PaperlayerView exact revision on arXiv

Source-linked AI summary

SWE-Trail: History-Guided Trajectory Collection for Coding Agent Training

Ziming Li, Lu Yu, Jiatan Huang, Qing Cui, Jun Zhou, Chuxu Zhang

arXiv:2611.02074v1cs.AIcs.SE

TL;DR

Independent rollouts increasingly revisit repair targets, and advice drawn from history may make training trajectories depend on information unavailable at test time. SWE-Trail uses history to direct collection while keeping it outside the coding agent’s context. It improves student resolution over independent sampling on two benchmarks, with gains also reported for a second student model.

  • Problem

    Later independent trajectories become increasingly redundant, while putting history in the agent’s context can make training reasoning depend on advice unavailable at test time.

    Source: §1 · pp. 1–2

  • Method

    SWE-Trail uses earlier attempts to identify states for added collection effort, then compares briefly executed candidate actions proposed from the agent’s ordinary context.

    Source: §1 · p. 2, §3.1 · p. 5, §3.3 · p. 5

  • Results

    2.33 and 1.60 percentage points: H5 improves the 124B student’s mean resolution over SeedBank on SWE-bench Multilingual and SWE-bench Pro, respectively.

    Source: §1 · p. 2

  • Takeaways & Limitations

    The results support using history as a collector-side signal to improve the training value of new trajectories, rather than showing it to the coding agent as advice.

    Source: §6 · p. 10

    Experiments use two student models but only one collection model and one agent scaffold; history is fixed within each collection round and effort is uniform across issues.

    Source: §6 · p. 10

Abstract

from arXiv · show

Coding agents solve repository issues through long sequences of exploration, editing, and validation. These trajectories are increasingly used as post-training data, so how they are collected shapes what the trained model learns. A common approach is to sample multiple trajectories independently for each issue and then select or curate them. However, as more trajectories are sampled independently, later trajectories become increasingly redundant, even though earlier trajectories could guide where subsequent collection effort should be spent. A natural way to use this history is to expose it to the coding agent as advice. But training on such trajectories can make the model rely on advice that is unavailable at test time, creating a train-test mismatch. Instead, we argue that history should guide where collection effort is spent, while the coding agent decides what to generate. We introduce SWE-Trail, a trajectory collector that uses earlier trajectories to guide subsequent collection while keeping that history outside the coding agent's context. During a new rollout, SWE-Trail uses history to identify a state where additional collection effort may be useful. There, the coding agent proposes candidate next actions from its ordinary context, and SWE-Trail briefly executes them from the same checkpoint to decide which route continues. We evaluate collection approaches by training student models on the trajectories they produce. For a 124B student initialized from Ling-3.0-flash-base, SWE-Trail improves mean resolution over independent sampling by 2.33 and 1.60 percentage points on SWE-bench Multilingual and SWE-bench Pro, respectively. Without history, the same pipeline produces more resolved trajectories than independent sampling, but improves by only 0.11 points on Multilingual. These results show the value of using history to guide collection without exposing it to the coding agent.

1 Introduction

Independent trajectory collection becomes increasingly redundant, while using history as advice can make training trajectories depend on information unavailable at test time. SWE-Trail uses history to guide collection without showing it to the coding agent, improving student resolution over independent sampling.

1 Introduction

  • Later independent samples increasingly revisit repair targets and behaviors already explored, while advice-conditioned trajectories may rely on information unavailable at test time.The resulting question is how history can guide collection effort while the agent generates actions from its ordinary context.Source: §1 · pp. 1–2
  • SWE-Trail uses history to identify states for added collection effort, then compares candidate actions proposed from the agent’s ordinary context and continues from the route better supported by repository evidence.History stays outside the agent’s context throughout collection.Source: §1 · p. 2
  • 2.33 and 1.60 percentage points: H5 improves mean resolution over SeedBank on SWE-bench Multilingual and SWE-bench Pro, respectively, for a 124B Ling-3.0-flash-base student.Without history, H0 produces more resolved trajectories than SeedBank but improves Multilingual mean resolution by only 0.11 points.Source: §1 · p. 2

2 Motivation

Independent attempts show diminishing coverage gains and substantial overlap in repair targets. Earlier attempts can also provide signals for where additional collection effort may be useful.

2 Motivation

  • Independent attempts are analyzed for both diminishing coverage as attempts accumulate and evidence that earlier attempts can guide later collection.An attempt is one independent try at solving an issue.Source: §2 · p. 2

2.1 Diminishing Coverage Gains from Independent Rollouts

  • 58.4 issues are solved in expectation with one attempt; marginal additions fall from 13.1 issues for attempt two to 8.0, 5.1, and 3.4 for attempts three through five.The expected number is averaged over all k-attempt subsets of the five recorded attempts.Source: §2.1 · p. 2
  • One attempt covers 90.0% of patch-target files on stable-solved issues, 73.2% on mixed issues, and 80.1% on never-solved issues, with later attempts covering fewer new files.Even failed attempts often modify the same files, so independent sampling does not automatically redirect effort after repeated failures.Source: §2.1 · p. 3

2.2 History Provides Allocation Signals

  • Mixed histories can reveal decision points where successful and failed attempts take different next actions; failed-only histories also show repeated target focus, with mean pairwise Jaccard overlap of 0.841.Among 198 never-solved issues with defined overlap, 93 have a score of 1.0; overlap identifies where effort recurs, not why attempts fail.Source: §2.2 · p. 3

3 SWE-Trail

SWE-Trail compiles earlier attempts into signals for where added collection effort may help, while keeping history and collector information outside the coding agent’s context. At a detected state, it compares briefly executed candidate routes and continues the selected route.

3 SWE-Trail

  • Earlier attempts and read-only repository checks guide where additional collection effort should be spent, while candidate actions come from the coding agent’s ordinary context.The coding agent does not receive the issue history, and reference patches are unavailable during collection.Source: §3 · p. 4

3.1 Compiling History into Detectors

  • The compiler extracts allocation signals from history, resolves open questions with bounded read-only checks, and uses the evidence to define detector rules, comparison objectives, and sampling temperatures.It focuses on uncertainty in stable-solved histories, differences between successful and failed routes in mixed histories, and recurring failure patterns in never-solved histories.Source: §3.1 · p. 5

3.2 Sampling Comparable Alternatives

  • At a detected state, SWE-Trail samples candidate next actions from ordinary context without adding the history or comparison objective to the agent’s context.The first sample uses the agent’s ordinary temperature; remaining samples use the history-guided schedule.Source: §3.2 · p. 5

3.3 Selecting a Route by Executed Evidence

  • SWE-Trail executes the default and one alternative from the same checkpoint, compares their progress under a local objective, and continues the better-supported route or keeps the default when evidence is unclear.Only the selected route is retained, preserving the agent’s original reasoning and actions while removing private collector events.Source: §3.3 · p. 5

4 Experiments

The experiments compare training on trajectories from fixed collection conditions, using student resolution on two benchmarks as the outcome. H5 improves over independent sampling, while comparisons with H0 and a second student examine history’s contribution and generality.

4.1 Setting

  • The study holds the 300 collection issues and training data and evaluation protocols fixed, with five trajectories per issue; evaluation covers 300 Multilingual tasks and 731 SWE-Pro tasks.All trajectories are used for training, resolved or not, and the reported resolution is the mean across three training runs.Source: §4.1 · pp. 6–7
  • The study compares SeedBank with controlled adaptations of prior collection methods, an agent-visible experience-message method, and H0 and H5 variants that test history and collection decisions.H10 tests additional history; random detection and random retention test where comparisons occur and which route continues.Source: §4.1 · p. 6

4.2 Downstream Training Utility

  • H0 improves mean resolution over SeedBank by 0.11 points on Multilingual and 0.73 on SWE-Pro; H5 adds a further 2.22 and 0.87 points over H0, respectively.Task-paired bootstrap 95% intervals for H5 versus H0 span zero on both benchmarks.Source: §4.2 · p. 7
  • With Qwen3-14B-Base, H5 improves mean resolution over SeedBank by 3.11 points on Multilingual and 2.42 points on SWE-Pro.H5 also exceeds H0 by 1.67 and 1.60 points, respectively.Source: §4.2 · p. 7

Appendix E.4 gives the full results.

History brings modest additional gains beyond the no-history pipeline, while selection from independent rollouts does not match H5 and can complement it. The results also connect downstream gains to changes in collection effort and repair trajectories, not simply to more resolved collection attempts.

4.3 History Scale and Cost

  • 2.22 points on Multilingual and 0.87 points on SWE-Pro: H5 improves mean resolution over H0, while increasing history from five to ten attempts adds only 0.22 and 0.14 points, respectively.H10 uses 66.0% more tokens beyond the ordinary rollout than H5; H5 remains close to H0 in cost.Source: §4.3 · p. 8
  • Top-5-of-10 remains below H5 by 1.56 points on Multilingual and 0.96 points on SWE-Pro, so selecting from more independent rollouts does not recover H5’s gains.Top-5-of-10 samples ten independent trajectories per issue and selects five for training.Source: §4.3 · p. 8
  • Top-5-hybrid reaches 66.22 on Multilingual and 43.14 on SWE-Pro, exceeding H5 by 0.56 and 0.32 points, respectively.It selects five trajectories from five SeedBank and five H5 trajectories per issue, showing post-hoc selection can complement SWE-Trail.Source: §4.3 · p. 8

4.4 What History Adds to Collection

  • H0 resolves 37 more trajectories than SeedBank but improves Multilingual by only 0.11 points; H5 adds 16 resolved trajectories and six solved issues over H0 while improving the student by a further 2.22 points.H5’s downstream gains are not proportional to the increase in resolved trajectories; mixed histories account for 13 of its 16 additional resolved trajectories and five of six additional solved issues.Source: §4.4 · pp. 8–9
  • Compared with H0, H5 shortens repair processes from 40.89 to 39.51 mean actions and reduces repeated file reads from 7.59 to 7.01, while post-final-edit checking changes only slightly.Similar reductions in actions and repeated reads appear in H5-trained students’ evaluation trajectories.Source: §4.4 · p. 8

4.5 How the Collection Mechanism Changes

  • H5 samples candidates in 1,147 of 1,500 rollouts versus 664 under H0; candidate diversity and alternative-selection rates remain similar, while local progress on one route is more frequent under H5.Local progress occurs in 40.7% of H5 rollouts versus 28.7% of H0 rollouts; alternative-selection rates are 37.4% and 39.0%, respectively.Source: §4.5 · p. 9, Figure 3 · p. 9

5 Related Work

  • Unlike test-time search, trajectory curation, and experience-guided agents that use prior experience to guide actions, SWE-Trail uses history to guide collection effort while the coding agent acts without seeing it.It briefly compares candidate actions from the same checkpoint and adds execution evidence rather than only selecting or modifying completed trajectories.Source: §5 · pp. 9–10

6 Discussion and Limitations

  • The experiments support history as a collector-side signal rather than advice to the coding agent, but use only one collection model and one agent scaffold.The authors identify testing other collection models and scaffolds as future work.Source: §6 · p. 10

A.1 Coverage and Overlap of Independent Attempts

Independent attempts add less new issue coverage as their number grows and often revisit the same repair targets. Audits of mixed histories identify decision points where different next actions may be worth comparing, with a second model finding a similar number of candidate points.

A.1 Coverage and Overlap of Independent Attempts

  • The study quantifies how issue and patch-target coverage grows as more independent attempts are included, with issue coverage reported in Figure 1A and patch-target coverage in Table 6’s panel B.Source: §A.1 · p. 15
  • Pairwise patch-target Jaccard overlap averages 0.841 on never-solved issues; 93 of 198 issues with defined overlap have a score of 1.0.The overlap shows repeated repair targets, but does not explain why attempts fail.Source: §A.1 · pp. 15–16
  • Mixed histories have the lowest issue-level mean behavioral similarity on every measure, followed by never-solved and stable-solved histories.The ordering persists when failed or timed-out attempts are excluded, although mixed and never-solved histories reverse under pooled-pair medians.Source: §A.1 · p. 16, Table 7 · p. 16

A.2 Decision Differences in Mixed Histories

  • The mixed-history audit compares a successful and failed attempt at similar observed situations, first locating a decision difference and then judging both actions without outcome or future feedback.The audit asks whether the states are comparable and both immediate actions are defensible from information available at that point.Source: §A.2 · pp. 16–17
  • Among 44 localized cases, the observed divergence types are hypothesis, implementation, and validation.The full label space also includes localization, recovery, finish, resource/infra, and other.Source: §A.2 · p. 17
  • 31 primary mixed-history decision points—21 strict and ten possible—show evidence that more than one next action may be worth comparing.Order-sensitive cases are excluded from this count.Source: §A.2 · p. 18

A.3 Independent Cross-Model Audit Replication

  • Auditor B independently identifies 26 comparison points across the 43 primary pairs, versus Auditor A’s 31, but aggregate counts do not establish case-level agreement.The auditors agree on found versus no meaningful divergence in 40/43 primary pairs.Source: §A.3 · p. 18
  • The auditors identify candidate locations for comparison, not which action to take; runtime alternatives instead execute from the same checkpoint.The audits have stronger agreement on locating meaningful differences than on whether both actions are reasonable.Source: §A.3 · p. 19

B Collector Algorithm, Schemas, and Prompts

SWE-Trail compiles prior attempts and repository evidence into private rules that identify where collection may be useful, while the coding agent proposes actions from its ordinary context. It briefly executes and compares candidate routes, then exports one continuous trajectory without collector-private information.

B.1 End-to-End Collection Algorithm

  • The collection algorithm analyzes issue history, checks the repository read-only, freezes acquisition rules, and monitors the agent’s ordinary rollout for a matching state.At a match, it samples candidate actions, privately shortlists an alternative, executes both routes from one checkpoint, and continues from the selected route.Source: §B.1 · p. 20

B.2 Collection Cost Accounting

  • Collection cost separates additional overhead from total bank cost: Cbank = Crollout + Coverhead.Coverhead is compute beyond an ordinary rollout; accounting also distinguishes summed trajectory-hours from concurrent wall-clock time.Source: §B.2 · pp. 20–21

B.3 Offline Compiler

  • The offline compiler extracts supported patterns from ordered attempts, runs bounded read-only checks for unresolved questions, and turns the evidence into candidate detector rules.Rules are parsed, linted, deduplicated, and replayed before collection; the private acquisition record never enters a coding-agent request.Source: §B.3 · p. 21

B.4 Detector Schema and Admission Checks

  • A detector rule specifies observable conditions, supporting evidence, a local route-comparison objective, confidence and admission status, and abstention conditions.Predicates use only features available during the rollout; H5 and H10 rules require history evidence and replay, while H0 rules require source-clean repository evidence.Source: §B.4 · p. 21
  • Of 687 H5 candidate detector rules, 569 match a source trajectory during replay and 397 pass admission checks for collection.Only admitted, fixed rules can guide collection.Source: §B.4 · pp. 21–22

B.5 Online Runtime Details

  • At a matching state, the runtime samples and deduplicates candidate continuations, executes the default and at most one alternative from isolated copies of the same checkpoint, then selects only with clear evidence of better progress.Otherwise, it retains the default route and exports one continuous trajectory.Source: §B.5 · p. 22

B.6 Information Boundaries and Export Checks

  • The exported trajectory preserves the agent’s original reasoning and actions while removing private collector events and verifying that retained actions and tool results form one continuous record.The export checks reject records containing private collector content or a patch that differs from the selected workspace.Source: §B.6 · pp. 22–23

B.7 Prompts and Inputs

  • The model-assisted collector stages use private inputs and outputs, while candidate generation uses the ordinary coding-agent prompt without history or SWE-Trail instructions.The study uses 300 issues and five trajectories per collection condition; H5 draws on five SeedBank attempts, and H10 on ten.Source: §B.7 · p. 23, Table 12 · p. 23, §C.1 · p. 24

C.2 Evaluation Benchmarks and Overlap Audit

The evaluation uses SWE-bench Multilingual and SWE-bench Pro, with no shared issues or repositories between either benchmark and the collection pool. The controlled conditions compare trajectory curation, history shown to the agent, and history kept private to guide collection.

C.2 Evaluation Benchmarks and Overlap Audit

  • SWE-bench Multilingual contains 300 issues across nine languages, while SWE-bench Pro’s public split contains 731 longer-horizon repository tasks.The collection and evaluation sets share neither issues nor repositories.Source: §C.2 · p. 25

C.3 Experimental Conditions

  • The conditions adapt three curation methods, expose compiled experience to the agent in SWE-Exp, or keep prior attempts private to guide SWE-Trail collection.H0 removes history; H5 and H10 use five and ten prior attempts, while randomized controls alter detection location or route retention.Source: §C.3 · p. 25, Table 13 · p. 26

D.1 Controlled SFT Protocol

  • The primary student is a 124B Ling-3.0-flash-base model trained on 1,500 trajectories per condition, with training settings fixed across conditions.Every trajectory is used whether or not it resolves the issue; each condition is trained independently three times.Source: §D.1 · p. 26, Table 14 · p. 27
  • A second comparison trains Qwen3-14B-Base on SeedBank, H0, H5, and H10 trajectory sets with the other training settings fixed.Each dataset is trained independently three times.Source: §D.1 · p. 26

D.2 Evaluation Protocol

The evaluation compares downstream resolution across trained students and examines task-paired contrasts, using fixed evaluation conditions and defined collection metrics. H5 exceeds H0 on both benchmarks in point estimates, while the task-bootstrap intervals include zero.

D.2 Evaluation Protocol

  • All students share the same evaluation scaffold, prompt, tools, budgets, repository construction, and verifier; only the trained student varies.Errors and timeouts remain in the denominator, and condition means use resolved counts before rounding.Source: §D.2 · p. 27

D.3 Metric Definitions

  • A resolved trajectory passes benchmark tests; a collection issue is solved if any of its five new trajectories resolves, while a verified positive also passes process and export checks.Repeated reads track repeated access to a normalized file target, and validation counts executable checks.Source: §D.3 · p. 27
  • Candidate sampling marks a detector match that triggers proposals, while an executed comparison requires both candidate suffixes to run from the same checkpoint.Candidate coverage measures whether a useful action appears among the eight proposals; collection-level bootstrap resamples issues because the five trajectories per issue are not independent.Source: §D.3 · pp. 27–28

E.1 Per-run resolution

  • Every H5 run exceeds every SeedBank run on both benchmarks, yielding a one-sided permutation-test p = 0.05 on each.With three runs per condition, 0.05 is the smallest attainable p-value for this test.Source: §E.1 · p. 28, Table 15 · p. 28

E.2 Task-paired contrasts

  • H5–H0 gains are 2.22 points on Multilingual and 0.87 on SWE-Pro, with both 95% task-bootstrap intervals including zero.Contrasts are conditional on the trained models and resample evaluation tasks.Source: §E.2 · p. 29, Table 16 · p. 29, Figure 4 · p. 29

E.3 Multilingual results by language

H5 improves resolution across repository languages and for a second student model, while collected and held-out trajectories show changes in repeated reads and actions. The held-out behavior differences are descriptive and do not identify the cause of resolution gains.

E.3 Multilingual results by language

  • H5 is no lower than H0 on seven of nine languages; the largest increase is on C, despite only three C issues in the collection set.The language-level breakdown is descriptive because each language has only 12–44 tasks.Source: §E.3 · pp. 29–30, Table 17 · p. 30

E.4 Results for the Qwen3-14B Student

  • For Qwen3-14B-Base, H5 exceeds SeedBank by 3.11 points on Multilingual and 2.42 on SWE-Pro, and every H5 run exceeds every SeedBank and H0 run.Increasing history from five to ten attempts adds 1.00 point on Multilingual and 0.36 on SWE-Pro.Source: §E.4 · p. 30, Table 18 · p. 30

F.1 Actions, Repeated Reads, and Validation

  • H5 lowers mean actions and repeated reads from SeedBank through H0 to H5, while post-final-edit validation increases; H10 partly reverses these changes.The process means cover all 1,500 collected trajectories per condition.Source: §F.1 · p. 31, Table 19 · p. 31

F.2 Student Behavior on Held-Out Issues

  • H5-trained students take fewer actions and repeat fewer file reads on both benchmarks, but post-final-edit checking moves in opposite directions across benchmarks.Checking decreases on Multilingual and increases on SWE-Pro; these behavior differences do not identify the cause of resolution gains.Source: §F.2 · p. 32, Table 20 · p. 32

F.3 Collection Yield by History Type

H5’s additional collection yield is concentrated in mixed histories, while runtime stages, compiler behavior, selector tests, and token costs qualify how the collection process works. The stricter selector produces lower student means despite nearly unchanged collection yield.

F.3 Collection Yield by History Type

  • Mixed histories account for 13 of H5’s 16 additional resolved trajectories and five of six additional solved issues; H5 preserves H0’s solved-issue coverage for this group.The overall H5–H0 collection differences are descriptive: +1.07 percentage points in resolved trajectories and +2.0 points in solved issues, with intervals including zero.Source: §F.3 · p. 32, Table 21 · p. 32

F.4 Runtime Comparison Stages

  • Candidate sampling occurs in 1,147 H5 rollouts versus 664 H0 rollouts, while conditional diversity, execution, and alternative-selection rates are similar.Because comparisons occur at different points in the rollouts, similar conditional rates do not imply identical candidate distributions.Source: §F.4 · p. 33, Table 22 · p. 33

F.5 Compiler Classes and Representative Records

  • Mixed and never-solved compiler classes have more active rules per issue than stable-solved histories, and the compiler abstains more often on stable-solved histories.An active rule passed compiler admission; it does not mean the rule matched a new rollout.Source: §F.5 · pp. 33–34, Table 23 · p. 33

G.1 Randomized Controls and Selector Replay

  • The stricter selector leaves collection yield nearly unchanged but lowers student means by 1.78 points on Multilingual and 0.91 on SWE-Pro.In 18 fixed-comparison replays, the stricter criterion increases ties from 5 to 12 and mainly converts uncertain preferences into abstentions; candidate coverage is a limitation in 13 of those 18.Source: §G.1 · pp. 34–35

G.2 History Scale and Collection Cost

  • H10 uses 2.160B total tokens, 12.9% above H5, while adding little mean improvement over H5 as collection cost rises.H5 uses 1.914B total collection tokens; H0’s 1.845B total is a lower bound because offline compiler calls were unrecorded.Source: Figure 6 · p. 35, §G.2 · pp. 35–36

G.3 Verifier-Based Training Corpus Selection

Verifier-based post-hoc selection from independent samples does not recover H5’s downstream gains, while selection that includes H5 trajectories can complement them. A repair case traces how earlier failed attempts inform route selection without entering the coding agent’s context.

G.3 Verifier-Based Training Corpus Selection

  • Top-5-of-10 remains below H5 by 1.56 points on Multilingual and 0.96 on SWE-Pro, while Top-5-hybrid exceeds H5 by 0.56 and 0.32 points, respectively.Both selection conditions rank 3,000 candidates and retain 1,500 trajectories; the hybrid set also has a 1.27-point higher trajectory-resolution rate under issue-cluster pairing, with a 95% interval including zero.Source: §G.3 · pp. 36–37, Table 25 · p. 36, Table 26 · p. 37

H.1 A Repair Case after Five Failed Attempts

  • In the schemars-304 case, four of five H5 trajectories resolve after all five prior SeedBank attempts failed to validate a one-line edit.The history leaves the edit’s sufficiency open; candidate 6 runs relevant Rust tests while candidate 0 continues reading, and the selector uses repository evidence to choose the route.Source: §H.1 · p. 37, Figure 7 · p. 38

H.2 Retaining and Removing Advice in Training Trajectories

The advice-export experiment finds that removing advice does not consistently improve downstream performance. The related-work discussion distinguishes SWE-Trail by using prior attempts to guide new trajectory collection rather than to condition the coding agent or select completed data.

H.2 Retaining and Removing Advice in Training Trajectories

  • The retained export keeps aligned advice before agent reasoning, while the removed export deletes advice and reasoning explicitly attributed to it; actions remain those generated with advice visible.Thus, removing advice changes the training record, not the rollout context.Source: §H.2 · p. 38
  • Advice removal lowers Multilingual resolution but raises SWE-Pro resolution relative to retaining advice, so it does not consistently improve performance.The comparison uses two exports of the same advice-conditioned trajectories.Source: §H.2 · pp. 38–39, Table 27 · p. 39

I Detailed Related Work

I.1 Executable Coding Agents and Trajectory Corpora

I.2 Inference-Time Search and Selection

  • Inference-time selection, prefix continuation, summaries, and tree search reuse computation to improve the current solution, whereas SWE-Trail collects training trajectories using prior independent attempts and at most one comparison per rollout.Its H0 control tests the bounded collection procedure without history, not a full tree-search policy.Source: §I.2 · pp. 39–40

I.3 Trajectory Construction and Post-Hoc Curation

  • Post-hoc curation filters or emphasizes completed data, while SWE-Trail uses prior attempts during new collection, where executing an alternative can produce repository evidence otherwise unavailable.The approaches can be combined; Appendix G.3 evaluates verifier-based selection over SeedBank and H5 trajectories.Source: §I.3 · p. 40

I.4 Training-Time Diversity and Adaptive Curation

SWE-Trail changes how new trajectories are collected: prior attempts guide where effort is spent, while the coding agent generates actions without seeing that history. Its fixed, issue-specific rules separate this approach from methods that change agent context, behavior, or workflow.

I.4 Training-Time Diversity and Adaptive Curation

  • SWE-Trail improves the training distribution by changing how new trajectories are collected, with prior attempts guiding effort while the coding agent proposes actions without that history.The candidate actions come from the agent’s ordinary context.Source: §I.4 · p. 40

I.5 Experience-Guided Agents

  • Unlike experience-guided agents that change visible context, workflow, or behavior, SWE-Trail keeps prior information outside the coding agent so exported trajectories do not depend on history being available at test time.The related work includes Reflexion, ExpeL, and methods that derive workflows, skills, or procedural knowledge from prior executions.Source: §I.5 · p. 40

I.6 Adaptive Acquisition and Harness Optimization

  • SWE-Trail allocates effort within one issue using history compiled before a rollout; its rules stay fixed and do not change the agent’s prompt, tools, or model weights.This is narrower than adaptive acquisition or harness optimization approaches that adapt prompts, workflows, tools, or control procedures.Source: §I.6 · p. 41
Loading 2611.02074v1…