library. For a new experimental decision, EEM retrieves relevant historical experience and assesses whether it provides sufficient support for deciding whether a candidate direction warrants further investment. When historical experience is insufficient, EEM conducts a targeted, low-cost pilot experiment to acquire the missing decision-relevant experience on demand. It then combines this newly acquired experience with retrieved historical experience to determine whether the direction warrants full-scale evaluation, which requires substantial resources. The resulting experimental outcomes are further distilled into reusable experience, allowing the library to continually grow through iterative accumulation. Experiments on autonomous research benchmarks show that EEM improves research performance while reducing model interaction overhead, demonstrating the value of reusing accumulated experience and acquiring additional experience only when needed.

1 INTRODUCTION

Autonomous research makes experimentation a major computational cost, so agents must decide when existing evidence is enough and when more evidence is warranted. EEM addresses this by organizing experimental experience for reuse and targeted acquisition, and the paper reports improved research performance with lower token consumption.

1 INTRODUCTION

  • Experimentation is a major computational cost, making it important to avoid expensive runs when evidence supports a direction without over-relying on limited or uncertain evidence.The decision is not only what to experiment with, but which experiments are worth running.Source: §1 · p. 1
  • Experimental experience captures what works under particular conditions and how findings should influence later choices, with low-cost pilots offering targeted evidence before substantial investment.The resulting evidence is incorporated into subsequent decisions.Source: §1 · pp. 1–2
  • Existing agents learn from feedback and execution traces, but do not systematically distill reusable experience or assess whether it suffices for a new investment decision.When historical experience is insufficient, they also do not acquire only the missing decision-relevant experience through a targeted low-cost experiment.Source: §1 · p. 2
  • EEM distills decision-relevant records into reusable experience, checks whether retrieved experience supports a decision, and uses a targeted low-cost pilot when it does not.Pilot and full-scale outcomes are added to the experience library for later decisions.Source: §1 · p. 2
  • 13.1% overall-score improvement over AutoResearchClaw, the strongest evaluated baseline, with result analysis up 31.7% and average total token consumption down 11.4% across 25 ARC-Bench topics (Liu et al., 2026).The evaluation reports improved research performance alongside reduced model interaction overhead.Source: §1 · p. 2

2 METHODOLOGY

EEM focuses on experimentation, reusing historical experience to guide investment decisions and running low-cost pilots when that experience is insufficient. It turns experimental outcomes into reusable records and lessons for later decisions.

2.1 OVERVIEW

  • EEM focuses on experimentation within a three-stage autonomous research workflow that starts with planning and ends with paper writing.The workflow takes a scientific question as input and produces a research paper.Source: §2 · p. 2, §2.1 · p. 3

2.2 HISTORICAL EXPERIENCE ACQUISITION

  • Experience items pair a source record with a lesson stating applicability conditions, the evidence-bounded finding, and its implication for future investment.Records can come from pilots, full-scale evaluations, refinement attempts, or interrupted runs with informative partial results or failures.Source: §2.2 · pp. 3–4
  • EEM localizes decision-relevant records, grounds observed consequences in attempted directions, and abstracts them into reusable guidance about conditions, findings, and future investment.This separates empirical support from the agent’s original expectations.Source: §2.2 · p. 4

2.3 EXPERIENCE-GUIDED EXPERIMENTAL DECISION MAKING

  • EEM tests whether available experience supports investing in a candidate direction, using historical evidence when sufficient and targeted, low-cost pilots when it is not.The resulting evidence supports a decision to proceed to full-scale evaluation, revise the direction, or abandon it before substantial resources are committed.Source: §2.3 · pp. 4–5

2.4 ITERATIVE EXPERIENCE ACCUMULATION

  • Pilot and full-scale outcomes are accumulated in the experience library for reuse, and retained pilot evidence can inform later decisions even when a direction is not selected for full-scale execution.Within this paper’s setting, the library starts empty and accumulates experience within the same research run; cross-task transfer is left for future work.Source: §2.4 · p. 5
  • Table 1 reports ARC-Bench results using an overall score that weights Code Development, Code Execution, and Result Analysis at 25:25:50, with higher scores better.The caption identifies the table as comparing EEM with baselines but does not provide cell values in the supplied passage.Source: Table 1 · p. 6

3 RESULTS

The results evaluate EEM on ARC-Bench across overall performance, token use, component ablations, backbone applicability, and repeated-run stability. EEM outperforms the strongest evaluated baseline overall, while the ablations and additional analyses examine its components and consistency.

3.1 EXPERIMENTAL SETUP

  • The evaluation covers 25 ARC-Bench topics, three experimentation metrics, matched model and sandbox settings, and pilots limited to 120 seconds.The overall score weights code development and execution at 0.25 each and result analysis at 0.50.Source: §3 · p. 6, §3.1 · p. 6

3.2 OVERALL PERFORMANCE

  • EEM scores 0.674 overall versus AutoResearchClaw’s 0.596, with code execution rising from 0.562 to 0.619 and result analysis from 0.442 to 0.582.AutoResearchClaw is the strongest evaluated baseline; EEM’s code-development score is lower than its score.Source: §3.2 · pp. 6–7

3.3 EFFICIENCY

  • Table 2 compares EEM and AutoResearchClaw’s input and output token consumption across planning, experimentation, and writing.The supplied passages describe the comparison scope but provide no token-consumption values.Source: Table 2 · p. 7, §3.3 · p. 7

3.4 ABLATION STUDY

  • Removing historical experience lowers the overall score from 0.674 to 0.506, a 24.9% relative decrease; removing on-demand acquisition yields 0.565.Retaining source records without experience distillation yields 0.649, indicating additional value from distilling records into reusable guidance.Source: Table 3 · p. 7, §3.4 · pp. 7–8

3.5 APPLICABILITY ACROSS BACKBONE MODELS

  • With GPT-5.5, EEM completes the workflow on ML10, but scores lower than with GPT-5.3-Codex, especially in code development.The result-analysis score declines less than the code-development score.Source: Figure 2 · p. 8, §3.5 · p. 8

3.6 STABILITY ACROSS RUNS

  • Across four ML10 runs, overall scores remain approximately 0.68–0.77, while result analysis is unchanged and code development and execution vary.The authors describe this as repeatability under the evaluated configuration, rather than dependence on one successful run.Source: Figure 2 · p. 8, §3.6 · p. 8

4 RELATED WORK

Prior work integrates autonomous research workflows and uses feedback, memory, or lower-cost evaluations to improve research agents. EEM focuses on decision-specific experimental experience, reusing it and acquiring missing evidence through targeted pilots when needed.

Autonomous research

  • Autonomous research systems integrate literature review, experimentation, and writing, while several systems refine ideas or implementations from experimental feedback.EEM focuses specifically on experimental decision making within this broader direction.Source: §4 · pp. 8–9

Experience Reuse in Language Agents

  • Language-agent memory and experience methods retain, retrieve, or distill information, whereas EEM distills experimental records into guidance for resource-investment decisions.EEM’s experience specifies applicability conditions, empirical findings, and implications for future investment.Source: §4 · p. 9

Cost-Efficient Experimental Decision Making

  • Multi-fidelity methods allocate resources among predefined candidates or fidelity levels; EEM reuses historical experience and pilots only when decision-relevant evidence is missing.EEM addresses open-ended research directions that may require different evidence to decide whether they are worth pursuing.Source: §4 · p. 9
  • Unlike research preference models that predict which candidates to execute, EEM supports decisions to pursue, revise, or reject a direction using accumulated experience and targeted acquisition.An agentic preference-model variant uses small-scale pilots to improve candidate ranking.Source: §4 · p. 9

5 CONCLUSION

EEM uses accumulated experimental experience to guide resource investment in autonomous research, and ARC-Bench experiments show improved research performance with reduced model token consumption.

5 CONCLUSION

  • EEM improves research performance while reducing model token consumption on ARC-Bench, with ablations supporting the effectiveness of its design.Source: §5 · p. 9
  • EEM distills experimental trajectories into reusable experience, retrieves applicable experience for investment decisions, and acquires missing evidence through targeted, low-cost pilots.Experience from both pilot and full-scale experiments is accumulated for subsequent decisions.Source: §5 · p. 9
  • Future work will investigate transferring experience across research tasks and balancing experience reuse with evidence acquisition.Source: §5 · p. 9

Related papers

Loading 2609.39392v1…