Sample-Efficient Learning from Agent Experience

Motivation

• Environment interaction is increasingly costly for realistic tasks, while reinforcement learning often requires large numbers of interactions to achieve strong performance.The introduction contrasts costly scientific and legal evaluation with settings where interaction is cheap and scalable.Source: §1 · p. 1

Method

• Experience Distillation consolidates experience-conditioned behavior into model weights without requiring further environment interaction.The implementation uses single-step branched rollouts from collected trajectories, stopping after the teacher’s action, and therefore needs neither a world model nor new environment samples.Source: §1 · pp. 1–2

Motivation

• In-context learning adapts agents from demonstrations or prior-attempt feedback but loses its improvement when the context is removed, motivating context distillation.Context distillation trains a context-free student to reproduce a teacher conditioned on informative context.Source: §1 · p. 1

Empirical results

• At least 64.8% of in-context-learning gains are retained after Experience Distillation, versus only 3.8% after direct SFT, while matching classical RL baselines with at least 9.6× fewer environment samples.Results cover 749 curated software-engineering tasks and six text-adventure games.Source: §1 · p. 2

• Context distillation transfers knowledge from a teacher that receives privileged context to a student that observes only the input.The auxiliary context may be a demonstration, document, memory, or other informative signal.Source: §2 · p. 2
• Interactive agent tasks are modeled as partially observable Markov decision processes in which actions depend on the observable history.At step t, the history is h_t = (o_0, a_0, ..., o_t).Source: §2 · p. 2
• Rewards and other environment feedback are included in observations, so data collection yields a finite observable trajectory prefix.The prefix is τ = (o_0, a_0, ..., a_{T−1}, o_T), where T is its length.Source: §2 · p. 2
• A policy’s interaction with the environment induces a distribution over trajectory prefixes determined by the environment’s initial-observation and history-conditioned transition distributions.These environment distributions marginalize over latent states.Source: §2 · p. 2

3.1 NAIVE CONTEXT DISTILLATION FROM EXPERIENCE

• Naive experience distillation reduces trajectory-level context matching to local policy matching, but estimating its objective still requires teacher rollouts in the environment.Teacher and student share initial-observation and environment-response factors, leaving their policy factors to be matched.Source: §3.1 · p. 3

3.2 CONTEXT DISTILLATION WITH 1-STEP BRANCH ROLLOUT

• One-step branched rollouts eliminate the world model and require no additional environment interaction, preserving environment sample efficiency.The method starts from recorded branch points, limits the continuation to k = 1, and distills the teacher’s local decision.Source: §3.2 · p. 3

3.3 PRACTICAL IMPLEMENTATION

• The practical implementation preprocesses noisy histories, elicits longer teacher reasoning, and packs successive branches to increase useful supervision within the context window.Preprocessing concentrates task-relevant information and preserves salient outcomes, while a fixed reasoning prompt asks the teacher to reason more thoroughly before deciding.Source: §3.3 · pp. 3–4
• Table 3 shows that branch packing retains distillation performance while substantially reducing teacher-generation and training time.The approximation packs branches from successive points of one recorded trajectory, with only teacher-generated decisions contributing to the loss.Source: §3.3 · p. 4

4.1 EXPERIMENTAL SETTING

• Experiments use 749 curated SWE tasks and six TaleSuite tasks, collecting task-specific multi-trial experience and evaluating task performance alongside the ICL-normalized gain GICL.ICL retains experience only at evaluation, whereas weight-updating methods are evaluated without that context; Experience Distillation uses a context-conditioned teacher and context-free student.Source: §4.1 · pp. 4–5

4.2 DISTILLING MULTI-TASK EXPERIENCE INTO MODEL WEIGHTS

• Experience Distillation reaches 51.4% pass@1 while retaining 64.8% of ICL gains on SWE, and score 43.8 while retaining 93.4% on TaleSuite, unlike SFT.SFT retains only 3.8% and −2.6% of ICL gains on SWE and TaleSuite, respectively.Source: §4.2 · p. 5, Table 1 · p. 5

4.3 ENVIRONMENT-SAMPLE EFFICIENCY OF ICL FOLLOWED BY EXPERIENCE DISTILLATION

• ICL followed by Experience Distillation achieves 51.4% SWE pass@1 versus PPO’s 17.7% with 9.6× fewer samples, and TaleSuite score 43.8 versus GRPO’s 29.9 with 57.2× fewer.The pipeline therefore reduces environment-sample requirements by approximately an order of magnitude relative to classical RL baselines.Source: §4.3 · p. 6, Figure 1 · p. 6

4.4 EVALUATING MODEL-FREE ONE-STEP DISTILLATION

• Model-free one-step distillation attains 83.8% average task-level GICL versus 57.3% for model-based branch rollout, with shorter rollouts and less world-model participation improving effectiveness.On Detective, GICL rises from 19.2% with full rollout to 68.1% with branch rollout and 86.7% without the world-model observation.Source: Table 2 · p. 6, §4.4 · p. 7

4.5 EFFICIENT BRANCH PACKING FOR LANGUAGE AGENTS

• Branch packing replaces 4096 examples with 128 sequences and 768 steps with 64, cutting combined generation and training time by over an order of magnitude while raising average GICL from 84.2% to 90.1%.Average normalized score changes from 43.1 to 43.8, remaining comparable across six TaleSuite tasks.Source: §4.5 · p. 7, Table 3 · p. 7

4.6 GENERALIZATION TO OOD TASKS / 4.7 CONTINUAL EXPERIENCE DISTILLATION

• Experience Distillation improves OOD SWE pass@1 from 4.62% to 8.84% and pass@5 from 20.39% to 26.13% across 494 tasks, including unseen repositories and task types.Across five continual cycles, mean TaleSuite score rises from 7.1 to 47.0 using 20 cumulative environment samples per task after prior contexts are discarded.Source: §4.6 · p. 8, Table 4 · p. 8, §4.7 · p. 8

4.8 TEACHER-SAMPLED FORWARD KL VERSUS STUDENT-SAMPLED REVERSE KL

• Teacher-sampled forward KL reaches 96.7% and 98.4% GICL on Balances and Detective, whereas student-sampled reverse KL reaches only 9.1% and 0.4%.The reverse-KL variant rarely samples the high-quality experience-induced decisions because the student lacks the experience context.Source: §4.8 · p. 9, Table 5 · p. 9

4.9 SAMPLED-TOKEN VERSUS FULL-DISTRIBUTION DISTILLATION / 4.10 SCALING AND COST OF TEACHER-GENERATED DATA

• Sampled-token NTP slightly outperforms full-distribution KL in mean task-level GICL (84.2% versus 82.0%) while avoiding teacher-logit storage for long agent contexts; scaling teacher-generated data raises GICL from 31.8% to 64.8%.This scaling corresponds to 0.17× to 2.79× the 61.7M-token experience corpus, and teacher-generation cost remains a limitation.Source: §4.9 · p. 9, §4.10 · p. 9

5.1 CONTEXT DISTILLATION

• Experience Distillation targets real agent-environment trajectories and improved future decision behavior, using one-step branch rollouts to create supervised decisions without additional interaction or long world-model rollouts.This distinguishes it from prior context distillation, which transfers context-conditioned behavior, and from methods requiring future trajectory sampling or simulated rollouts.Source: §5.1 · p. 10

5.2 IN-CONTEXT LEARNING

• Experience Distillation consolidates behavioral gains from collected experience into model weights, unlike prior agent methods whose adaptation depends on inference-time context or external memory.The surrounding literature includes recurrent and transformer-based history adaptation, reflections, extracted insights, semantic memories, and reusable workflows.Source: §5.2 · p. 10

5.3 SAMPLE-EFFICIENT REINFORCEMENT LEARNING

• Unlike sample-efficient reinforcement learning, Experience Distillation distills in-context decisions at recorded history points without estimating returns, predicting future observations, or collecting additional environment interactions.Prior approaches reuse data through replay, synthetic rollouts, offline datasets, or richer interaction-feedback supervision, often relying on reward/value estimation or learned dynamics.Source: §5.3 · p. 11

6 CONCLUSION

• With at least 9.6× fewer environment samples, in-context learning followed by Experience Distillation matches classical reinforcement-learning baselines across software-engineering tasks and text-adventure games.The method requires neither further environment interaction nor a world model, and retains substantial in-context-learning gains while outperforming direct supervised fine-tuning.Source: §6 · p. 11
• Experience Distillation resamples only the teacher’s next decision at histories already observed in collected experience, consolidating trial-and-error learning into model weights without additional environment interaction.The conclusion frames this as a paradigm where agents learn quickly in context and periodically consolidate those gains into their weights.Source: §6 · p. 11
• Distilled capabilities transfer to out-of-distribution tasks, and experience-derived gains accumulate across repeated collect-and-distill cycles.These findings extend the reported benefits beyond the original collected experience.Source: §6 · p. 11

APPENDIX OVERVIEW

• The appendices expand the evaluation with ablations, frontier-model references, curated SWE pass@10 results, diagnostics, learning curves, exploration analysis, behavior comparisons, training trajectories, and SWE case studies.The case studies trace task-specific knowledge from accumulated experience through teacher decisions to evaluation behavior without the experience context.Source: Appendix APPENDIX OVERVIEW · p. 16

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; GICL maps zeroshot to 0% and ICL to 100%.Source: §A.1 · p. 16, Table 7 · p. 16

A.2 EFFECT OF LEARNING RATE

• The smaller 10^-6 learning rate with more epochs yields the strongest evaluation performance among the evaluated schedules.At similar training losses around 0.15–0.16, 10^-6 after 8 epochs substantially outperforms 10^-5 after 4 epochs, while further 10^-5 fitting degrades evaluation performance.Source: §A.2 · p. 17, Figure 4 · p. 17
• Final training loss alone does not determine distillation quality: 10^-6 checkpoints improve from 4 to 16 epochs as loss decreases, unlike the overfit 10^-5 schedule.Figure 4 compares learning rates and training durations using reversed training-loss and GICL axes.Source: §A.2 · p. 17, Figure 4 · p. 17

A.3 TALESUITE FRONTIER-MODEL REFERENCE RESULTS

• 43.77 average normalized score matches Claude Opus 4.5 at 43.76 and exceeds the other frontier-model references across six TaleSuite tasks.The trained model is evaluated without the experience context, while frontier-model rows use inference-only zeroshot evaluation under the same protocol.Source: §A.3 · p. 18, Table 8 · p. 18
• Experience Distillation raises the model’s average normalized score from 18.50 to 43.77 under the reported six-task TaleSuite protocol.Scores use the normalization defined in Appendix A.7.Source: §A.3 · p. 18, Table 8 · p. 18

A.4 CURATED SWE PASS@10 RESULTS

• Table 9 reports pass@1 and pass@10 on 749 curated SWE tasks.The table complements the pass@1 comparison in Table 1.Source: Table 9 · p. 18, §A.4 · p. 18
• Pass@10 measures whether at least one of ten independently sampled candidates per task is accepted.Source: Table 9 · p. 18, §A.4 · p. 18
• Only the ICL reference receives task-specific experience context during evaluation.Source: Table 9 · p. 18, §A.4 · p. 18

A.5 OBSERVED ERRORS IN MODEL-BASED ROLLOUTS

• Fabricated rewards in model-based rollouts changed subsequent policy behavior, turning an unsupported dance action into evidence for another unsupported sing action.The world model said dancing was impossible but awarded ten points, which the teacher then used to select sing.Source: §A.5 · p. 19
• Policy-output tokens embedded in generated observations contaminated the role boundary and caused later teacher decisions to condition on an off-distribution prefix.The generated text began as an environment observation after go west but ended with an answer-tagged take small black pistol action.Source: §A.5 · p. 19
• Longer model-based rollouts provide more opportunities for deviations to compound, whereas one-step model-free distillation stops before generating a world-model observation.These examples illustrate the failure mode but do not estimate its frequency.Source: §A.5 · p. 19

A.6 BEHAVIOR AFTER ON-POLICY DISTILLATION ON TALESUITE

• After on-policy distillation, Detective alternated between take paper and task-appropriate read paper while inserting many wait tokens.The opening Director’s Office state required reading the paper before going north or west, but the displayed reasoning repeatedly revised the action.Source: §A.6 · p. 20
• After on-policy distillation, Balances cycled among generic actions such as look, look outside, and examine furniture before selecting examine furniture.The choice followed extended self-correction rather than a stable task-specific plan.Source: §A.6 · p. 20
• Both TaleSuite examples show repeated self-revision and conflicting or generic action selection after on-policy distillation training.The examples document displayed outputs, not the frequency or composition of the training signal.Source: §A.6 · p. 20
• These outputs are consistent with the student-sampling hypothesis: reverse-KL supervision on student-generated trajectories cannot directly include task-specific decisions the student never generates.On-policy distillation samples trajectories from the student without the experience context.Source: §A.6 · p. 20

TaleSuite score normalization

• TaleSuite scores are normalized by each task’s maximum, and cross-task results are arithmetic means of task-level normalized scores.The normalized score uses achieved score Rm,t and task maximum Rmax_t.Source: §A.7 · p. 21

TaleSuite experience

• TaleSuite experience uses repeated trials with up to 100 interaction steps per trial; game state resets between trials while accumulated experience remains in context.The main comparison uses Balances, Detective, Enter, Inhumane, Library, and Reverb, with fixed task subsets within each comparison.Source: §A.7 · p. 21
• Detective records 288 interaction turns and 40.4k tokens, whereas Inhumane records 1,212 turns and 143.9k tokens across their multi-trial histories.These are task-level records used in the main six-task TaleSuite comparison.Source: §A.7 · p. 21

Curated SWE experience

• Curated SWE experiences retain trajectories reaching an accepted commit after more than one trial, selected from 8–12 rollout processes with up to 10 trials.When multiple trajectories qualify, one is selected uniformly at random across the 749 tasks.Source: §A.7 · p. 21
• 60.5 interaction turns and 82.4k tokens are average curated SWE experience lengths, with maxima of 140 turns and 130.0k tokens across 749 tasks.Experience Distillation’s aggregate corpora contain 502.0k tokens across six TaleSuite tasks and 61.7M tokens across 749 curated SWE tasks; these sum separate task histories.Source: §A.7 · pp. 21–22

B ADDITIONAL ANALYSIS OF RL BASELINES IN EXPERIENCE-LEARNING TASKS

• The aggregate RL comparison is analyzed through complete GRPO and PPO trajectories, task-level GRPO dynamics, exploration diagnostics, and matched TaleSuite responses.These analyses connect endpoint sample efficiency with learning dynamics and observed behavior.Source: B ADDITIONAL ANALYSIS OF RL BASELINES… · p. 23

B.1 GRPO LEARNING DYNAMICS ON TALESUITE

• GRPO’s six-task mean averages the same TaleSuite tasks as Figure 1, omits incomplete checkpoints, and uses dashed connectors only across missing observations.The learning curve is shown by update and cumulative environment samples.Source: §B.1 · p. 23, Figure 5 · p. 23
• 29.9 best six-task mean at update 16 and 240 environment samples, falling to 26.0 after 448 samples, indicating non-monotonic GRPO improvement.The best-so-far score is nevertheless associated with the full 448-sample run budget in the endpoint comparison.Source: Figure 5 · p. 23, §B.1 · p. 23
• Task-level GRPO curves preserve observed scores without imputing missing checkpoints, exposing heterogeneous dynamics across the six TaleSuite tasks.Figure 6 disaggregates the same run by task.Source: §B.1 · p. 23, Figure 6 · p. 24
• Balances trends upward but remains variable; Enter fluctuates at a higher level; Inhumane and Library are non-monotonic; Detective stays near its initial range; and Reverb remains at zero.These task-specific patterns show that the aggregate curve does not represent uniform improvement.Source: §B.1 · p. 24
• Detective combines a large zeroshot-to-ICL gap when task-specific experience is available with little sustained GRPO improvement, motivating a controlled exploration test.The comparison is reported in Table 2, while the exploration test examines whether rewarding behavior is unlikely under the initial policy.Source: §B.1 · p. 24

B.2 COLORBUTTON DIAGNOSTIC OF ON-POLICY EXPLORATION

• ColorButton uses 18 three-button sequences with only terminal binary feedback, so failed episodes reveal that the complete sequence was wrong without identifying the erroneous decision.The controlled setup tests whether on-policy sampling discovers the successful sequence needed for a reward-dependent update.Source: §B.2 · p. 25
• Even G = 256 does not ensure coverage of the 18 possible sequences because on-policy sampling is with replacement and increasing G helps only when p⋆ > 0.Thus, large best-of-N or pass@k-style budgets need not expand behavioral coverage, unlike systematic experience accumulation, which can enumerate all sequences within at most 18 episodes.Source: §B.2 · p. 25
• When every on-policy trajectory misses the passcode, the reward-dependent GRPO gradient is exactly zero, so learning requires a group containing informative reward variation.Auxiliary KL terms may still produce gradients, but they do not identify which unseen sequence is rewarding.Source: §B.2 · p. 26
• 60% initial success rose to 100% within eight GRPO updates for green–blue–red, but success stayed at 0% for red–red–green despite G = 256.The required sample count depends on the initial policy mass p⋆ on the rewarding path, not only the finite action-space size.Source: §B.2 · p. 26, Table 12 · p. 26

EXAMPLE 1: SAME IMMEDIATE ACTION, DIFFERENT USE OF EXPERIENCE

• In both TaleSuite states, Experience Distillation preserves the same immediate action while connecting it to downstream locations, rewards, and consequences from earlier trials; GRPO remains local to the current observation.The comparison evaluates explicit task-specific consequences and decisions that use them, rather than response length or fluency.Source: EXAMPLE 1: SAME IMMEDIATE ACTION,… · p. 27, §B.3 · p. 27

EXAMPLE 2: EXPERIENCE CHANGES THE NEXT ACTION

• After GRPO, the model acknowledges an eastern dead end but still proposes checking it, whereas Experience Distillation uses prior failure and clue locations to choose the westward route.The distilled model plans through Main Street to the mayor’s house while avoiding the known deadly restaurant path.Source: EXAMPLE 2: EXPERIENCE CHANGES THE NEXT… · p. 28

B.3 TALESUITE BEHAVIOR AFTER GRPO AND EXPERIENCE DISTILLATION

• Across the selected states, Experience Distillation explicitly uses trial-specific locations, rewards, and failure consequences, while GRPO provides locally plausible reasoning without the same evidence of experience use.These examples illustrate a behavioral contrast; aggregate tables and learning curves supply the performance evidence.Source: EXAMPLE 2: EXPERIENCE CHANGES THE NEXT… · p. 28

B.4 PPO LEARNING DYNAMICS ON CURATED SWE TASKS

• 76.4% pass@1 with experience-conditioned ICL falls to 51.4% after Experience Distillation without experience context, while PPO reaches 17.74% only after 504.9 environment samples per task.The PPO trajectory shows its best supplied checkpoint is the final one, so the sample-efficiency gap is not caused by early-checkpoint selection.Source: §B.4 · p. 29, C QUALITATIVE ANALYSIS OF EXPERIENCE… · p. 30

C QUALITATIVE ANALYSIS OF EXPERIENCE DISTILLATION ON CURATED SWE TASKS

• The qualitative analysis asks whether repeated-trial repair knowledge appears in teacher-generated targets and in EPD patches produced without experience context.Each case compares the issue, accumulated trials and accepted ICL behavior, teacher-sequence excerpts, and one experience-free EPD evaluation patch.Source: C QUALITATIVE ANALYSIS OF EXPERIENCE… · p. 30

C.2 CASE PORTFOLIO AND OUTCOMES

• The eight cases span UI state, interaction logic, resource ownership, protocols, schema transformation, command-line behavior, parser state, and URL handling, with Tables 13–14 summarizing their portfolios and outcomes.The cases provide complementary diagnostic contrasts rather than prevalence estimates.Source: C QUALITATIVE ANALYSIS OF EXPERIENCE… · p. 30, §C.2 · p. 30, Table 13 · p. 31

C.1 SWE CASE-STUDY PROTOCOL

• Evaluation gives every candidate the original repository and task issue, while only the ICL reference receives that task’s repeated-trial history.The protocol therefore isolates behavior produced with experience context from zeroshot, SFT, PPO, and EPD behavior without it.Source: §C.1 · p. 30

C.2 CASE PORTFOLIO AND OUTCOMES

• Across eight selected cases, ICL solves all tasks with 79 accepted candidates out of 80, EPD solves all with 72 out of 79 valid candidates, and zeroshot, SFT, and PPO solve none.These cases are diagnostic rather than prevalence estimates; overall effectiveness is instead measured by aggregate pass@1.Source: §C.2 · p. 31, Table 14 · p. 31

C.3.1 TASK AND REPAIR OBJECTIVE

• In the MapStore2 task, the repair objective is to make the x control remove the geometry constraint and restore unfiltered feature-grid results after an advanced-filter-panel round trip.The failure occurs when the filter remains applied after the panel is opened and closed.Source: §C.3.1 · pp. 31–32

C.3.2 ACCUMULATED EXPERIENCE AND EXPERIENCE-CONDITIONED ICL

• After five rejected sequential trials, Trial 6 combines a component-side value: null reset with a query-pipeline correction, producing the accepted MapStore2 repair; all ten sampled ICL candidates are accepted.The accumulated history includes hypotheses, inspections, edits, tests, verifier outcomes, and evolving working-tree state.Source: §C.3.2 · pp. 32–33

C.3.3 TEACHER-GENERATED DISTILLATION TARGETS

• Teacher Targets 1 and 4 encode the accepted repair by making the control emit an explicit null reset and propagating it through the query epic, whereas Targets 2 and 3 revisit rejected hypotheses.The targets therefore evidence task-specific knowledge in the branch-packed sequence without making every generated decision an independently validated requirement.Source: §C.3.3 · pp. 33–34

C.3.4 EPD EVALUATION WITHOUT EXPERIENCE CONTEXT

• 10/10 sampled EPD candidates produce accepted MapStore2 patches, while none of 10 zeroshot candidates does, retaining the experience-conditioned coordinated repair pattern without context.The patch resets null-valued geometry, preserves the update through the epic, and clears surviving geometry on panel return; external verification passed, but no local browser suite ran.Source: §C.3.4 · pp. 34–35, Table 16 · p. 35

C.4.2 ACCUMULATED EXPERIENCE AND EXPERIENCE-CONDITIONED ICL

• All 10 ICL Billboard.js candidates are accepted after six trials exposing representation-aware lookup, explicit absence handling, and event-rectangle safety as a coordinated cross-layer repair.Zeroshot candidates propose local nearest-x changes but do not coordinate propagation, the public API, and DOM dispatch.Source: §C.4.2 · p. 36

C.4.3 TEACHER-GENERATED DISTILLATION TARGETS

• The teacher-generated Billboard.js targets retain representation-aware lookup and the DOM guard, while omitting the public-tooltip null-index exit and linked-chart precheck included in the EPD patch.The lookup distinguishes binary search on shared sorted x values from full search across per-series values.Source: §C.4.3 · p. 36

C.4.4 EPD EVALUATION WITHOUT EXPERIENCE CONTEXT

• 9/10 EPD candidates are accepted for Billboard.js, versus 10/10 ICL and 0/10 for zeroshot, SFT, and PPO, despite omitting trial experience at evaluation.The EPD patch coordinates representation-aware nearest-x lookup, explicit absence handling, linked-chart propagation, and a downstream event-rectangle guard; the legacy browser suite did not run locally.Source: §C.4.4 · p. 37

C.5.3 TEACHER-GENERATED DISTILLATION TARGETS

• Repeated Relay trials teach EPD to separate cache-retained references from active references, preserving release-buffer records while unsubscribing inactive live subscriptions during garbage collection.The teacher explicitly rejects Trial 1’s all-or-nothing ownership model and adopts the two-set structure reached in Trial 2.Source: §C.5.2 · pp. 38–39, §C.5.3 · p. 39

C.5.4 EPD EVALUATION WITHOUT EXPERIENCE CONTEXT

• 9/10 EPD candidates are accepted for Relay, versus 10/10 ICL and 0/10 for zeroshot, SFT, and PPO, with five LiveResolvers and 162 RelayModernStore tests passing.The repair preserves cache-retained records while unsubscribing inactive live state during garbage collection; verification covered store cleanup rather than the complete React unmount chain.Source: §C.5.4 · pp. 39–40

C.6.3 TEACHER-GENERATED DISTILLATION TARGETS

• Repeated Auth0 trials show that cookie emission alone is insufficient: the accepted protocol reconstructs cookie boundaries while preserving commas inside attributes, then writes separate response-header values.The teacher targets both the omitted read path and the stateful rule requiring a new name/value prefix before starting another cookie.Source: §C.6.2 · p. 40, §C.6.3 · p. 41

C.6.4 EPD EVALUATION WITHOUT EXPERIENCE CONTEXT

• 8/9 retained EPD candidates are accepted for Auth0, versus 10/10 ICL and 0/10 for zeroshot, SFT, and PPO; all 49 reported utility tests pass.The patch reconstructs cookies by distinguishing name/value boundaries from attribute continuations, then emits each reconstructed cookie separately without the experience context.Source: §C.6.4 · pp. 41–42

C.8.1 TASK AND REPAIR OBJECTIVE

• The reported tasks span GraphQL Mesh root wrapping, GitHub CLI squash-merge messages, Gauge circular-reference recovery, and Strapi relative media URLs.Their failures involve schema-root naming, command/API message flow, parser-state cleanup and discovery idempotence, and backend-relative URL handling, respectively.Source: §C.7.4 · p. 44, §C.8.1 · p. 45, §C.9.1 · p. 47

C.7.4 EPD EVALUATION WITHOUT EXPERIENCE CONTEXT

• Nine of ten EPD candidates are accepted for the GraphQL Mesh repair without experience context, matching ICL and exceeding zeroshot, SFT, and PPO, which each accept zero of ten.The patch passes 146 tests with one skipped and preserves separate outer-root resolution, stable generated names, and public configuration overrides for query, mutation, and subscription roots.Source: §C.7.4 · p. 44

C.8.2 ACCUMULATED EXPERIENCE AND EXPERIENCE-CONDITIONED ICL

• Repeated trials expose task-specific contracts that issue-and-repository-only trajectories fail to compose, while ICL candidates are accepted at ten of ten on the GitHub CLI and Gauge tasks.The GitHub CLI trials establish the default, TTY-gated interaction, last-commit retrieval, and mutation propagation; Gauge trials establish complete dictionary cleanup and shared file deduplication.Source: §C.8.2 · p. 45, §C.9.2 · p. 48

C.8.3 TEACHER-GENERATED DISTILLATION TARGETS

• Teacher-generated targets preserve accepted trial components while correcting scope or policy errors, including shared discovery deduplication, strict user-input validation, and schema-time root resolution.For GitHub CLI, targets specify the deterministic default, interactive TTY gate, and query-to-mutation path; for Gauge, they retain two-map cleanup while moving deduplication to shared utilities.Source: §C.7.3 · p. 43, §C.8.3 · pp. 45–46, §C.9.3 · pp. 48–49

C.8.4 EPD EVALUATION WITHOUT EXPERIENCE CONTEXT

• Nine of ten EPD candidates are accepted for the GitHub CLI repair, versus ten of ten ICL and zero of ten zeroshot, SFT, and PPO candidates.Without the five-trial context, the patch combines a pull-request title/body headless default, an interactive TTY-gated three-way choice, and propagation of the selected message through the GraphQL mutation.Source: §C.8.2 · p. 45, §C.8.3 · p. 46, §C.8.4 · p. 47

C.10.2 ACCUMULATED EXPERIENCE AND EXPERIENCE-CONDITIONED ICL

• The Strapi trials identify backend-prefix normalization and defensive provider handling as the accepted stored-path repair, while the teacher separately restores absolute-URL validation for user input.All ten ICL candidates are accepted, but the selected trial and ICL repair also relax the add-from-URL schema; the teacher retains stored-path and provider fixes while restoring the original policy.Source: §C.10.2 · p. 50, §C.10.3 · p. 50

C.10.3 TEACHER-GENERATED DISTILLATION TARGETS

• The three teacher-generated targets normalize stored paths with the backend-prefix helper, preserve absolute-only validation for user-entered URLs, and guard provider-side parsing.The teacher restores urlYupSchema.js to its original state while retaining defensive provider behavior from accumulated trials.Source: §C.10.3 · p. 51

C.10.4 EPD EVALUATION WITHOUT EXPERIENCE CONTEXT

• 9/10 EPD candidates are accepted without the four-trial experience context, versus 10/10 ICL candidates and 0/10 candidates for zeroshot, SFT, and PPO.The evaluated patch prefixes stored media paths, reuses the helper for filename extraction, guards invalid provider URLs, and leaves urlYupSchema.js unchanged.Source: §C.10.4 · p. 51

C.11 CROSS-CASE FINDINGS AND LIMITATIONS

• Across the eight cases, issue-and-repository-only samples produced no accepted repairs, whereas selected EPD patches reproduced task-specific behavior without the experience context.Selected sequences sometimes revisited rejected hypotheses or refined accepted patches, and did not always regenerate every accepted-tree edit.Source: §C.11 · p. 52, Table 24 · p. 52
• 79/80 accepted candidates let experience-conditioned ICL solve all eight cases, while EPD solved all eight with 72/79 valid samples without experience context.Zeroshot, SFT, and PPO produced no accepted candidates in 80 samples per method.Source: Table 24 · p. 52, §C.11 · p. 53
• Each teacher excerpt and EPD patch is a selected qualitative sample, so it documents observed repair patterns without identifying causality or estimating strategy frequency.Aggregate pass@1 in Table 1 remains the evidence for overall performance because the cases were selected for outcome contrast.Source: §C.11 · p. 53