PaperlayerView exact revision on arXiv

Source-linked AI summary

Sample-Efficient Learning from Agent Experience

Chenhui Gou, Haoqin Tu, Yunhao Fang, Jianfei Cai, Hamid Rezatofighi

arXiv:2607.21051v1cs.CL

TL;DR

Agents can learn efficiently from interaction histories in context, but those gains vanish when the context is removed. Experience Distillation internalizes these gains into model weights without further environment interaction, retaining at least 64.8% of in-context learning gains across software-engineering and text-adventure tasks.

  • Problem

    How to internalize agents’ interaction-history learning into model weights without sacrificing environment sample efficiency remains underexplored.

  • Method

    Experience Distillation resamples teacher decisions at histories already observed in collected experience to train a context-free student without further environment interaction.

  • Results

    64.8% of in-context learning gains were retained across 749 software-engineering tasks and six text-adventure games, versus 3.8% recovered by direct supervised fine-tuning.

  • Takeaways & Limitations

    Experience Distillation matches classical reinforcement-learning baselines with at least 9.6× fewer environment samples while transferring distilled capabilities to out-of-distribution tasks.

  • Takeaways & Limitations

    Long synthetic trajectories can provide poor supervision because world-model errors compound and out-of-support predictions become unreliable.

Abstract

from arXiv · show

Real-world agent learning is often constrained by costly environment interactions, such as running time-consuming experiments or obtaining human feedback. In-context learning offers a highly sample-efficient way for agents to learn from their own interaction histories, but its gains disappear once that experience is removed from the context. Separately, context distillation provides a mechanism for internalizing contextual information into model weights. However, applying it to agents' interaction histories without sacrificing environment sample efficiency remains underexplored. We term this problem Experience Distillation and develop an implementation that requires no further environment interaction beyond the collected experience. Experiments on 749 curated software-engineering tasks and six text-adventure games show that it retains at least 64.8\% of the gains from in-context learning across both domains, whereas direct supervised fine-tuning on the collected experience recovers only 3.8\%. Compared with classical reinforcement-learning baselines, in-context learning from trial-and-error experience followed by Experience Distillation matches their performance with at least \(9.6\times\) fewer environment samples.

1 INTRODUCTION

Experience Distillation addresses the high cost of real-world agent interaction by internalizing interaction histories without further environment samples. Across software-engineering and text-adventure tasks, it preserves substantially more of in-context learning’s gains than direct supervised fine-tuning.

  • Motivation: Real-world rollouts are increasingly costly as agents tackle longer, more realistic tasks requiring substantial time, resources, or human expertise.Examples include laboratory testing of battery materials and qualified-lawyer review of legal analyses.
  • Motivation: In-context learning adapts agents from task-specific demonstrations or prior-attempt feedback without large training sets or extensive environment interaction.Its gains disappear once the contextual experience is removed because ICL does not update model parameters.
  • Method: World-model rollouts are proposed as a way to avoid further environment interaction, but compounding errors over long rollouts can make generated trajectories unreliable.Branched rollouts are described as reducing this problem by beginning from points sampled along collected trajectories.
  • Evaluation: 749 software-engineering tasks and six text-adventure games provide the evaluation setting for Experience Distillation.Experience contexts typically contain roughly 60–600 interaction turns and over 80k tokens, while one multi-task run aggregates up to 61.7M experience tokens.
  • Results: 64.8% of ICL’s performance gains are retained after Experience Distillation, versus 3.8% after direct SFT on the same experience.The comparison is reported across the software-engineering and text-adventure domains.

2 PRELIMINARIES

This section defines context distillation as distilling knowledge from privileged context into a student model, and formalizes interactive agent tasks as POMDPs with observable histories and trajectory-prefix distributions.

  • Context Distillation: Context distillation transfers information from teacher-only auxiliary context, such as demonstrations, documents, or memories, to a student model.The teacher models pθ0(y | x, c), where c is unavailable to the student.
  • Interactive Agent Tasks: Interactive agent tasks are modeled as partially observable Markov decision processes in which actions depend on the observable history.At step t, the agent receives observation o_t and selects action a_t based on h_t = (o_0, a_0, ..., o_t).
  • Interactive Agent Tasks: Rewards and other environment feedback are included in observations, and data collection yields a finite observable trajectory prefix τ of length T.The prefix is τ = (o_0, a_0, ..., a_T−1, o_T).
  • Interactive Agent Tasks: A policy πθ interacting with environment M induces a trajectory-prefix distribution determined by the initial-observation and history-conditioned next-observation distributions.These environment distributions marginalize over latent states.

3 EXPERIENCE DISTILLATION

Experience Distillation internalizes collected interaction histories into model weights through trajectory-level context distillation, using one-step branched rollouts without additional environment interaction or a world model. The implementation preprocesses experience, elicits enhanced teacher reasoning, and packs branches to retain distillation performance while reducing computation.

  • Trajectory-Level Distillation: The method formulates experience internalization as trajectory-level context distillation, implemented with one-step branched rollouts requiring neither further environment interaction nor a world model.The objective reduces to local policy matching between teacher and student distributions, though direct estimation would otherwise require teacher rollouts.
  • Sample Efficiency: Distilling recorded experience under the one-step objective requires no additional environment interaction, preserving environment sample efficiency.The approach replaces costly additional teacher trajectories with local teacher decisions generated from recorded histories.
  • One-Step Branched Rollouts: World-model rollouts are avoided because long synthetic trajectories compound prediction errors and become unreliable outside real-data support.Branched rollouts instead begin at recorded branch points and simulate only short continuations; the implementation uses k = 1, eliminating the world model.
  • Implementation: Experience preprocessing removes noisy, repetitive, irrelevant, or overly long history content before teacher generation, while a fixed prompt elicits more thorough reasoning.The prompt instructs the teacher to examine the preprocessed experience and reason more thoroughly before deciding.
  • Implementation: Branch packing combines successive branch points from one trajectory into a single sequence, using recorded context while applying loss only to teacher-generated decisions.This practical approximation retains distillation performance while substantially reducing teacher-generation and training time.

4 EXPERIMENTS

Experiments across curated software-engineering tasks and TaleSuite text-adventure games evaluate Experience Distillation’s ability to retain in-context learning gains, reduce environment samples, and improve efficiency and transfer. Experience Distillation consistently outperforms direct supervised fine-tuning and approaches or matches strong baselines across multiple ablations.

  • Experimental setup: Experiments use 749 curated software-engineering tasks and six TaleSuite text-adventure tasks, with task-specific multi-trial experience collected through repeated attempts.SWE trials preserve interaction history and the evolving working tree, while TaleSuite resets game state between trials.
  • Experience Distillation: On curated SWE, Experience Distillation reaches 51.4% average pass@1 and retains 64.8% of the ICL gain; on TaleSuite, it reaches 43.8 and retains 93.4%.SFT instead achieves GICL = 3.8% and −2.6% on curated SWE and TaleSuite, respectively.
  • Sample efficiency: 51.4% average pass@1 on curated SWE versus 17.7% for PPO uses 9.6× fewer environment samples; TaleSuite reaches 43.8 versus 29.9 for GRPO with 57.2× fewer samples.The comparison evaluates ICL followed by Experience Distillation against classical reinforcement-learning baselines.
  • Distillation variants: 83.8% average task-level GICL across five tasks for model-free one-step distillation exceeds 57.3% for model-based branch rollout.On Detective, branch rollout raises GICL from 19.2% to 68.1%, and removing the world-model observation raises it to 86.7%.
  • Branch packing: Branch packing replaces 4096 branch examples with 128 packed sequences, reduces training steps from 768 to 64, and changes average normalized score from 43.1 to 43.8.Average task-level GICL changes from 84.2% to 90.1%, while combined teacher-generation and student-training time falls by more than an order of magnitude.
  • Transfer and accumulation: Experience Distillation improves OOD SWE pass@1 from 4.62% to 8.84% and pass@5 from 20.39% to 26.13%, while five cycles raise TaleSuite score from 7.1 to 47.0.The OOD gains include cross-repository tasks, and the cyclic result uses 20 cumulative environment samples per task.

5 RELATED WORK

Related work frames Experience Distillation at the intersection of context distillation, in-context learning, and sample-efficient reinforcement learning. Context distillation transfers context-induced behavior into model parameters, while reinforcement-learning methods reduce environment interactions by reusing data, modeling rollouts, or learning offline.

  • Context distillation: Context distillation transfers behavior induced by informative contexts into model parameters, enabling students to reproduce context-conditioned teachers without repeatedly consuming those contexts.The approach builds on knowledge distillation and has been applied to prompt-conditioned behavior, instructions, demonstrations, and scratchpads.
  • In-context learning: In-context learning lets fixed-parameter models adapt from demonstrations and other task-specific evidence provided at inference time.Prior work studies mechanisms including implicit Bayesian inference, induction heads, and learning algorithms implemented in transformer activations.
  • Sample-efficient reinforcement learning: Reinforcement-learning methods improve sample efficiency through off-policy learning and experience replay, learned world models for synthetic rollouts, or fixed offline datasets.These approaches address the longstanding challenge of achieving strong performance with fewer environment interactions.

6 CONCLUSION … A.2 EFFECT OF LEARNING RATE

The paper concludes that Experience Distillation consolidates interaction-history learning into model weights without further environment interaction or a world model. Ablations show that enhanced teacher reasoning improves distillation outcomes, while smaller learning rates with longer training produce stronger evaluation performance than final training loss alone predicts.

  • 6 CONCLUSION: Experience Distillation consolidates agents’ interaction-history learning into model weights while requiring neither further environment interaction nor a world model.The implementation resamples only the teacher’s next decision at histories already observed in collected experience.
  • 6 CONCLUSION: The method retains substantial in-context-learning gains across software-engineering tasks and text-adventure games and clearly outperforms dire…
  • APPENDIX OVERVIEW: Appendix A adds Experience Distillation ablations, frontier-model references, curated SWE pass@10 results, qualitative diagnostics, and detailed experience-collection statistics.
  • A.1 EFFECT OF ENHANCED TEACHER REASONING: The ablation evaluates enhanced teacher reasoning by disabling long teacher reasoning while holding the teacher-generation budget fixed, using GICL to map zeroshot to 0% and ICL to 100%.
  • A.1 EFFECT OF ENHANCED TEACHER REASONING: Enhanced Teacher Reasoning raises Detective’s normalized score from 50.7 to 71.8 and GICL from 34.9% to 72.5%.Both variants use 4096 teacher rollouts, three rollout steps, and one to five collected histories in context.
  • A.2 EFFECT OF LEARNING RATE: The smaller learning rate with more epochs yields the strongest evaluation performance among the evaluated schedules.At similar training losses around 0.15–0.16, the 10^-6 run after 8 epochs substantially outperforms the 10^-5 run after 4 epochs.
  • A.2 EFFECT OF LEARNING RATE: Further fitting at 10^-5 lowers training loss but degrades evaluation performance, whereas 10^-6 checkpoints improve from 4 to 16 epochs as training loss decreases.Final training loss alone therefore does not determine distillation quality.

A.3 TALESUITE FRONTIER-MODEL REFERENCE RESULTS

Under a common six-task TaleSuite protocol, ICL followed by Experience Distillation substantially improves the model and reaches the performance range of frontier references without the experience context. Its average normalized score is comparable to Claude Opus 4.5 and exceeds the other listed frontier models.

  • Results: 43.77 average normalized score follows ICL plus Experience Distillation, up from 18.50.The trained model is evaluated without the experience context.
  • Results: 43.77 is comparable to Claude Opus 4.5 at 43.76 under the six-task protocol.The comparison uses the normalization defined in Appendix A.7.
  • Results: ICL plus Experience Distillation exceeds the other frontier-model references in Table 8.Those frontier-model rows are inference-only zeroshot evaluations under the same six-task TaleSuite protocol.

A.4 CURATED SWE PASS@10 RESULTS

Table 9 reports pass@1 and pass@10 results on 749 curated software-engineering tasks. Pass@10 evaluates ten independently sampled candidates per task, with only ICL receiving task-specific experience context at evaluation.

  • Table 9 evaluates performance on 749 curated SWE tasks using both pass@1 and pass@10.Pass@10 measures whether at least one of ten independently sampled candidates per task is accepted.
  • Pass@10 is computed from ten independently sampled candidates for each task.A task counts as successful when at least one candidate is accepted.
  • Only the ICL reference receives task-specific experience context during evaluation.

A.5 OBSERVED ERRORS IN MODEL-BASED ROLLOUTS · A.6 BEHAVIOR AFTER ON-POLICY DISTILLATION ON TALESUITE

Model-based rollouts can propagate fabricated rewards and policy text into later decisions, creating increasingly off-distribution prefixes. After on-policy distillation, TaleSuite outputs instead show repeated self-revision, generic actions, and unstable task-specific behavior.

  • A.5 OBSERVED ERRORS IN MODEL-BASED ROLLOUTS: A world-model observation assigns ten points to the unsupported action dance, prompting the teacher to select another unsupported action, sing.The error changes the subsequent policy distribution rather than remaining isolated.
  • A.5 OBSERVED ERRORS IN MODEL-BASED ROLLOUTS: A generated observation can embed answer-tagged policy output, breaking the separation between environment observations and policy decisions.Subsequent teacher decisions in a multi-step rollout then condition on the contaminated prefix.
  • A.5 OBSERVED ERRORS IN MODEL-BASED ROLLOUTS: Longer model-based rollouts provide more opportunities for deviations from real dynamics or role structure to compound, unlike model-free one-step distillation.These examples illustrate the failure mode but do not estimate its frequency.
  • A.6 BEHAVIOR AFTER ON-POLICY DISTILLATION ON TALESUITE: After on-policy distillation, the Detective model repeatedly alternates between take paper and the task-appropriate read paper while inserting many wait tokens.The opening Director’s Office state makes read paper the task-appropriate action that advances the episode.
  • A.6 BEHAVIOR AFTER ON-POLICY DISTILLATION ON TALESUITE: In Balances, the model cycles among generic actions such as look, look outside, and examine furniture before selecting examine furniture after extended self-correction.The output reflects no stable task-specific plan.
  • A.6 BEHAVIOR AFTER ON-POLICY DISTILLATION ON TALESUITE: Both TaleSuite examples exhibit repeated self-revision and conflicting or generic action selection after OPD training.This pattern is consistent with student sampling: reverse-KL supervision on student-generated trajectories cannot directly include task-specific decisions the student does not generate.

A.7 EXPERIMENTAL DETAILS AND EXPERIENCE STATISTICS

The section defines TaleSuite score normalization and reports how repeated-trial experience is collected and sized across TaleSuite and 749 curated SWE tasks. It also distinguishes aggregate corpus totals used for multi-task teacher generation from individual context lengths.

  • TaleSuite score normalization: TaleSuite scores are normalized per method and task using the achieved score and task maximum, then averaged arithmetically across tasks.Cross-task results use arithmetic means of task-level normalized scores.
  • TaleSuite experience: TaleSuite experience uses repeated trials on the same task, allowing up to 100 interaction steps per trial while retaining accumulated experience after each reset.The game state resets between trials, but the accumulated experience remains in context.
  • TaleSuite experience: Detective records contain 288 interaction turns and 40.4k tokens across six trials, while Inhumane records contain 1,212 turns and 143.9k tokens across twelve trials.These are complete multi-trial task-level records used in the main TaleSuite comparison.
  • Curated SWE experience: Curated SWE experience is selected from 8–12 independent rollout processes, retaining trajectories that reach an accepted commit after more than one trial.Each process permits up to 10 trials; when multiple trajectories qualify, one is selected uniformly at random.
  • Experience statistics: 60.5 interaction turns and 82.4k tokens are the average curated SWE experience lengths, with maxima of 140 turns and 130.0k tokens across 749 tasks.Aggregate corpus totals are 502.0k tokens across six TaleSuite tasks and 61.7M tokens across 749 curated SWE tasks, representing separate task-level histories.

B ADDITIONAL ANALYSIS OF RL BASELINES IN EXPERIENCE-LEARNING TASKS · B.1 GRPO LEARNING DYNAMICS ON TALESUITE

The additional analysis explains the RL-baseline comparison through complete GRPO/PPO trajectories, task-level dynamics, exploration diagnostics, and matched responses. On six TaleSuite tasks, GRPO improves unevenly: it peaks at 29.9 after 240 environment samples but ends at 26.0 after 448, with gains concentrated in some tasks and Reverb remaining at zero.

  • B ADDITIONAL ANALYSIS OF RL BASELINES IN EXPERIENCE-LEARNING TASKS: The analysis examines complete GRPO and PPO trajectories, task-level GRPO dynamics, controlled exploration, and matched TaleSuite responses after GRPO and Experience Distillation.These analyses are intended to explain the aggregate RL-baseline comparison through learning dynamics and observed behavior.
  • B.1 GRPO LEARNING DYNAMICS ON TALESUITE: 29.9 is the best observed six-task mean at GRPO update 16 and 240 environment samples, while the trajectory ends at 26.0 after 448 samples.The six-task mean is computed over the same six TaleSuite tasks used in Figure 1.
  • B.1 GRPO LEARNING DYNAMICS ON TALESUITE: GRPO learning is highly non-monotonic, indicating that its aggregate improvement remains variable throughout training.The curve rises above its initial level but does not improve steadily.
  • B.1 GRPO LEARNING DYNAMICS ON TALESUITE: Figure 5 reports mean normalized score by GRPO update and by cumulative environment samples, connecting adjacent complete checkpoints and marking omitted incomplete checkpoints with dashed lines.The outlined diamond identifies the best observed complete checkpoint used in Figure 1.
  • B.1 GRPO LEARNING DYNAMICS ON TALESUITE: Task-level curves show that the aggregate improvement is not shared uniformly across the six TaleSuite tasks, with every available checkpoint observed rather than imputed.Missing checkpoints remain unobserved in the task-level disaggregation.
  • B.1 GRPO LEARNING DYNAMICS ON TALESUITE: Balances trends upward but remains variable; Enter stays at a higher score level while fluctuating, Inhumane and Library are non-monotonic, Detective stays near its initial range, and Reverb remains at zero.Detective shows little sustained GRPO improvement despite a large zeroshot-to-ICL gap when task-specific experience is available.

B.2 COLORBUTTON DIAGNOSTIC OF ON-POLICY EXPLORATION · B.3 TALESUITE BEHAVIOR AFTER GRPO AND EXPERIENCE DISTILLATION

ColorButton shows that GRPO’s reward-driven learning depends on sampling successful trajectories and informative reward variation, whereas large on-policy groups do not guarantee exploration coverage. In TaleSuite, Experience Distillation uses trial-specific consequences in decisions more reliably than GRPO training, including rejecting known failures and planning around learned locations and rewards.

  • B.2 COLORBUTTON DIAGNOSTIC OF ON-POLICY EXPLORATION: ColorButton contains 18 complete three-button sequences, with reward granted only for one hidden passcode after the final selection.No intermediate feedback identifies which earlier decision was incorrect.
  • B.2 COLORBUTTON DIAGNOSTIC OF ON-POLICY EXPLORATION: An ideal experience-accumulating agent discovers the passcode within at most 18 episodes by enumerating sequences without replacement.This guarantee comes from retaining failures rather than from high initial probability on the successful sequence.
  • B.2 COLORBUTTON DIAGNOSTIC OF ON-POLICY EXPLORATION: Even G = 256 need not cover the passcode, because GRPO samples on-policy with replacement and increasing G helps only when p⋆ > 0.Large best-of-N or pass@k-style budgets therefore do not imply behavioral coverage.
  • B.2 COLORBUTTON DIAGNOSTIC OF ON-POLICY EXPLORATION: bAi = 0 implies Jreward(θ) = 0 and ∇θJreward(θ) = 0 when every sampled trajectory misses the passcode.The task-reward update can reinforce the passcode only after a group contains informative reward variation; auxiliary KL terms do not identify the rewarding unseen sequence.
  • B.2 COLORBUTTON DIAGNOSTIC OF ON-POLICY EXPLORATION: 60% initial success with G = 16 rose to 100% within eight updates, whereas G = 256 still achieved zero success for the red–red–green passcode.The contrast shows that GRPO’s sample needs depend on initial policy mass p⋆, not only on the finite action-space size.
  • EXAMPLE 1: SAME IMMEDIATE ACTION, DIFFERENT USE OF EXPERIENCE: In Example 1, both models read the paper, but only Experience Distillation connects that action to downstream locations and rewards learned from earlier trials.The distilled output links reading the document to unlocking movement and to a westward route containing a crucial pistol.
  • B.3 TALESUITE BEHAVIOR AFTER GRPO AND EXPERIENCE DISTILLATION: Across the selected TaleSuite states, Experience Distillation explicitly uses earlier locations, rewards, and failure consequences, whereas GRPO produces locally plausible reasoning without the same trial-specific use.These examples illustrate behavioral differences; aggregate tables and learning curves provide the performance evidence.
  • EXAMPLE 2: EXPERIENCE CHANGES THE NEXT ACTION: In Example 2, GRPO acknowledges that east is a dead end but proposes checking it, while Experience Distillation rejects east and plans west using prior failures and clue locations.The distilled output also incorporates consequences associated with previously visited locations, including immediate death at the restaurant and clues at the mayor’s house.

B.4 PPO LEARNING DYNAMICS ON CURATED SWE TASKS … C.3.3 TEACHER-GENERATED DISTILLATION TARGETS

PPO reaches its best curated-SWE result only at the largest evaluated sample cost, while Experience Distillation retains substantial ICL gains without experience context. Eight case studies further show that distilled targets can encode task-specific repair knowledge, including accepted and rejected hypotheses.

  • B.4 PPO LEARNING DYNAMICS ON CURATED SWE TASKS: 17.74% pass@1 is PPO’s best supplied checkpoint at step 375, requiring 504.9 environment samples per task.PPO improves gradually overall, with occasional local declines, and its final checkpoint is also its best observed result.
  • C QUALITATIVE ANALYSIS OF EXPERIENCE DISTILLATION ON CURATED SWE TASKS: Experience-conditioned ICL raises curated-SWE average pass@1 from 5.3% to 76.4%, while Experience Distillation reaches 51.4% without experience context.These aggregate results establish overall performance but do not identify the task-specific repair knowledge underlying the gains.
  • C QUALITATIVE ANALYSIS OF EXPERIENCE DISTILLATION ON CURATED SWE TASKS: Across eight case studies, the analysis tests whether accumulated trials supply repair knowledge through ICL and whether EPD-trained models reproduce it without experience context.The cases provide diagnostic contrasts rather than prevalence estimates; aggregate pass@1 remains the evidence for overall effectiveness.
  • C.1 SWE CASE-STUDY PROTOCOL: Every evaluation candidate starts from the original repository and receives the same issue, while only the ICL reference additionally receives repeated-trial history.Zeroshot, SFT, PPO, and EPD are evaluated without that history.
  • C.2 CASE PORTFOLIO AND OUTCOMES: The eight cases span distinct repair patterns, with Table 13 describing task context and collection outcomes and Table 14 reporting accepted candidates under pass@10.The portfolio includes implementation language, task type, repeated-trial outcomes, and evaluation outcomes.
  • C.2 CASE PORTFOLIO AND OUTCOMES: 79/80 ICL candidates are accepted, compared with 72/79 valid EPD candidates (91.1%), while zeroshot, SFT, and PPO produce none on the selected tasks.ICL and EPD solve all eight selected cases; the Auth0 case has nine valid EPD candidates.
  • C.3.1 TASK AND REPAIR OBJECTIVE: The MapStore2 repair objective is to make the feature-grid x control remove a geometry constraint after the advanced-filter panel is opened and closed.MapStore2 is a JavaScript geospatial-layer application whose feature grid supports map-geometry filtering.
  • C.3.2 ACCUMULATED EXPERIENCE AND EXPERIENCE-CONDITIONED ICL: Ten zeroshot samples identify related state or geometry issues, but none composes an actionable x-control reset; all ten sampled ICL candidates produce accepted patches.The successful trajectory combines an explicit null reset, deactivation state, and propagation of that reset to refresh query results.

C.3.4 EPD EVALUATION WITHOUT EXPERIENCE CONTEXT … C.10.4 EPD EVALUATION WITHOUT EXPERIENCE CONTEXT

Across seven software-engineering case studies, Experience Distillation preserves experience-conditioned repair behavior when the collected experience is absent at evaluation. The distilled patches reproduce multi-component fixes and achieve high acceptance rates, generally matching or closely approaching ICL while outperforming zeroshot, SFT, and PPO.

  • C.3.4 EPD EVALUATION WITHOUT EXPERIENCE CONTEXT: EPD reproduces the coordinated MapStore2 repair without experience context: all ten EPD candidates are accepted, versus zero of ten zeroshot candidates.The selected patch emits a null-valued geometry reset, preserves it through the epic, and clears surviving geometry before reopening the grid.
  • C.4 CASE STUDY 2 ON CURATED SWE TASK: LINKED TOOLTIPS WITH NONMATCHING TIME AXES: All ten ICL candidates are accepted for linked tooltips, while nine of ten EPD candidates are accepted without experience context.EPD combines representation-aware nearest-x lookup, explicit absence handling, linked-chart propagation, and a downstream event-rectangle guard.
  • C.5 CASE STUDY 3 ON CURATED SWE TASK: LIVE RESOLVER CLEANUP AFTER COMPONENT UNMOUNT: Nine of ten EPD and all ten ICL candidates are accepted for Relay cleanup, whereas zeroshot, SFT, and PPO each yield zero of ten.The distilled repair separates cache-retained references from active references, unsubscribing inactive live state while preserving release-buffer records.
  • C.6 CASE STUDY 4 ON CURATED SWE TASK: OVERSIZED SESSION COOKIES IN THE EDGE RUNTIME: Eight of nine retained EPD candidates are accepted for Edge-Runtime cookies, compared with ten of ten ICL candidates and zero of ten zeroshot, SFT, and PPO candidates.The patch fixes both protocol directions by recognizing cookie name/value boundaries, preserving attribute commas, and appending reconstructed cookies separately.
  • C.7 CASE STUDY 5 ON CURATED SWE TASK: CONFIGURABLE GRAPHQL ROOT TYPE NAMES: Nine of ten EPD candidates and nine of ten ICL candidates are accepted for configurable GraphQL root names, versus zero of ten for zeroshot, SFT, and PPO.EPD preserves the distinction between configured source-root names and generated wrapper names across query, mutation, and subscription, with 146 tests passed and one skipped.
  • C.8 CASE STUDY 6 ON CURATED SWE TASK: SQUASH-MERGE COMMIT MESSAGE SELECTION: Nine of ten EPD candidates are accepted for squash-merge message selection, compared with ten of ten ICL candidates and zero of ten zeroshot, SFT, and PPO candidates.The distilled patch implements headless defaults, interactive overrides, and propagation from query selection to mutation, passing squash, non-TTY, and interactive tests.
  • C.9 CASE STUDY 7 ON CURATED SWE TASK: STALE DUPLICATE ERROR AFTER CIRCULAR-REFERENCE RECOVERY: EPD reproduces the two-part Gauge repair without experience context: it updates both parser maps through the dictionary abstraction and deduplicates both discovery paths.The parser reports 198 passing tests, and utility and Gauge suites pass while intentional duplicates remain detectable.
  • C.10 CASE STUDY 8 ON CURATED SWE TASK: BACKEND-RELATIVE STORED MEDIA URLS IN STRAPI: Nine of ten EPD candidates are accepted for Strapi media URLs, compared with ten of ten ICL candidates and zero of ten zeroshot, SFT, and PPO candidates.The patch prefixes stored paths with the backend helper, guards provider parsing, reuses the helper for filename extraction, and leaves user-input validation unchanged.

C.11 CROSS-CASE FINDINGS AND LIMITATIONS

Across eight selected cases, experience-conditioned ICL and EPD produce task-specific repair behavior and solve all cases, unlike zeroshot, SFT, and PPO. The qualitative examples document observed patterns but do not establish causality or overall frequency.

  • Cross-case findings: Across eight cases, repeated-trial experience changes the base model’s repair behavior, while issue-and-repository-only zeroshot samples produce no accepted repair.ICL and EPD samples implement task-specific behaviors summarized in the cross-case comparison.
  • Cross-case findings: 79 accepted candidates out of 80 let experience-conditioned ICL solve all eight cases, while zeroshot solves none under pass@10.These counts summarize the selected-case outcome contrast.
  • Cross-case findings: 72 accepted candidates out of 79 valid samples let EPD solve all eight cases without the experience context, whereas SFT and PPO produce no accepted candidates in 80 samples each.EPD patches and selected teacher decisions also exhibit task-specific repair behavior.
  • Limitations: Selected teacher excerpts and EPD patches document observed repair patterns but do not identify a unique causal path or estimate how frequently EPD uses each strategy.Because cases were selected for outcome contrast, aggregate pass@1 in Table 1 remains the evidence for overall performance.
Loading 2607.21051v1…