Source-linked AI summary

Efficient SWE Agent Benchmarking via Trajectory-Aware Evaluation

Kefeng Duan, Dewu Zheng, Yanlin Wang, Xiwen Wang, Ensheng Shi, Xilin Liu, Yuchi Ma, Jiachi Chen, Mingwei Liu, Zibin Zheng

arXiv:2609.01603v1cs.SEcs.AIcs.CL

TL;DR

Full SWE-agent benchmarks are costly, while result-only efficient evaluation discards process signals from how agents solve tasks. PTA-IRT uses historical trajectories as privileged information for calibration selection and ability estimation, and consistently outperforms prior IRT baselines on score and ranking recovery across four SWE benchmarks.

  • Problem

    Efficient SWE evaluation methods largely discard trajectory information by relying on pass/fail outcomes, despite costly multi-step benchmark tasks.

  • Method

    PTA-IRT uses historical execution trajectories as privileged information for trajectory-aware item selection and full-benchmark performance estimation.

  • Results

    PTA-IRT consistently outperforms prior IRT baselines on score and ranking recovery across four SWE benchmarks under low calibration budgets.

  • Takeaways & Limitations

    Privileged trajectory information makes budgeted SWE-agent evaluation more informative within the paper’s evaluation protocol.

  • Takeaways & Limitations

    Historical trajectories are available for training interactions but are not available or required when estimating a new agent’s performance at inference time.

Abstract

from arXiv · show

Evaluating software engineering agents on realistic benchmarks is costly, since each task may require multi-step code exploration, modification, and test execution. Existing efficient evaluation methods select representative subsets to estimate full-benchmark performance, but are largely result-only: they fit historical pass/fail response matrices or static task semantics, discarding how agents solve problems. We propose PTA-IRT, a Privileged Trajectory-Aware Item Response Theory framework that fuses process and outcome signals. Historical execution trajectories supply process-level evidence beyond pass/fail, such as explored context, attempted edits, and solving paths, which PTA-IRT uses as privileged information for calibration subset selection and ability estimation. Under low calibration budgets, PTA-IRT consistently outperforms prior IRT baselines on score and ranking recovery across four SWE benchmarks. Code and data are publicly available at https://github.com/DeepSoftwareAnalytics/PTA-IRT.

1 Introduction

Realistic SWE-agent benchmarks are costly because tasks require long, multi-step interaction, while prior efficient evaluation methods largely discard trajectory information. PTA-IRT addresses this gap by using historical trajectories as privileged information and consistently improves budgeted score and ranking recovery across four benchmarks.

  • Motivation: Full-benchmark SWE-agent evaluation is expensive because repository-level tasks require code exploration, tool use, editing, and test execution.SWE-bench contains more than 2,000 programming tasks, and a full SWE-agent run can exceed $8,000 under a $4-per-task limit.
  • Motivation: Prior IRT methods estimate benchmark performance from small subsets but typically represent each task only through final pass/fail outcomes.This discards process-level signals about explored context, actions, edits, and verification paths.
  • Approach: PTA-IRT incorporates historical agent trajectories as privileged information for trajectory-aware item selection and performance estimation.It summarizes trajectories, learns trajectory-aware 4PL item characteristics, selects a difficulty-stratified calibration subset, and transfers measurement knowledge through LUPI.
  • Results: PTA-IRT consistently outperforms prior IRT baselines on score and ranking recovery under low calibration budgets across four SWE benchmarks.The evaluation compares PTA-IRT with prior IRT baselines under a budgeted subset-to-full benchmark protocol.
  • Evaluation setting: The paper reports experiments on four SWE benchmarks and frames efficient benchmarking as recovering full-benchmark performance from a limited calibration subset.Historical trajectories are used offline, while a new agent is executed only on the calibration subset.

2 Background

SWE benchmarks increasingly require costly long-horizon agent interactions, motivating smaller evaluation subsets and IRT-based performance estimation. The background introduces IRT and LUPI, while positioning trajectories as privileged process information unavailable when evaluating a new agent.

  • Benchmark context: Modern SWE-agent benchmarks require long-horizon reasoning, tool use, editing, and testing, making them more expensive than conventional single-turn evaluations.These requirements arise in repository-level and agentic benchmark settings.
  • Efficient evaluation: IRT reduces evaluation cost by modeling agent–instance responses to estimate item properties and latent agent abilities from fewer test instances.Its item-level properties include difficulty and discriminability.
  • Prior work: Prior efficient SWE evaluation includes neural IRT extensions and agent-driven item selection, while SWE benchmarks also use rule-based or human-filtered subsets.Examples include Deep-IRT, PSN-IRT, AutoJudger, SWE-bench Lite, and SWE-bench Verified.
  • Item Response Theory: The IRT Item Characteristic Curve models the conditional probability that an agent correctly solves an instance from the agent’s latent ability and item difficulty.The one-parameter logistic formulation uses θ_i for agent ability and b_j for instance difficulty.
  • Learning Using Privileged Information: LUPI uses privileged information during training to improve a predictor that relies only on standard inputs at inference time.In this setting, historical agent trajectories are privileged information and are not required when estimating a new agent’s full-benchmark performance.

3 PTA-IRT

PTA-IRT uses historical agent trajectories as privileged process evidence to select an informative, difficulty-stratified calibration subset and estimate full-benchmark ability from limited outcomes.

  • Trajectory representations: PTA-IRT compresses heterogeneous execution logs into structured summaries covering task goals, explored context, executed edits, and solution paths.The summaries describe observable actions without inferring correctness or unobserved reasoning.
  • Calibration selection: The calibration subset allocates items across pass-rate difficulty bins and selects high-information tasks within each bin to cover easy-to-hard regimes.This balances informativeness with representativeness for recovering full-benchmark scores from a small budget.
  • Trajectory-aware scoring: A trajectory-aware 4PL IRT model injects summary-derived residuals into item discriminability and difficulty while keeping guessing and feasibility bounds shared.The shared item parameters are (aj, bj, cj, dj); trajectory information adjusts only discriminability and difficulty.
  • Calibration selection: PTA-IRT scores tasks with trajectory-aware Fisher information weighted by effective sample size, reducing priority for items with low-quality privileged evidence.Incomplete or malformed summaries are down-weighted, and the ESS factor lowers scores when privileged evidence is sparse or unreliable.
  • Ability estimation: A teacher–student LUPI estimator uses privileged summaries offline, while the deployable student estimates ability from calibration outcomes and extrapolates across all tasks.At test time, network weights are frozen and a scalar ability θ* is fit on the calibration subset before predicting benchmark-wide pass probabilities.

4 Experimental Setup

The experiments evaluate PTA-IRT across four SWE-bench versions against classical, neural, and agent-driven IRT baselines using score-error and ranking-recovery metrics with four-fold cross-validation.

  • Benchmarks: The evaluation covers SWE-bench Lite, Verified, Full, and Pro, spanning curated subsets and longer-horizon enterprise-style tasks.Lite and Verified are curated subsets of Full, while Pro extends the benchmark across broader repositories and languages.
  • Baselines: Baselines span classical IRT methods, neural IRT methods, and AutoJudger’s agent-driven calibration-item selection.Classical baselines include MLE, MCMC, VI, and VIBO; advanced methods include Deep-IRT and PSN-IRT.
  • Metrics: The study measures Mean Absolute Error, Kendall’s τ, and Spearman’s ρ to assess score accuracy and ranking recovery.MAE compares predicted with ground-truth full-benchmark performance, while Kendall’s τ and Spearman’s ρ evaluate ranking agreement and monotonic correlation.
  • Evaluation protocol: Four-fold cross-validation trains on 75% of evaluated models and tests on the remaining 25%.Process summaries are generated with DeepSeek-V4-Flash1 and embedded with all-MiniLM-L6-v2.

5 Main Results

Across four SWE benchmarks, PTA-IRT improves budgeted score and ranking recovery, with gains supported by trajectory-aware selection and process-aware representations. Ablations and trajectory analyses show that summary content, stratified selection, and privileged supervision each contribute to performance.

  • RQ1: Effectiveness: At 10% calibration, PTA-IRT achieves the best MAE, Kendall’s τ, and Spearman’s ρ across all four SWE benchmarks.Average performance is MAE 0.041±0.015, τ = 0.888, and ρ = 0.973.
  • RQ1: Effectiveness: 5% calibration already reaches τ = 0.768 on SWE-bench Lite, while 20% raises τ from 0.768 to 0.886 and tends to reduce MAE.Figure 2 evaluates calibration budgets from 5–25% using MAE, Kendall’s τ, and Spearman’s ρ.
  • RQ2: Ablation Analysis: Removing trajectory-aware scoring, LUPI, Top-K selection, or clustering worsens average score and ranking recovery relative to the full configuration.The only local exception is a marginal MAE improvement on Verified without LUPI; it does not overturn the average result.
  • RQ3: Trajectory Summaries: Summary dropout generally weakens recovery, and full dropout without ESS approaches the trajectory-scorer ablation on Lite, indicating that gains depend on summary content.At extreme dropout, both ESS settings degrade as usable process signal disappears.
  • RQ4: Summary Characteristics: Trajectory summaries divide labor across channels: Task Goal aligns with issue text, Edits Executed with submitted patches, and Context Explored and Path Overview remain process-specific.Whole-summary question similarity is 0.714 with Δ = +0.473 versus a random-question baseline; Task Goal reaches 0.776 and Edits Executed reaches 0.670 against submitted answers.
  • RQ4: Summary Characteristics: Submitted answers concentrate outcome geometry, whereas summaries preserve diverse solving paths that binary labels and patches alone do not collapse.Answer distances are PP 0.175 versus FF 0.290 with Δout = 0.033; full-summary distances are PP 0.143 versus FF 0.163 with Δout = 0.005.

6 Conclusion

PTA-IRT addresses the cost of full SWE-agent evaluation by using historical execution trajectories as privileged information in efficient IRT-based evaluation. Across four SWE benchmarks and low calibration budgets, it consistently outperforms prior IRT baselines on score and ranking recovery.

  • Conclusion: PTA-IRT builds structured trajectory summaries, selects a difficulty-stratified calibration subset, and estimates full-benchmark score and ranking from limited observations via LUPI.Historical execution trajectories provide privileged information for trajectory-aware measurement and ability estimation.
  • Conclusion: Under low calibration budgets, PTA-IRT consistently outperforms prior IRT baselines on score and ranking recovery across four SWE benchmarks.The evaluation protocol uses limited calibration observations to estimate full-benchmark performance.
Loading 2609.01603v1…