SWE-bench-Live-Verified: Automatic Filtering of SWE-bench-Live-Full

July 18, 2025 · View on GitHub

The process of automatically filtering the SWE-bench-Live-Full dataset to produce SWE-bench-Live-Verified is now live.


Methodology

We input the issue description, gold patch, and Fail2Pass test cases of each instance into a large language model (LLM) to assess whether the instance is suitable ("good") for evaluating agents' debugging capabilities.

Based on insights from several previous analyses:

We prompt the LLM to classify instances into the following categories:

CategoryDescription
1Issue description has minor vagueness or missing details, making it difficult (but possible) to understand, reproduce, and solve the bug.
2Issue description is highly vague, unclear, or incomplete, making reproduction and resolution impossible.
3Issue description includes proposed solutions, but at least one solution is misleading or incorrect relative to the gold patch.
4Issue description is sufficient, but the provided test cases are inadequate, overly broad, missing, or underspecified, such that passing tests does not reliably indicate issue resolution.
5Gold patch or test cases require specific outputs, errors, or formats not described in the issue description, or impose unnecessarily narrow or incorrect constraints, causing valid solutions to fail.
6Ground truth fix (patch or natural-language description) is explicitly provided in the issue, making the fix trivial.
7Issue description is complete, precise, and suitable for evaluation—neither overly simple nor overly complex (Good instance).
8Other issues (e.g., environmental constraints, flaky tests, licensing) render the instance unsolvable or unreliable for evaluation.

Results

We used GPT-o3 to categorize 1699 SWE-bench-full instances previously sampled by SWE-bench-Verified:

Negative Instance Filtering (Including Category 6):

  • SWE-bench-Verified identified 1160 negative instances.
  • GPT-o3 identified 652 negative instances, with a filter ratio of 38.4%.
  • Overlap between methods: 469 instances.
MetricValue
Precision469 / 652 ≈ 72%
Recall469 / 1160 ≈ 40%

Negative Instance Filtering (Excluding Category 6):

Since our prompt explicitly excludes too easy cases (Category 6), but SWE-bench-Verified did not, we also analyzed the filtered instances exluding Category 6:

  • GPT-o3 identified 433 negative instances (excluding Category 6).
  • Overlap with SWE-bench-Verified: 400 instances.
MetricValue
Precision400 / 433 ≈ 92%
Recall400 / 1160 ≈ 35%

Model Comparison (GPT-4.1 vs. GPT-o3)

We additionally evaluated GPT-4.1:

Negative Instances with Category 6:

  • GPT-4.1 Precision: 122 / 161 ≈ 76%
  • GPT-4.1 Recall: 122 / 1160 ≈ 10.5%

Negative Instances without Category 6:

  • GPT-4.1 Precision: 115 / 133 ≈ 86.5%
  • GPT-4.1 Recall: 115 / 1160 ≈ 10%

Indications

  • Reasoning models (GPT-o3) demonstrate superior precision and recall compared to general-purpose models (GPT-4.1).
  • LLMs generally show very high precision identifying problematic instances but adopt a conservative, cautious approach, resulting in lower recall (fewer filtered instances).

SWE-bench-Live-Verified and Agents' Performance on it

The initial Verified subset includes 500 instances filtered from the SWE-bench-Live-full set from July 2024 to April 2025. The filter ratio is around 38%. See filtering details at Live-Verified-log.

Due to time constraints and computational limitations, we simply retain instances from the Live-Lite subset (version 2025-04-20) that overlap with the Live-Verified subset to recalculate the success rates of various agents based on these remaining instances. After filtering, 174 instances remain.

Success Rates on Filtered Live-Verified (174 instances)

Agent TypeGPT-4oGPT-4.1Claude-3.7-SonnetDeepSeek-V3
SWE-agent14.94%16.09%19.54%13.22%
Agentless13.22%11.49%12.07%13.79%
Openhands+CodeAct6.32%12.07%20.69%13.22%

Observations

  • The results do not demonstrate a significant increase of resolved rate compared to the previously evaluated SWE-bench-Verified dataset, suggesting current LLMs and debugging agents are likely overfitting to the existing SWE-bench-Verified benchmark.
  • Additionally, because trivial instances have been explicitly excluded from our new Verified subset, which SWE-bench-Verified did not, it is reasonable that overall success rate increase remains modest.

SWE-bench-Live-Verified: Current Limitations and Future Directions

Currently, retrieving Fail2Pass instance content relies on hardcoded parsers, missing some test cases.

To address this, future developments will include an LLM-based agent capable of automatically extracting Fail2Pass test case contents, enabling seamless scalability across programming languages.