Benchmarks · coding
MLS-Bench-Lite
AI agents get a fixed ML research scaffold with strong baselines and must invent one algorithmic change — a loss, optimizer, sampler or training procedure — whose gain holds across settings, seeds, datasets and scales; Lite is the 30-task subset over all 12 research domains. Maintained by MLS-Bench authors. Score: The board's MLS-Bench-Lite score: the paper's normalized task metric, averaged (arithmetic mean) over the 30 Lite tasks, each agent run under Harbor with a 5-hour exploration budget.
14 models · 14 results · data as of 2026-10-29 · mls-bench.com ↗
Leaderboard
Fable 5.1 max
All results
14 rows| # | Model | Effort | MLS-Bench-Lite | Harness |
|---|---|---|---|---|
| 1 | Fable 5.1 | max | 50.3% | Claude Code (max effort) |
| 2 | Qwen 3.8 Max | max | 50.1% | Claude Code |
| 3 | Fable 5 | max | 49.9% | Claude Code (max effort, with fallback) |
| 4 | Opus 5 | max | 49.8% | Claude Code (max effort) |
| 5 | Kimi K3 | max | 48.3% | Kimi-Code (max) |
| 6 | GPT-5.6 Sol | max | 46.2% | Codex (max) |
| 7 | Opus 4.8 | max | 42.8% | Claude Code (max effort) |
| 8 | GLM 5.2 | max | 40.4% | Claude Code (max effort) |
| 9 | GPT-5.5 | xhigh | 35.5% | Codex (xhigh) |
| 10 | Kimi K2.7 Code | max | 35.1% | Kimi-Code |
| 11 | Qwen 3.7 Max | max | 31.7% | Claude Code |
| 12 | Sonnet 5 | max | 31.4% | Claude Code (max effort) |
| 13 | Kimi K2.6 | max | 26.7% | Kimi-Code |
| 14 | DeepSeek V4 Pro | max | 24.4% | Claude Code |