Debate Benchmark update: 8 new models added
retained text
[https://github.com/lechmazur/debate](https://github.com/lechmazur/debate) \- GPT-6.2 Astra scores 1673, up 67 over Astra 6. \- Claude Fable 5.2 scores 1688, 18 points below Fable 5.1. \- Gemini 4 Carbon scores 1676, 27 points ahead of Argon. \- Qwen 4 Max makes the biggest gain: 1535 → 1679 (+144) over Qwen 3.8 Max. \- Tencent Hy4 (high) gains 85 points over Hy4 Preview, reaching 1646. \- Step 5 Preview (high) debuts at 1521. \- MiniMax M3.1 Flash (high) reaches 1509 (+27 over M3). \- Seed 2.1 Pro reaches 1468 (+34 over Seed 2.0 Pro). The benchmark now includes 10,928 rated debates across 683 topics, backed by 33,342 judge decisions. Debate Benchmark tests whether a model can defend a position through sustained opposition. A strong opening helps, but a model also has to answer objections, ask pointed questions, and carry its argument through the closing. Every matchup runs twice on the same motion, with PRO and CON swapped to control for side advantage. Three judges from distinct model families independently evaluate each debate’s winner and margin. The topic bank spans ten subject areas. Law and regulation form the largest group (135 motions), followed by labor, education, and social policy (122). Of the 683 motions, 419 are global in scope and 48 focus specifically on the U.S.
document fetched 2026-11-19 00:30 · canonical https://reddit.com/r/singularity/comments/1xzshsa/debate_benchmark_update_8_new_models_added