PaperlayerView exact revision on arXiv

Source-linked AI summary

ReGE: Stronger Agents through Experience Synthesis

Jiarong Liang, Yubo Wang, Zijie Liu, Yuxuan Zhang, Ping Nie, Wenhu Chen

arXiv:2610.33861v1cs.AI

TL;DR

Interactive attempts can reveal facts, useful operations, and failures, but agents need a way to turn that experience into guidance for later runs. ReGE-Syn jointly synthesizes recorded attempts into guidance for a fresh student execution, and ReGE improves accuracy over unguided execution across four benchmarks.

  • Problem

    Interactive attempts reveal facts, useful operations, and failures, but the question is how an agent can turn its own experience into more effective subsequent execution.

    Source: §1 · p. 1

  • Method

    ReGE-Syn uses one coach call to jointly synthesize bounded execution records and artifacts into self-contained guidance that a fresh student adapts to current observations.

    Source: §1 · p. 2

  • Results

    ReGE improves accuracy by an average of 9.9 percentage points for GPT and 8.8 percentage points for Claude compared with unguided execution across four benchmarks.

    Source: §1 · p. 2

  • Takeaways & Limitations

    ReGE offers a practical way to make an agent’s prior attempts useful for subsequent action while keeping the base model fixed.

    Source: §6 · p. 10

    Evaluation is limited to four benchmarks, with ablations on selected subsets; broader task settings and repeated executions with aligned settings remain for future evaluation.

    Source: §6 · p. 10

Abstract

from arXiv · show

Repeated attempts at interactive tasks uncover facts, useful operations, and failures that can inform subsequent execution. We introduce ReGE, a framework for turning this recorded experience into guidance for a fresh attempt. Its main implementation, ReGE-Syn, uses a single coach call to jointly synthesize bounded execution records and available artifacts into self-contained guidance, connecting task-specific findings with proposed actions and relevant pitfalls. A student starts from the task's initial environment state and adapts this guidance to current observations. Source generation, synthesis, and student execution use the same base model, without parameter updates or a stronger teacher. Across four benchmarks, ReGE improves accuracy by an average of 9.9 percentage points for GPT and 8.8 percentage points for Claude compared with unguided execution. On matched BrowseComp-Plus subsets, ReGE-Syn reduces guidance-construction tokens by 98.2\% on GPT and 95.5\% on Claude relative to adapted AggAgent, with observed success. Combined construction and execution tokens decrease by 15.7\% and 35.4\%, respectively, counting input plus output and excluding source generation. BrowseComp-Plus ablations further associate access to process information with fewer student steps and show that joint synthesis produces substantially shorter guidance than Direct history or Independent summaries, with observed success. These results demonstrate how agents can use their own recorded experience to improve subsequent task completion, and suggest a low-cost way to consolidate the execution records of parallel agent systems such as agent swarms.

1 INTRODUCTION

ReGE turns recorded attempts into guidance for a fresh execution, where the student must perform operations in the current environment. Across four benchmarks, ReGE improves accuracy over unguided execution, while ReGE-Syn reduces guidance-construction and combined usage on matched BrowseComp-Plus subsets.

1 INTRODUCTION

  • Useful findings may be distributed across long records, making it challenging to retain actionable details while reducing the history a new agent processes.Candidate selection, response aggregation, and trajectory summarization offer distinct ways to reuse prior attempts.Source: §1 · p. 1
  • A fresh student must perform required operations again because files and application states from previous attempts do not carry over after reset.Repeated attempts can supply facts, useful operations, intermediate results, and failed strategies that inform subsequent actions.Source: §1 · pp. 1–2
  • ReGE improves accuracy by an average of 9.9 percentage points for GPT and 8.8 percentage points for Claude compared with unguided execution.On matched BrowseComp-Plus subsets, ReGE-Syn reduces guidance-construction tokens by 98.2% on GPT and 95.5% on Claude relative to adapted AggAgent, with observed success; combined construction and execution tokens decrease by 15.7% and 35.4%, respectively, excluding source generation.Source: §1 · p. 2
  • ReGE-Syn jointly connects findings, operations, and failure lessons from recorded attempts into actionable guidance for a fresh execution by the same base model.Source generation, synthesis, and student execution use the same base model, without parameter updates or a stronger teacher.Source: §1 · p. 2

2 REGE: EXPERIENCE-GUIDED AGENT EXECUTION

ReGE constructs guidance from recorded experience for a fresh student execution. ReGE-Syn jointly synthesizes bounded records and artifacts, while the student adapts the resulting guidance to its own environment.

2 REGE: EXPERIENCE-GUIDED AGENT EXECUTION

  • ReGE constructs guidance from recorded experience for a fresh execution; the student starts from the task’s initial state and uses guidance alongside current observations.The framework also includes ReGE-Sel, which forwards one selected trajectory; the paper focuses on joint synthesis.Source: §2 · p. 2

2.1 RECORDED EXPERIENCE

  • Each source attempt is rendered as a separate, step-ordered block with available narratives, commands, and final artifact paths and contents; fields are copied and truncated to bound input length.No per-trajectory LLM summary is generated before the coach sees the rendered records.Source: §2.1 · p. 3

2.2 JOINT EXPERIENCE SYNTHESIS

  • A single coach call sees all retained source blocks together, preserving attempt boundaries and step order to connect findings with proposed operations and pitfalls.The default guide is an execution-ordered plan with paths, values, commands, format requirements, problems, and verification steps placed where they matter; recommendations remain subject to student checks because correctness labels are unavailable.Source: §2.2 · p. 3

2.3 GUIDED STUDENT EXECUTION

  • The student begins from the initial state and uses current observations and interaction history to adopt, adjust, or skip proposed operations, performing required actions in its own environment.Files and environment changes from source attempts do not automatically carry over.Source: §2.3 · p. 4

3 EXPERIMENTAL SETUP

The evaluation compares ReGE with unguided execution and alternative reuse methods across four benchmarks, using fixed base models and matched execution limits. It reports token use after source-pool generation as its primary inference-cost measure.

3 EXPERIMENTAL SETUP

  • Evaluation covers Terminal-Bench 2.0, Spider2-Lite, AndroidWorld, and BrowseComp-Plus, with benchmark-specific task, tool, and step budgets.BrowseComp-Plus uses a fixed corpus and retrieval tools without live web access; AndroidWorld evaluation excludes two templates that crash the emulator.Source: §3 · p. 4
  • GPT and Claude base models remain fixed across source generation, guidance construction, and student execution; each source pool contains four attempts per task.ReGE and AggAgent share source pools and each receive one fresh student execution under the same benchmark-specific step limit.Source: §3 · p. 4
  • Baselines include one unguided execution, a Reflexion adaptation, adapted AggAgent, independent summaries, and references without student re-execution.The references are Self-Consistency voting over source outputs and Best-of-N reporting the selected source’s original score.Source: §3 · p. 4
  • The primary inference-cost metric sums input and output tokens used after source-pool generation, including method-specific guidance construction and execution calls.Input-only usage and costs including source generation are identified separately.Source: §3 · p. 4

4 RESULTS

Across four benchmarks, ReGE improves task completion over unguided execution, while joint synthesis reduces guidance-construction and combined usage. Results also show that process information, mixed source pools, and fresh student adaptation matter to the observed outcomes.

4 RESULTS

  • Across four benchmarks, ReGE improves task completion over Control; on GPT Terminal-Bench 2.0, success rises from 41.6% to 56.2%.Table 2 reports task success and source-inclusive token costs across the four benchmarks.Source: §4 · p. 4, Table 2 · p. 5

4.1 EFFICIENT EXPERIENCE SYNTHESIS

  • ReGE-Syn cuts BrowseComp-Plus guidance-construction tokens by more than 95% on both models versus AggAgent, and has the fewest combined construction and execution tokens among four methods.Students use more tokens than with AggAgent, but the construction savings more than compensate; savings vary by task, with Spider2-Lite SQL calls down 50% and output tokens down 30%, while AndroidWorld measures remain nearly unchanged.Source: §4.1 · p. 5, Table 4 · p. 6

4.2 COMPARISON WITH OTHER MODELS

  • On the 468-task subset, ReGE raises gpt-5.6-luna success from 56.8% to 67.4%, matching gpt-5.6-sol’s 67.0% at 28% of its cost.With gpt-6-luna, ReGE is 18.1 points more accurate on average, while costing more: gpt-6-luna costs 4% as much including source generation and 26% as much with an existing pool.Source: §4.2 · p. 6, Table 5 · p. 6

4.3 WHAT INFORMATION MAKES EXPERIENCE USEFUL?

  • On BrowseComp-Plus, providing one recorded attempt substantially improves success over coaching without source experience, and four attempts add another 4.0 percentage points.Adding process records also helps: 6.3% of queries with only one submitted source answer account for 90% of net additional successes; with four answering sources, student steps fall from 8.9 to 4.8 despite nearly identical success.Source: §4.3 · pp. 6–7, Table 6 · p. 7, Table 7 · p. 7

4.4 WHEN DOES RECORDED EXPERIENCE HELP?

  • Mixed pools of successful and failed attempts yield the largest observed gains over Control in all eight model–benchmark combinations, reaching 48.0 points on GPT and 35.2 on Claude BrowseComp-Plus.When all sources fail, improvements are smaller and inconsistent; when all succeed, Control already reaches at least 90% success in every model–benchmark combination.Source: §4.4 · p. 7, Table 8 · p. 8

4.5 JOINT SYNTHESIS FOR COMPACT GUIDANCE

  • On matched BrowseComp-Plus queries, ReGE-Syn guidance is roughly one quarter as long as Direct history or Independent summaries, with combined tokens down 19.8% and 31.8%, respectively.Direct history concatenates trajectories; Independent summaries concatenate separately generated summaries, while ReGE-Syn jointly constructs one guide from the same four-source pools.Source: §4.5 · pp. 7–8, Table 9 · p. 8

4.6 GUIDANCE VARIANTS

  • ReGE-A reduces combined tokens by 32.1% while observed success rises from 88.3% to 90.7%; ReGE-Sel reaches 89.3% success with 21% fewer student tokens than ReGE-Syn.ReGE-A’s answer format is not best in every tested setting: on another 100-query comparison, it solves 80% versus ReGE-Syn’s 82%.Source: §4.6 · p. 8, Table 10 · p. 9

4.7 FROM RECORDED EXPERIENCE TO FRESH EXECUTION

  • In database recovery, the student used the guide’s XOR key and backup advice, adjusted the cleanup order, and passed all seven official tests.For BrowserMultiply, unsuccessful source attempts still supplied a route to open the page; the student then obtained the values and completed the task.Source: §4.7 · p. 9

5 RELATED WORK

Prior work reuses repeated attempts through aggregation, distillation, reflection, and memory, while ReGE-Syn jointly synthesizes records from attempts at the same task into guidance for a fresh execution. The paper also frames this as complementary to consolidating parallel agent-swarm attempts.

5 RELATED WORK

  • Unlike iterative aggregation that selects or combines sampled responses (Cobbe et al., 2021; Jiang et al., 2023; Wang et al., 2024b), ReGE-Syn guides a fresh agent using execution records and artifacts.Self-MoA jointly aggregates responses from one model (Li et al., 2025), while AggAgent inspects candidate solutions and retrieves trajectory information with tools (Lee et al., 2026).Source: §5 · pp. 9–10
  • ReGE-Syn jointly reads bounded execution records and artifacts to produce guidance for a fresh environment interaction, complementing agent swarms that execute heterogeneous subtasks in parallel (Kimi Team, 2026).The paper positions it as consolidation of parallel attempts at the same task for one fresh execution.Source: §5 · p. 10
  • Unlike independent rollout summaries concatenated for new rollouts (Madaan et al., 2025; Kim et al., 2026), ReGE-Syn jointly chooses content from retained source records before producing guidance.Source: §5 · p. 10
  • Prior reflection and memory methods reuse feedback, insights, workflows, guidelines, manuals, skills, or linked notes (Wang et al., 2024c; Fu et al., 2024); ReGE focuses on records from multiple attempts at the same problem for a new execution.Source: §5 · p. 10

6 LIMITATIONS AND CONCLUSION

The paper concludes that recorded attempts can guide fresh execution, while noting limits in benchmark coverage and ablation scope. Its reported safety context uses public benchmarks without human subjects or real user accounts.

6 LIMITATIONS AND CONCLUSION

  • The evaluation covers four benchmarks, with ablations on selected subsets; the authors call for broader task settings and repeated executions with aligned settings.Source: §6 · p. 10
  • ReGE turns recorded attempts into fresh-execution guidance, and its joint synthesis improves mean task completion while keeping the base model fixed.On matched BrowseComp-Plus subsets, it reduces guidance-construction tokens relative to adapted iterative aggregation and lowers combined construction and execution usage, excluding source generation.Source: §6 · p. 10
  • Generative AI tools assisted with manuscript language polishing and textual revisions, while the authors retain responsibility for the final content.Source: AI USE STATEMENT · p. 10
  • The benchmarks involve no human subjects or real user accounts, and the authors report no direct potential for harm beyond that of stronger autonomous agents generally.Source: ETHICS STATEMENT · p. 10

A EXPERIMENTAL PROTOCOLS

The protocols specify what source attempts contain and how they are bounded, then describe the prompts and delivery formats used to construct or select guidance. Students receive guidance in a reset environment and use current observations during execution.

A.1 SOURCE REPRESENTATION

  • Source artifacts include produced files, submitted SQL and query results, submitted answers, and AndroidWorld action narratives and final action summaries; empty or non-submitting attempts remain in the pool.Source: §A.1 · p. 13
  • ReGE-Syn, ReGE-Sel, and ReGE-A cap each source at 120 rendered steps, truncate per-step reasoning and commands, cap artifacts at 6,000 characters, and remove artifact bodies if serialized sources exceed 350,000 characters.Direct history instead uses step-only rendering, omitting tool observations and separate final-artifact blocks.Source: §A.1 · p. 13

A.2 PROMPTS AND GUIDANCE DELIVERY

  • The ReGE-Syn coach receives the task and source pool, then writes a self-contained executable plan with concrete steps, discovered facts, and pitfalls for a fresh agent that never sees the attempts.Source: §A.2 · p. 13
  • Alternative constructions request a self-contained explanation and answer in exactly two XML sections, while ReGE-Sel asks a judge to choose the rollout most likely fully correct without success labels.ReGE-Sel permits up to eight parsing attempts and selects the first source if all attempts fail.Source: §A.2 · p. 13
  • Guidance is delivered before instructions on Terminal-Bench 2.0 and Spider2-Lite, through AndroidWorld’s native guideline interface, or appended to BrowseComp-Plus user messages; source pools use the same model.Students are told the environment has reset and to use current observations and verify completion.Source: §A.2 · p. 14
  • Independent-summary calls summarize bounded source blocks into task requirements, actions, artifact changes, and commands; retries enforce shorter outputs, with no quality-based selection or raw-trajectory fallback.Usage counts all recorded calls for each delivered summary, including original generation when a cached summary is reused.Source: §A.2 · p. 14

A.3 EVALUATION AND TOKEN ACCOUNTING

The protocol distinguishes requested generation limits from serving evidence and measures realized usage under each method’s construction and stopping procedures.

A.3 EVALUATION AND TOKEN ACCOUNTING

  • Table 11 reports requested generation limits alongside available serving evidence; comparisons use realized usage under method-specific construction and stopping procedures.Source: Table 11 · p. 14, §A.3 · p. 14

B.1 SOURCE-POOL CHARACTERISTICS

The appendix defines source-pool coverage and examines how source availability relates to fresh-student success. It also describes conditional-success and repeated-execution analyses and specifies how student execution is accounted for.

B.1 SOURCE-POOL CHARACTERISTICS

  • Coverage is the fraction of tasks with at least one successful source attempt, calculated using all source pools.Later student analyses use common valid execution records.Source: §B.1 · p. 15, Table 12 · p. 15
  • With one, two, four, and six sources, student success is 81%, 82%, 85%, and 89%, respectively, versus 70% for Control; six versus four gives +4/−0 (p = .125).The separate 100-query BrowseComp-Plus subset had 87% source-pool coverage.Source: §B.1 · p. 15

B.2 RE-EXECUTION FROM A SUCCESSFUL SELECTED SOURCE

  • Table 13 reports fresh-student success conditional on a judge-selected successful source, using the same selected subset for every method.The table caption defines Tasks (%) as the share of benchmark tasks with a successful selected source and other columns as fresh-student success rates.Source: §B.2 · p. 15, Table 13 · p. 15

B.3 REPEATED STUDENT EXECUTIONS

  • On 35 Spider2-Lite tasks with mixed four-source pools, Control and ReGE-Syn each have three claude-sonnet-5 student executions; mean success averages all three.The other two rows are percentages of the same 35 tasks; the supplied passages do not report the outcome values.Source: §B.3 · p. 15, Table 14 · p. 15

C STUDENT EXECUTION ACCOUNTING

  • Table 4 sums recorded model output tokens per student execution and counts task-specific trace entries, SQL calls, or agent steps.Ratios divide method means on the common valid task set, and the calculation uses output tokens only.Source: C STUDENT EXECUTION ACCOUNTING · p. 16

D COMPARISON WITH OTHER MODELS

The appendix specifies task selection and execution conditions for model comparisons, then describes the recorded cases and their starting-condition and record-linkage documentation.

D COMPARISON WITH OTHER MODELS

  • The comparison selects tasks by alphabetical or instance order, independently of results, and uses 300 BrowseComp-Plus ablation queries.AndroidWorld uses each template’s first parameterization, also used for claude-sonnet-5.Source: D COMPARISON WITH OTHER MODELS · p. 16
  • gpt-5.6-terra, gpt-5.6-sol, and gpt-6-luna each ran once per task without guidance, using the prompts, tools, step limits, and graders of the main gpt-5.6-luna Control runs.Each used serving default-medium effort. Listed input/output token prices per million tokens are 0.20/1.20 for gpt-5.6-luna, 2.00/12.00 for gpt-5.6-terra, 4.00/20.00 for gpt-5.6-sol, and 0.10/0.50 for gpt-6-luna.Source: D COMPARISON WITH OTHER MODELS · p. 16

E ADDITIONAL EXECUTION CASES AND EVIDENCE

E.1 EVIDENCE FOR THE MAIN-TEXT CASES

E.2 ADAPTING A DATABASE QUERY TO STORED VALUES

In the database case, the student adapts query preprocessing to stored year formats while retaining the guidance’s query structure and reaches an accepted result. In the document-research case, the student verifies a recorded lead and succeeds.

E.2 ADAPTING A DATABASE QUERY TO STORED VALUES

  • The four recorded sources produce different counts after varying how they parse Movie.year and bound gaps between credited years.The synthesized guidance retains distinct actor–year pairs and a LAG calculation, recommends inspecting stored formats, and uses a four-year maximum gap.Source: §E.2 · p. 16
  • 28,699 actors matches an accepted official result after the student adapts year preprocessing and retains the proposed gap calculation.The final query uses CAST(substr(year,-4) AS INTEGER) rather than the guide’s strict raw-string filter, after checking nonstandard year strings, nulls, and invalid suffixes.Source: §E.2 · p. 17

E.3 REUSING DOCUMENT LEADS FOR FRESH VERIFICATION

  • One recorded attempt reaches the Citrus Experiment Station and Herbert J. Webber before submitting Leon Dexter Batchelor; two others find relevant documents but submit no answer.The synthesized guide prioritizes the lead, retains document IDs 9606 and 26507, and flags clues still requiring checking.Source: §E.3 · p. 17
  • The student submits “Leon Dexter Batchelor” and receives an exact-match success grade after six recorded tool steps, including submission.The student checks document 9606 and Webber’s biography in document 26507; the guide preserves those document locations and a candidate answer for verification.Source: §E.3 · p. 17

Related papers

Loading 2610.33861v1…