AutoBench: Automating LLM Evaluation through Reciprocal Peer Assessment
Paper • 2510.22593 • Published
Datasets:
Tasks:
Text Generation
Modalities:
Text
Formats:
parquet
Languages:
English
Size:
10K - 100K
ArXiv:
License:
prompt_id stringclasses 252
values | model stringclasses 52
values | answer stringclasses 217
values | score stringlengths 0 18 ⌀ | score_sd stringlengths 0 19 ⌀ | n_judges stringclasses 11
values | cost stringlengths 0 22 ⌀ | judges_json stringlengths 2 249 ⌀ |
|---|---|---|---|---|---|---|---|
q-corpus-1/q100_1/1 | Muse-glimmer-30b | Error answering - Failed to respond | 1.8 | 0.42163702135578396 | 10 | 0.00218066 | {"Deepseek-v4-flash-0731": 2, "GLM-5.2": 2, "Gemini-3.7-flash": 1, "Gpt-oss-120b": 2, "Grok-4.3": 2, "Hy3": 2, "Minimax-m3": 2, "Muse-glimmer-30b": 2, "Nemotron-3-ultra-550b-a55b": 1, "Nemotron-3.5-lightning": 2} |
q-corpus-1/q100_1/1 | Grok-4.3 | Error answering - Failed to respond | 3.822222222222222 | 0.856997342145496 | 9 | 0.00838935 | {"Deepseek-v4-flash-0731": 4, "GLM-5.2": 4.8, "Gemini-3.7-flash": 4.8, "Gpt-oss-120b": 4.2, "Grok-4.3": 4, "Hy3": 3.6, "Muse-glimmer-30b": 3.2, "Nemotron-3-ultra-550b-a55b": 3.8, "Nemotron-3.5-lightning": 2, "Qwen3.5-9b": null} |
q-corpus-1/q100_1/1 | Qwen3.5-9b | Error answering - Failed to respond | 1.34 | 0.5081557066709201 | 10 | 0.0002625 | {"Deepseek-v4-flash-0731": 1, "GLM-5.2": 2.2, "Gemini-3.7-flash": 1, "Gemma-4-31b-it": 1.2, "Gpt-oss-120b": 2, "Hy3": 1, "Muse-glimmer-30b": 1, "Nemotron-3-ultra-550b-a55b": 1, "Nemotron-3.5-lightning": 2, "Qwen3.5-9b": 1} |
q-corpus-1/q100_1/1 | Claude-sonnet-5 | Error answering - Failed to respond | 4.34 | 1.0458383134011575 | 10 | 0.219224 | {"Deepseek-v4-flash-0731": 5, "Gemini-3.7-flash": 5, "Gemma-4-31b-it": 4.8, "Gpt-5.6-luna": 2, "Grok-4.3": 5, "Hy3": 4.6, "Minimax-m3": 5, "Muse-glimmer-30b": 3, "Nemotron-3-ultra-550b-a55b": 5, "Nemotron-3.5-lightning": 4} |
q-corpus-1/q100_1/1 | GLM-5.3 | Error answering - Failed to respond | 3.6800000000000006 | 1.3070747662029303 | 10 | 0.009636 | {"Deepseek-v4-flash-0731": 3.8, "GLM-5.2": 2, "Gemini-3.7-flash": 4.8, "Gpt-oss-120b": 3, "Grok-4.3": 5, "Hy3": 4.6, "Muse-glimmer-30b": 2, "Nemotron-3-ultra-550b-a55b": 2, "Nemotron-3.5-lightning": 4.6, "Qwen3.5-9b": 5} |
q-corpus-1/q100_1/1 | GLM-5.2 | Error answering - Failed to respond | 3.64 | 1.2103259432437565 | 10 | 0.00486486 | {"Deepseek-v4-flash-0731": 4.2, "Gemini-3.7-flash": 5, "Gemma-4-31b-it": 4.8, "Gpt-5.6-luna": 3.4, "Grok-4.3": 4.6, "Hy3": 2, "Minimax-m3": 4.2, "Muse-glimmer-30b": 2, "Nemotron-3.5-lightning": 4.2, "Qwen3.5-9b": 2} |
q-corpus-1/q100_1/1 | Hy3 | Error answering - Failed to respond | 3.4399999999999995 | 1.1805836597961947 | 10 | 0.003068208 | {"Deepseek-v4-flash-0731": 2, "GLM-5.2": 2, "Gemini-3.7-flash": 5, "Gemma-4-31b-it": 4.6, "Gpt-5.6-luna": 3.2, "Gpt-oss-120b": 4, "Hy3": 4.8, "Muse-glimmer-30b": 3, "Nemotron-3.5-lightning": 3.8, "Qwen3.5-9b": 2} |
q-corpus-1/q100_1/1 | Gemini-3.7-flash | Error answering - Failed to respond | 4.22 | 0.9818350166906862 | 10 | 0.0182505 | {"GLM-5.2": 5, "Gemini-3.7-flash": 5, "Gemma-4-31b-it": 4.8, "Gpt-5.6-luna": 2, "Gpt-oss-120b": 4.4, "Grok-4.3": 3, "Minimax-m3": 4.6, "Nemotron-3-ultra-550b-a55b": 4.8, "Nemotron-3.5-lightning": 4, "Qwen3.5-9b": 4.6} |
q-corpus-1/q100_1/1 | Nemotron-3-ultra-550b-a55b | Error answering - Failed to respond | 2.82 | 0.8189424074174367 | 10 | 0.0148086 | {"Deepseek-v4-flash-0731": 2, "Gemini-3.7-flash": 2, "Gpt-5.6-luna": 3.2, "Grok-4.3": 3.6, "Hy3": 3, "Minimax-m3": 2, "Muse-glimmer-30b": 2, "Nemotron-3-ultra-550b-a55b": 4.4, "Nemotron-3.5-lightning": 3, "Qwen3.5-9b": 3} |
q-corpus-1/q100_1/1 | Qwen3.8-max | Error answering - Failed to respond | 4.16 | 1.1654279137819816 | 10 | 0.030646 | {"Deepseek-v4-flash-0731": 2, "Gemini-3.7-flash": 5, "Gpt-5.6-luna": 2, "Gpt-oss-120b": 4.6, "Grok-4.3": 4.4, "Hy3": 4.8, "Minimax-m3": 4.8, "Muse-glimmer-30b": 4.2, "Nemotron-3-ultra-550b-a55b": 4.8, "Nemotron-3.5-lightning": 5} |
q-corpus-1/q100_1/1 | Nemotron-3.5-lightning | Error answering - Failed to respond | 2.64 | 0.9131143289740763 | 10 | 0.00695648 | {"Deepseek-v4-flash-0731": 2, "GLM-5.2": 3.6, "Gemma-4-31b-it": 2.2, "Gpt-5.6-luna": 2.4, "Gpt-oss-120b": 4, "Grok-4.3": 2, "Minimax-m3": 2.0, "Muse-glimmer-30b": 2, "Nemotron-3-ultra-550b-a55b": 2, "Nemotron-3.5-lightning": 4.2} |
q-corpus-1/q100_1/1 | Muse-spark-1.2 | Error answering - Failed to respond | 3.3111111111111104 | 1.0867893591267404 | 9 | 0.01011925 | {"Gemini-3.7-flash": 4.4, "Gemma-4-31b-it": 3.4, "Gpt-5.6-luna": 3.4, "Gpt-oss-120b": 4.8, "Grok-4.3": 4.2, "Hy3": 2, "Muse-glimmer-30b": 2, "Nemotron-3-ultra-550b-a55b": 2, "Nemotron-3.5-lightning": 3.6, "Qwen3.5-9b": null} |
q-corpus-1/q100_1/1 | Grok-4.6 | Error answering - Failed to respond | 4.333333333333334 | 0.9380831519646858 | 9 | 0.07565 | {"Deepseek-v4-flash-0731": 5, "GLM-5.2": 5, "Gemini-3.7-flash": 5, "Gpt-5.6-luna": 2, "Gpt-oss-120b": 4.4, "Grok-4.3": 4.8, "Hy3": 4.4, "Minimax-m3": 4.2, "Muse-glimmer-30b": 4.2, "Qwen3.5-9b": null} |
q-corpus-1/q100_1/1 | Gpt-5.6-luna | Error answering - Failed to respond | 4.779999999999999 | 0.304776785350999 | 10 | 0.015338 | {"Deepseek-v4-flash-0731": 5, "Gemini-3.7-flash": 5, "Gemma-4-31b-it": 4.8, "Gpt-oss-120b": 4.6, "Grok-4.3": 5, "Hy3": 4.8, "Minimax-m3": 4.8, "Muse-glimmer-30b": 4, "Nemotron-3-ultra-550b-a55b": 5, "Qwen3.5-9b": 4.8} |
q-corpus-1/q100_1/1 | Gpt-5.6-sol | Error answering - Failed to respond | 4.8 | 0.32071349029490914 | 8 | 0.715363 | {"Deepseek-v4-flash-0731": 5, "GLM-5.2": 5, "Gemini-3.7-flash": 5, "Gemma-4-31b-it": 4.8, "Gpt-oss-120b": 4.4, "Grok-4.3": 5, "Minimax-m3": 5, "Muse-glimmer-30b": 4.2, "Nemotron-3.5-lightning": null, "Qwen3.5-9b": null} |
q-corpus-1/q100_1/1 | Kimi-k2.6 | Error answering - Failed to respond | 2.8799999999999994 | 0.9761602782785668 | 10 | 0.08495205812 | {"Deepseek-v4-flash-0731": 2, "GLM-5.2": 4.8, "Gemini-3.7-flash": 3, "Gemma-4-31b-it": 2.8, "Gpt-5.6-luna": 3, "Gpt-oss-120b": 4.2, "Hy3": 2, "Nemotron-3-ultra-550b-a55b": 2, "Nemotron-3.5-lightning": 2, "Qwen3.5-9b": 3} |
q-corpus-1/q100_1/1 | Gemma-4-31b-it | Error answering - Failed to respond | 3.15 | 0.892828570971414 | 8 | 0.00048958 | {"Deepseek-v4-flash-0731": 2, "GLM-5.2": 2, "Gemma-4-31b-it": 3.4, "Gpt-5.6-luna": 2.6, "Gpt-oss-120b": 4.2, "Grok-4.3": 4, "Hy3": 3, "Nemotron-3-ultra-550b-a55b": 4, "Nemotron-3.5-lightning": null, "Qwen3.5-9b": null} |
q-corpus-1/q100_1/1 | Deepseek-v4-flash-0731 | Error answering - Failed to respond | 3.2888888888888883 | 0.9803627446568495 | 9 | 0.00095858 | {"Deepseek-v4-flash-0731": 2, "GLM-5.2": 2, "Gemma-4-31b-it": 3.8, "Gpt-5.6-luna": 3.6, "Gpt-oss-120b": 4, "Grok-4.3": 2, "Hy3": 4.2, "Minimax-m3": 4.0, "Nemotron-3.5-lightning": 4, "Qwen3.5-9b": null} |
q-corpus-1/q100_1/1 | Gpt-oss-120b | Error answering - Failed to respond | 2.2888888888888888 | 1.1185307823708344 | 9 | 0.0005981 | {"Deepseek-v4-flash-0731": 2, "GLM-5.2": 2, "Gemma-4-31b-it": 3, "Gpt-5.6-luna": 1, "Gpt-oss-120b": 2, "Grok-4.3": 3, "Hy3": 2, "Minimax-m3": 4.6, "Nemotron-3-ultra-550b-a55b": 1, "Qwen3.5-9b": null} |
q-corpus-1/q100_1/1 | Minimax-m3 | Error answering - Failed to respond | 0 | 0.0 | {} | ||
q-corpus-1/q100_1/2 | Gemma-4-31b-it | 1. Cardiac output by the Fick method: 4.72 L/min
2. Cardiac output by thermodilution: 2.29 L/min
3. The Fick method computes pulmonary blood flow (forward flow) from oxygen consumption and the arteriovenous oxygen content difference. In tricuspid regurgitation, the mixed venous sample from the pulmonary artery repr... | 4.819999999999999 | 0.11352924243950945 | 10 | 0.00028792 | {"Deepseek-v4-flash-0731": 4.8, "GLM-5.2": 4.8, "Gemini-3.7-flash": 5, "Gemma-4-31b-it": 4.8, "Gpt-5.6-luna": 4.8, "Grok-4.3": 4.8, "Hy3": 4.8, "Minimax-m3": 4.8, "Nemotron-3-ultra-550b-a55b": 4.6, "Qwen3.5-9b": 5} |
q-corpus-1/q100_1/2 | GLM-5.3 | 1. Cardiac output by the Fick method: 4.72 L/min
2. Cardiac output by thermodilution: 2.29 L/min
3. The Fick method computes pulmonary blood flow (forward flow) from oxygen consumption and the arteriovenous oxygen content difference. In tricuspid regurgitation, the mixed venous sample from the pulmonary artery repr... | 4.88 | 0.13984117975602034 | 10 | 0.0050818 | {"Deepseek-v4-flash-0731": 4.8, "GLM-5.2": 4.8, "Gemini-3.7-flash": 5, "Gpt-5.6-luna": 5, "Gpt-oss-120b": 5, "Grok-4.3": 4.8, "Hy3": 4.6, "Nemotron-3-ultra-550b-a55b": 5, "Nemotron-3.5-lightning": 4.8, "Qwen3.5-9b": 5} |
q-corpus-1/q100_1/2 | Muse-glimmer-30b | 1. Cardiac output by the Fick method: 4.72 L/min
2. Cardiac output by thermodilution: 2.29 L/min
3. The Fick method computes pulmonary blood flow (forward flow) from oxygen consumption and the arteriovenous oxygen content difference. In tricuspid regurgitation, the mixed venous sample from the pulmonary artery repr... | 4.688888888888889 | 0.17638342073763935 | 9 | 0.00142172 | {"Deepseek-v4-flash-0731": 4.6, "Gemini-3.7-flash": 4.8, "Gemma-4-31b-it": 4.8, "Gpt-oss-120b": 4.8, "Hy3": 4.6, "Minimax-m3": 4.6, "Muse-glimmer-30b": 4.4, "Nemotron-3-ultra-550b-a55b": 4.6, "Nemotron-3.5-lightning": null, "Qwen3.5-9b": 5} |
q-corpus-1/q100_1/2 | Gpt-5.6-sol | 1. Cardiac output by the Fick method: 4.72 L/min
2. Cardiac output by thermodilution: 2.29 L/min
3. The Fick method computes pulmonary blood flow (forward flow) from oxygen consumption and the arteriovenous oxygen content difference. In tricuspid regurgitation, the mixed venous sample from the pulmonary artery repr... | 4.859999999999999 | 0.13498971154211067 | 10 | 0.0559955 | {"GLM-5.2": 4.8, "Gemini-3.7-flash": 5, "Gemma-4-31b-it": 5, "Gpt-5.6-luna": 4.8, "Gpt-oss-120b": 4.8, "Grok-4.3": 4.8, "Minimax-m3": 5, "Muse-glimmer-30b": 4.6, "Nemotron-3-ultra-550b-a55b": 4.8, "Qwen3.5-9b": 5} |
q-corpus-1/q100_1/2 | Claude-sonnet-5 | 1. Cardiac output by the Fick method: 4.72 L/min
2. Cardiac output by thermodilution: 2.29 L/min
3. The Fick method computes pulmonary blood flow (forward flow) from oxygen consumption and the arteriovenous oxygen content difference. In tricuspid regurgitation, the mixed venous sample from the pulmonary artery repr... | 4.8 | 0.23094010767585016 | 10 | 0.03395 | {"Deepseek-v4-flash-0731": 4.8, "Gemini-3.7-flash": 5, "Gemma-4-31b-it": 4.8, "Gpt-oss-120b": 4.8, "Grok-4.3": 5, "Hy3": 5, "Minimax-m3": 4.8, "Muse-glimmer-30b": 4.4, "Nemotron-3.5-lightning": 4.4, "Qwen3.5-9b": 5} |
q-corpus-1/q100_1/2 | Gpt-5.6-luna | 1. Cardiac output by the Fick method: 4.72 L/min
2. Cardiac output by thermodilution: 2.29 L/min
3. The Fick method computes pulmonary blood flow (forward flow) from oxygen consumption and the arteriovenous oxygen content difference. In tricuspid regurgitation, the mixed venous sample from the pulmonary artery repr... | 4.92 | 0.1932183566158591 | 10 | 0.0029974 | {"GLM-5.2": 5, "Gemini-3.7-flash": 5, "Gemma-4-31b-it": 5, "Gpt-oss-120b": 5, "Grok-4.3": 4.8, "Minimax-m3": 5, "Muse-glimmer-30b": 4.4, "Nemotron-3-ultra-550b-a55b": 5, "Nemotron-3.5-lightning": 5, "Qwen3.5-9b": 5} |
q-corpus-1/q100_1/2 | Muse-spark-1.2 | 1. Cardiac output by the Fick method: 4.72 L/min
2. Cardiac output by thermodilution: 2.29 L/min
3. The Fick method computes pulmonary blood flow (forward flow) from oxygen consumption and the arteriovenous oxygen content difference. In tricuspid regurgitation, the mixed venous sample from the pulmonary artery repr... | 4.8999999999999995 | 0.19436506316150995 | 10 | 0.013889 | {"Deepseek-v4-flash-0731": 5, "GLM-5.2": 5, "Gemini-3.7-flash": 5, "Gemma-4-31b-it": 4.8, "Hy3": 5, "Minimax-m3": 4.8, "Muse-glimmer-30b": 4.4, "Nemotron-3-ultra-550b-a55b": 5, "Nemotron-3.5-lightning": 5, "Qwen3.5-9b": 5.0} |
q-corpus-1/q100_1/2 | Gpt-oss-120b | 1. Cardiac output by the Fick method: 4.72 L/min
2. Cardiac output by thermodilution: 2.29 L/min
3. The Fick method computes pulmonary blood flow (forward flow) from oxygen consumption and the arteriovenous oxygen content difference. In tricuspid regurgitation, the mixed venous sample from the pulmonary artery repr... | 4.92 | 0.10327955589886455 | 10 | 0.000264337 | {"Deepseek-v4-flash-0731": 4.8, "Gemini-3.7-flash": 5, "Gemma-4-31b-it": 4.8, "Gpt-oss-120b": 5, "Grok-4.3": 5, "Hy3": 4.8, "Minimax-m3": 5, "Nemotron-3-ultra-550b-a55b": 4.8, "Nemotron-3.5-lightning": 5, "Qwen3.5-9b": 5} |
q-corpus-1/q100_1/2 | Hy3 | 1. Cardiac output by the Fick method: 4.72 L/min
2. Cardiac output by thermodilution: 2.29 L/min
3. The Fick method computes pulmonary blood flow (forward flow) from oxygen consumption and the arteriovenous oxygen content difference. In tricuspid regurgitation, the mixed venous sample from the pulmonary artery repr... | 4.819999999999999 | 0.17511900715418255 | 10 | 0.001284624 | {"Deepseek-v4-flash-0731": 4.8, "GLM-5.2": 4.8, "Gemini-3.7-flash": 5, "Gemma-4-31b-it": 4.8, "Gpt-5.6-luna": 4.4, "Gpt-oss-120b": 4.8, "Hy3": 4.8, "Nemotron-3-ultra-550b-a55b": 4.8, "Nemotron-3.5-lightning": 5, "Qwen3.5-9b": 5} |
q-corpus-1/q100_1/2 | Gemini-3.7-flash | 1. Cardiac output by the Fick method: 4.72 L/min
2. Cardiac output by thermodilution: 2.29 L/min
3. The Fick method computes pulmonary blood flow (forward flow) from oxygen consumption and the arteriovenous oxygen content difference. In tricuspid regurgitation, the mixed venous sample from the pulmonary artery repr... | 4.8999999999999995 | 0.19436506316150995 | 10 | 0.0087495 | {"Gemini-3.7-flash": 5, "Gemma-4-31b-it": 5, "Gpt-5.6-luna": 4.8, "Gpt-oss-120b": 5, "Grok-4.3": 4.8, "Hy3": 5, "Minimax-m3": 5, "Muse-glimmer-30b": 4.4, "Nemotron-3.5-lightning": 5, "Qwen3.5-9b": 5} |
q-corpus-1/q100_1/2 | Minimax-m3 | 1. Cardiac output by the Fick method: 4.72 L/min
2. Cardiac output by thermodilution: 2.29 L/min
3. The Fick method computes pulmonary blood flow (forward flow) from oxygen consumption and the arteriovenous oxygen content difference. In tricuspid regurgitation, the mixed venous sample from the pulmonary artery repr... | 4.88 | 0.21499353995462794 | 10 | 0.002504688 | {"Deepseek-v4-flash-0731": 5, "Gemma-4-31b-it": 5, "Gpt-5.6-luna": 4.6, "Grok-4.3": 5, "Hy3": 4.8, "Minimax-m3": 5, "Muse-glimmer-30b": 4.4, "Nemotron-3-ultra-550b-a55b": 5, "Nemotron-3.5-lightning": 5, "Qwen3.5-9b": 5} |
q-corpus-1/q100_1/2 | Nemotron-3.5-lightning | 1. Cardiac output by the Fick method: 4.72 L/min
2. Cardiac output by thermodilution: 2.29 L/min
3. The Fick method computes pulmonary blood flow (forward flow) from oxygen consumption and the arteriovenous oxygen content difference. In tricuspid regurgitation, the mixed venous sample from the pulmonary artery repr... | 3.66 | 1.4938392000331078 | 10 | 0.00165888 | {"Deepseek-v4-flash-0731": 4.8, "GLM-5.2": 1, "Gemma-4-31b-it": 4.8, "Gpt-5.6-luna": 3.6, "Gpt-oss-120b": 5, "Grok-4.3": 4.8, "Hy3": 2, "Minimax-m3": 3.6, "Muse-glimmer-30b": 2, "Nemotron-3-ultra-550b-a55b": 5} |
q-corpus-1/q100_1/2 | GLM-5.2 | 1. Cardiac output by the Fick method: 4.72 L/min
2. Cardiac output by thermodilution: 2.29 L/min
3. The Fick method computes pulmonary blood flow (forward flow) from oxygen consumption and the arteriovenous oxygen content difference. In tricuspid regurgitation, the mixed venous sample from the pulmonary artery repr... | 4.777777777777778 | 0.23333333333333317 | 9 | 0.0079126125 | {"Deepseek-v4-flash-0731": 4.8, "Gemini-3.7-flash": 5, "Gemma-4-31b-it": 4.8, "Gpt-5.6-luna": 4.8, "Grok-4.3": 5, "Minimax-m3": 4.8, "Muse-glimmer-30b": 4.4, "Nemotron-3-ultra-550b-a55b": null, "Nemotron-3.5-lightning": 4.4, "Qwen3.5-9b": 5} |
q-corpus-1/q100_1/2 | Grok-4.6 | 1. Cardiac output by the Fick method: 4.72 L/min
2. Cardiac output by thermodilution: 2.29 L/min
3. The Fick method computes pulmonary blood flow (forward flow) from oxygen consumption and the arteriovenous oxygen content difference. In tricuspid regurgitation, the mixed venous sample from the pulmonary artery repr... | 4.688888888888889 | 0.22607766610417557 | 9 | 0.019258 | {"GLM-5.2": 4.6, "Gemini-3.7-flash": 5, "Gemma-4-31b-it": 4.6, "Gpt-5.6-luna": 4.8, "Gpt-oss-120b": 4.8, "Hy3": 4.6, "Minimax-m3": 4.8, "Muse-glimmer-30b": 4.2, "Nemotron-3-ultra-550b-a55b": null, "Qwen3.5-9b": 4.8} |
q-corpus-1/q100_1/2 | Deepseek-v4-flash-0731 | 1. Cardiac output by the Fick method: 4.72 L/min
2. Cardiac output by thermodilution: 2.29 L/min
3. The Fick method computes pulmonary blood flow (forward flow) from oxygen consumption and the arteriovenous oxygen content difference. In tricuspid regurgitation, the mixed venous sample from the pulmonary artery repr... | 4.819999999999999 | 0.17511900715418255 | 10 | 0.00073626 | {"Deepseek-v4-flash-0731": 5, "GLM-5.2": 5, "Gemma-4-31b-it": 5, "Gpt-oss-120b": 4.8, "Grok-4.3": 4.8, "Hy3": 4.8, "Minimax-m3": 4.8, "Muse-glimmer-30b": 4.8, "Nemotron-3-ultra-550b-a55b": 4.8, "Nemotron-3.5-lightning": 4.4} |
q-corpus-1/q100_1/2 | Grok-4.3 | 1. Cardiac output by the Fick method: 4.72 L/min
2. Cardiac output by thermodilution: 2.29 L/min
3. The Fick method computes pulmonary blood flow (forward flow) from oxygen consumption and the arteriovenous oxygen content difference. In tricuspid regurgitation, the mixed venous sample from the pulmonary artery repr... | 4.71111111111111 | 0.26666666666666655 | 9 | 0.0085456 | {"Deepseek-v4-flash-0731": 4.8, "Gemma-4-31b-it": 4.8, "Gpt-5.6-luna": 4.8, "Grok-4.3": 5, "Hy3": 4.6, "Minimax-m3": 4.8, "Muse-glimmer-30b": 4.2, "Nemotron-3-ultra-550b-a55b": null, "Nemotron-3.5-lightning": 4.4, "Qwen3.5-9b": 5} |
q-corpus-1/q100_1/2 | Qwen3.8-max | 1. Cardiac output by the Fick method: 4.72 L/min
2. Cardiac output by thermodilution: 2.29 L/min
3. The Fick method computes pulmonary blood flow (forward flow) from oxygen consumption and the arteriovenous oxygen content difference. In tricuspid regurgitation, the mixed venous sample from the pulmonary artery repr... | 4.911111111111111 | 0.1452966314513559 | 9 | 0.01888 | {"Deepseek-v4-flash-0731": 5, "GLM-5.2": 4.8, "Gpt-5.6-luna": 4.6, "Gpt-oss-120b": 5, "Grok-4.3": 5, "Hy3": 4.8, "Minimax-m3": 5, "Nemotron-3-ultra-550b-a55b": null, "Nemotron-3.5-lightning": 5, "Qwen3.5-9b": 5} |
q-corpus-1/q100_1/2 | Kimi-k2.6 | 1. Cardiac output by the Fick method: 4.72 L/min
2. Cardiac output by thermodilution: 2.29 L/min
3. The Fick method computes pulmonary blood flow (forward flow) from oxygen consumption and the arteriovenous oxygen content difference. In tricuspid regurgitation, the mixed venous sample from the pulmonary artery repr... | 4.879999999999999 | 0.1932183566158591 | 10 | 0.05816703098 | {"Deepseek-v4-flash-0731": 4.8, "GLM-5.2": 4.8, "Gemini-3.7-flash": 5, "Gemma-4-31b-it": 5, "Gpt-5.6-luna": 4.8, "Grok-4.3": 5, "Minimax-m3": 5, "Muse-glimmer-30b": 4.4, "Nemotron-3-ultra-550b-a55b": 5, "Nemotron-3.5-lightning": 5} |
q-corpus-1/q100_1/2 | Qwen3.5-9b | 1. Cardiac output by the Fick method: 4.72 L/min
2. Cardiac output by thermodilution: 2.29 L/min
3. The Fick method computes pulmonary blood flow (forward flow) from oxygen consumption and the arteriovenous oxygen content difference. In tricuspid regurgitation, the mixed venous sample from the pulmonary artery repr... | 4.779999999999999 | 0.23944379994757292 | 10 | 0.0018798 | {"Deepseek-v4-flash-0731": 4.8, "GLM-5.2": 4.8, "Gemini-3.7-flash": 5, "Gemma-4-31b-it": 4.8, "Gpt-5.6-luna": 4.6, "Gpt-oss-120b": 4.8, "Minimax-m3": 5, "Muse-glimmer-30b": 4.2, "Nemotron-3-ultra-550b-a55b": 4.8, "Qwen3.5-9b": 5} |
q-corpus-1/q100_1/2 | Nemotron-3-ultra-550b-a55b | 1. Cardiac output by the Fick method: 4.72 L/min
2. Cardiac output by thermodilution: 2.29 L/min
3. The Fick method computes pulmonary blood flow (forward flow) from oxygen consumption and the arteriovenous oxygen content difference. In tricuspid regurgitation, the mixed venous sample from the pulmonary artery repr... | 4.679999999999999 | 0.21499353995462783 | 10 | 0.0272712 | {"Deepseek-v4-flash-0731": 4.8, "GLM-5.2": 4.8, "Gemma-4-31b-it": 4.8, "Gpt-5.6-luna": 4.8, "Grok-4.3": 4.8, "Hy3": 4.8, "Minimax-m3": 4.4, "Muse-glimmer-30b": 4.2, "Nemotron-3-ultra-550b-a55b": 4.6, "Qwen3.5-9b": 4.8} |
q-corpus-1/q100_1/3 | Muse-spark-1.2 | **How the 1965‑1980 CPI‑inflation/unemployment experience was used to justify the natural‑rate hypothesis and the monetarist critique of the Phillips curve, and how that shaped the move to rules‑based policy in the early 1980s**
---
### 1. The empirical pattern (1965‑1980)
| Year | CPI YoY (inflation) | Unemployment... | 4.8 | 0.26457513110645897 | 9 | 0.0103925 | {"Deepseek-v4-flash-0731": 5, "GLM-5.2": 5, "Gemma-4-31b-it": 5, "Gpt-5.6-luna": 4.8, "Gpt-oss-120b": 4.6, "Grok-4.3": 5, "Hy3": 5, "Muse-glimmer-30b": 4.4, "Nemotron-3-ultra-550b-a55b": null, "Nemotron-3.5-lightning": 4.4} |
q-corpus-1/q100_1/3 | Muse-glimmer-30b | **How the 1965‑1980 CPI‑inflation/unemployment experience was used to justify the natural‑rate hypothesis and the monetarist critique of the Phillips curve, and how that shaped the move to rules‑based policy in the early 1980s**
---
### 1. The empirical pattern (1965‑1980)
| Year | CPI YoY (inflation) | Unemployment... | 4.866666666666667 | 0.19999999999999993 | 9 | 0.00111782 | {"GLM-5.2": 5, "Gemini-3.7-flash": 5, "Gemma-4-31b-it": 4.8, "Gpt-5.6-luna": 4.4, "Grok-4.3": 5, "Hy3": 4.8, "Minimax-m3": 4.8, "Nemotron-3-ultra-550b-a55b": null, "Nemotron-3.5-lightning": 5, "Qwen3.5-9b": 5} |
q-corpus-1/q100_1/3 | Nemotron-3-ultra-550b-a55b | **How the 1965‑1980 CPI‑inflation/unemployment experience was used to justify the natural‑rate hypothesis and the monetarist critique of the Phillips curve, and how that shaped the move to rules‑based policy in the early 1980s**
---
### 1. The empirical pattern (1965‑1980)
| Year | CPI YoY (inflation) | Unemployment... | 4.66 | 0.6397916327472041 | 10 | 0.0033516 | {"Deepseek-v4-flash-0731": 4.6, "GLM-5.2": 5, "Gemini-3.7-flash": 5, "Gemma-4-31b-it": 5, "Gpt-5.6-luna": 3, "Gpt-oss-120b": 4.8, "Grok-4.3": 5, "Hy3": 5, "Minimax-m3": 5, "Nemotron-3.5-lightning": 4.2} |
q-corpus-1/q100_1/3 | Gemma-4-31b-it | **How the 1965‑1980 CPI‑inflation/unemployment experience was used to justify the natural‑rate hypothesis and the monetarist critique of the Phillips curve, and how that shaped the move to rules‑based policy in the early 1980s**
---
### 1. The empirical pattern (1965‑1980)
| Year | CPI YoY (inflation) | Unemployment... | 4.71111111111111 | 0.26666666666666655 | 9 | 0.00021684 | {"Deepseek-v4-flash-0731": 4.8, "GLM-5.2": 4.8, "Gemini-3.7-flash": 5, "Gemma-4-31b-it": 5, "Gpt-oss-120b": 4.4, "Grok-4.3": 4.8, "Hy3": 4.6, "Muse-glimmer-30b": 4.2, "Nemotron-3-ultra-550b-a55b": null, "Nemotron-3.5-lightning": 4.8} |
q-corpus-1/q100_1/3 | Claude-sonnet-5 | **How the 1965‑1980 CPI‑inflation/unemployment experience was used to justify the natural‑rate hypothesis and the monetarist critique of the Phillips curve, and how that shaped the move to rules‑based policy in the early 1980s**
---
### 1. The empirical pattern (1965‑1980)
| Year | CPI YoY (inflation) | Unemployment... | 4.78 | 0.2898275349237886 | 10 | 0.035142 | {"GLM-5.2": 5, "Gemma-4-31b-it": 5, "Gpt-5.6-luna": 4.4, "Gpt-oss-120b": 4.4, "Grok-4.3": 5, "Hy3": 5, "Minimax-m3": 5, "Muse-glimmer-30b": 4.4, "Nemotron-3-ultra-550b-a55b": 5, "Nemotron-3.5-lightning": 4.6} |
q-corpus-1/q100_1/3 | Qwen3.8-max | **How the 1965‑1980 CPI‑inflation/unemployment experience was used to justify the natural‑rate hypothesis and the monetarist critique of the Phillips curve, and how that shaped the move to rules‑based policy in the early 1980s**
---
### 1. The empirical pattern (1965‑1980)
| Year | CPI YoY (inflation) | Unemployment... | 4.62 | 0.6562519841239848 | 10 | 0.009972 | {"Deepseek-v4-flash-0731": 4.8, "GLM-5.2": 5, "Gemma-4-31b-it": 5, "Gpt-5.6-luna": 3, "Gpt-oss-120b": 5, "Grok-4.3": 5, "Minimax-m3": 5, "Muse-glimmer-30b": 4.2, "Nemotron-3-ultra-550b-a55b": 5, "Nemotron-3.5-lightning": 4.2} |
q-corpus-1/q100_1/3 | Grok-4.6 | **How the 1965‑1980 CPI‑inflation/unemployment experience was used to justify the natural‑rate hypothesis and the monetarist critique of the Phillips curve, and how that shaped the move to rules‑based policy in the early 1980s**
---
### 1. The empirical pattern (1965‑1980)
| Year | CPI YoY (inflation) | Unemployment... | 4.84 | 0.2458545188611435 | 10 | 0.011364 | {"Deepseek-v4-flash-0731": 5, "GLM-5.2": 5, "Gemma-4-31b-it": 5, "Gpt-5.6-luna": 4.4, "Grok-4.3": 5, "Hy3": 4.8, "Minimax-m3": 5, "Muse-glimmer-30b": 4.4, "Nemotron-3-ultra-550b-a55b": 4.8, "Nemotron-3.5-lightning": 5} |
q-corpus-1/q100_1/3 | GLM-5.3 | **How the 1965‑1980 CPI‑inflation/unemployment experience was used to justify the natural‑rate hypothesis and the monetarist critique of the Phillips curve, and how that shaped the move to rules‑based policy in the early 1980s**
---
### 1. The empirical pattern (1965‑1980)
| Year | CPI YoY (inflation) | Unemployment... | 4.88 | 0.2529822128134702 | 10 | 0.0066696 | {"Deepseek-v4-flash-0731": 5, "Gemini-3.7-flash": 5, "Gemma-4-31b-it": 5, "Gpt-5.6-luna": 4.4, "Hy3": 5, "Minimax-m3": 5, "Muse-glimmer-30b": 4.4, "Nemotron-3-ultra-550b-a55b": 5, "Nemotron-3.5-lightning": 5, "Qwen3.5-9b": 5} |
q-corpus-1/q100_1/3 | Gpt-5.6-sol | **How the 1965‑1980 CPI‑inflation/unemployment experience was used to justify the natural‑rate hypothesis and the monetarist critique of the Phillips curve, and how that shaped the move to rules‑based policy in the early 1980s**
---
### 1. The empirical pattern (1965‑1980)
| Year | CPI YoY (inflation) | Unemployment... | 4.8999999999999995 | 0.19436506316150998 | 10 | 0.1283315 | {"Deepseek-v4-flash-0731": 5, "GLM-5.2": 5, "Gemini-3.7-flash": 5, "Gemma-4-31b-it": 4.8, "Gpt-5.6-luna": 5, "Gpt-oss-120b": 4.4, "Grok-4.3": 5, "Hy3": 5, "Minimax-m3": 5, "Nemotron-3.5-lightning": 4.8} |
q-corpus-1/q100_1/3 | Grok-4.3 | **How the 1965‑1980 CPI‑inflation/unemployment experience was used to justify the natural‑rate hypothesis and the monetarist critique of the Phillips curve, and how that shaped the move to rules‑based policy in the early 1980s**
---
### 1. The empirical pattern (1965‑1980)
| Year | CPI YoY (inflation) | Unemployment... | 4.88 | 0.21499353995462794 | 10 | 0.0028256 | {"GLM-5.2": 5, "Gemini-3.7-flash": 5, "Gemma-4-31b-it": 5, "Gpt-5.6-luna": 4.6, "Grok-4.3": 5, "Hy3": 5, "Minimax-m3": 4.8, "Muse-glimmer-30b": 4.4, "Nemotron-3-ultra-550b-a55b": 5, "Qwen3.5-9b": 5} |
q-corpus-1/q100_1/3 | GLM-5.2 | **How the 1965‑1980 CPI‑inflation/unemployment experience was used to justify the natural‑rate hypothesis and the monetarist critique of the Phillips curve, and how that shaped the move to rules‑based policy in the early 1980s**
---
### 1. The empirical pattern (1965‑1980)
| Year | CPI YoY (inflation) | Unemployment... | 4.7555555555555555 | 0.29627314724385284 | 9 | 0.00458718 | {"Deepseek-v4-flash-0731": 4.8, "GLM-5.2": 4.8, "Gemini-3.7-flash": 5, "Gemma-4-31b-it": 5, "Grok-4.3": 5, "Hy3": 5, "Minimax-m3": 4.6, "Muse-glimmer-30b": 4.2, "Nemotron-3-ultra-550b-a55b": null, "Nemotron-3.5-lightning": 4.4} |
q-corpus-1/q100_1/3 | Hy3 | **How the 1965‑1980 CPI‑inflation/unemployment experience was used to justify the natural‑rate hypothesis and the monetarist critique of the Phillips curve, and how that shaped the move to rules‑based policy in the early 1980s**
---
### 1. The empirical pattern (1965‑1980)
| Year | CPI YoY (inflation) | Unemployment... | 4.779999999999999 | 0.2573367875415838 | 10 | 0.001313664 | {"Deepseek-v4-flash-0731": 5, "Gemini-3.7-flash": 5, "Gpt-5.6-luna": 4.6, "Gpt-oss-120b": 4.8, "Grok-4.3": 5, "Hy3": 4.8, "Minimax-m3": 4.8, "Muse-glimmer-30b": 4.2, "Nemotron-3-ultra-550b-a55b": 5, "Nemotron-3.5-lightning": 4.6} |
q-corpus-1/q100_1/3 | Gemini-3.7-flash | **How the 1965‑1980 CPI‑inflation/unemployment experience was used to justify the natural‑rate hypothesis and the monetarist critique of the Phillips curve, and how that shaped the move to rules‑based policy in the early 1980s**
---
### 1. The empirical pattern (1965‑1980)
| Year | CPI YoY (inflation) | Unemployment... | 4.688888888888889 | 0.36209268304000714 | 9 | 0.005365875 | {"GLM-5.2": 5, "Gemini-3.7-flash": 5, "Gemma-4-31b-it": 4.8, "Gpt-5.6-luna": 4.4, "Gpt-oss-120b": 4.4, "Hy3": 5, "Muse-glimmer-30b": 4, "Nemotron-3-ultra-550b-a55b": null, "Nemotron-3.5-lightning": 4.6, "Qwen3.5-9b": 5} |
q-corpus-1/q100_1/3 | Deepseek-v4-flash-0731 | **How the 1965‑1980 CPI‑inflation/unemployment experience was used to justify the natural‑rate hypothesis and the monetarist critique of the Phillips curve, and how that shaped the move to rules‑based policy in the early 1980s**
---
### 1. The empirical pattern (1965‑1980)
| Year | CPI YoY (inflation) | Unemployment... | 4.859999999999999 | 0.23190036174568116 | 10 | 0.00017094 | {"Gemini-3.7-flash": 5, "Gemma-4-31b-it": 5, "Gpt-5.6-luna": 4.6, "Gpt-oss-120b": 4.6, "Grok-4.3": 5, "Hy3": 5, "Minimax-m3": 5, "Nemotron-3-ultra-550b-a55b": 5, "Nemotron-3.5-lightning": 4.4, "Qwen3.5-9b": 5} |
q-corpus-1/q100_1/3 | Minimax-m3 | **How the 1965‑1980 CPI‑inflation/unemployment experience was used to justify the natural‑rate hypothesis and the monetarist critique of the Phillips curve, and how that shaped the move to rules‑based policy in the early 1980s**
---
### 1. The empirical pattern (1965‑1980)
| Year | CPI YoY (inflation) | Unemployment... | 4.76 | 0.36270588023294503 | 10 | 0.0030345 | {"Deepseek-v4-flash-0731": 5, "GLM-5.2": 5, "Gemini-3.7-flash": 5, "Gemma-4-31b-it": 5, "Gpt-5.6-luna": 4, "Gpt-oss-120b": 4.8, "Grok-4.3": 5, "Hy3": 4.4, "Muse-glimmer-30b": 4.4, "Qwen3.5-9b": 5} |
q-corpus-1/q100_1/3 | Nemotron-3.5-lightning | **How the 1965‑1980 CPI‑inflation/unemployment experience was used to justify the natural‑rate hypothesis and the monetarist critique of the Phillips curve, and how that shaped the move to rules‑based policy in the early 1980s**
---
### 1. The empirical pattern (1965‑1980)
| Year | CPI YoY (inflation) | Unemployment... | 4.64 | 0.3627058802329451 | 10 | 0.00066116 | {"Deepseek-v4-flash-0731": 4.2, "GLM-5.2": 4.8, "Gemma-4-31b-it": 5, "Gpt-5.6-luna": 4, "Gpt-oss-120b": 4.4, "Hy3": 5, "Minimax-m3": 4.8, "Nemotron-3-ultra-550b-a55b": 5, "Nemotron-3.5-lightning": 4.4, "Qwen3.5-9b": 4.8} |
q-corpus-1/q100_1/3 | Gpt-5.6-luna | **How the 1965‑1980 CPI‑inflation/unemployment experience was used to justify the natural‑rate hypothesis and the monetarist critique of the Phillips curve, and how that shaped the move to rules‑based policy in the early 1980s**
---
### 1. The empirical pattern (1965‑1980)
| Year | CPI YoY (inflation) | Unemployment... | 4.76 | 0.35023801430836515 | 10 | 0.0063154 | {"Deepseek-v4-flash-0731": 5, "GLM-5.2": 5, "Gemini-3.7-flash": 5, "Gemma-4-31b-it": 4.8, "Gpt-oss-120b": 4.4, "Grok-4.3": 5, "Muse-glimmer-30b": 4.2, "Nemotron-3-ultra-550b-a55b": 5, "Nemotron-3.5-lightning": 4.2, "Qwen3.5-9b": 5} |
End of preview. Expand in Data Studio
Autobench Router Benchmark (ARB) 2026
The public dump that goes with the Autobench white paper Do LLM Routers Pay for Themselves?.
252 long, open-ended questions. 20 models answered each one. A ten-judge committee scored every answer and we stored the billed cost. Seven routers were then graded by lookup: a router only names a model; we read that cell.
You can recompute every table in the paper from this folder. You do not need to call a router again.
- Paper: [Do LLM Routers Pay for Themselves?] (https://autobench.org/reports/do-llm-routers-pay-for-themselves)
- Live board: autobench.org
- Leaderboard Space: AutoBench/AutoBench-Leaderboard
- Method (how questions are judged): arXiv:2510.22593
This is not the March 2025 methodology set AutoBench/AutoBench_Results_20_LLMs.
What is in here
| File | Rows | What |
|---|---|---|
questions.csv |
252 | Prompt text, domain, difficulty, skill |
matrix.csv |
5,428 | Per-question × model score, judge SD, billed USD, per-judge ranks (judges_json) |
answers.csv |
~5,040 | Answer text for the original freeze cells |
decisions.csv |
all Wave passes | Router name → selected model, plus Best Answer Rate (Delta-Hit) on that cell |
holes.csv |
varies | Unscored matrix cells and unscored decisions |
pool.json |
20 models | Display names, OpenRouter ids, price bands |
frozen_params.json |
— | δ bands, top-K=3, collapse threshold 0.90, lock time 2026-08-29 |
scoreboard.json |
28 passes | The published board (paths rewritten to this dump) |
SHA256SUMS |
— | Hashes of every file in this folder |
Parquet twins (.parquet) are the same tables, for the HF viewer.
matrix.csv is the fill sibling used to grade the paper (q_corpus_1_matrix_fill_7_4.csv, 5 Sep 2026). The original freeze (29 Aug, 5,040 cells) is source=original. The 388 later cells are source=fill_7_4. Answer text is present when has_answer=true. The 388 fill cells have scores and costs; their answer text was not kept in the fill ledger.
How to read a router
- Take one row of
decisions.csv. - Look up
(prompt_id, selected_model)inmatrix.csv. - A Delta-Hit is 1 when that cell’s score is within 0.2007 of the top three in-pool scores on that question (⅔ × median judge SD). The freeze-day official width is 0.3011 (1.0×). Both are in
frozen_params.json. - Collapse is a pass that sends ≥ 90% of prompts to one model.
- ARI (Autobench Router Index) uses β = 0.2. A pass that does not beat random quality (64.7%) scores 0.
Headline finding: no commercial pass beats always calling GPT-5.6 Luna on quality and cost together.
What is not here
- Any AutoBench / Botscanner product router, training triples, or encoder.
- Raw HTTP dumps, API keys, OpenRouter generation ids.
- Scoring source code (data-only release; the lookup rules are in the paper).
Dates
- Questions frozen 28 Aug 2026. Matrix frozen 29 Aug 2026. Fill 5 Sep 2026.
- Router calls 30 Aug–1 Sep 2026 (Wave 1/2 windows as in the paper).
- Vendor products have moved since then. Old decision logs must not be mixed with new questions.
License
CC-BY-4.0 on the questions, scores, costs, and our tables. Model answers are reproduced for audit. If a vendor objects to a row, we will take that row down. Cite the paper if you use the dump.
Autobench. Do LLM Routers Route? A Pre-Registered, Cost-Aware Audit of Commercial LLM Routers. 2026.
- Downloads last month
- 427