Fangcun DeepSeek 1017 Report - ImpossibleBench Conflicting SWE-bench: leaderboard
Fangcun report 2610.35671v1, hosted API evaluations Oct 4–25 2026; Appendix A.1 exact model identities, maximum reasoning effort unless named otherwise, temperature 1/top_p 0.95. GLM 5.3 judge where required. API/format failures excluded by source; common cohort versus successfully completed denominators retained. Primary PDF printed values, not estimated plot heights. Figure 2.5 89 repository tasks; passing deliberately faulty tests contrary to task specification is reward hacking; lower is source-defined desirable. One-off wrong expected output vs conflicting expected outputs kept distinct. Metric: Native rate (%).
Metric: Native rate (%). Source: arxiv.org. Saturation forecast: Around January 2027. 7 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3.8 Flash | 0 | #24 |
| 2 | GPT-6 Astra (Max) | 0 | #3 (GPT-6 Astra) |
| 3 | DeepSeek V4.1 Flash (Max) | 2.2 | #36 (DeepSeek V4.1 Flash) |
| 4 | GPT-5.6 Sol (Max) | 3.4 | #19 (GPT-5.6 Sol) |
| 5 | DeepSeek V4.1 Flash (1017) (Max) | 10.1 | #25 (DeepSeek V4.1 Flash (1017)) |
| 6 | DeepSeek V4 Pro (Max) | 18 | #109 (DeepSeek V4 Pro) |
| 7 | DeepSeek V4 Flash (Max) | 19.1 | #147 (DeepSeek V4 Flash) |
No result here: #2 Claude Opus 5.5, #6 GPT-6.1 Sol, #8 Gemini 4 Argon.
Interactive version: theaggregate.ai/benchmark?slug=fangcun-deepseek-1017-report-impossiblebench-conflicting-swe-bench · How It Works · Data refreshed daily, snapshot 2026-10-28.