Multi-step reasoning on TriviaQA
80.09Task PerformanceLAR
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| LARBackbone=Qwen3-8B2026.05 | 80.09 | -27.1 | |
| ReActBackbone=Qwen3-8B2026.05 | 77.84 | — | |
| COTBackbone=Llama-3.1 8B-Instruct2026.05 | 75.5 | — | |
| VanillaBackbone=Llama-3.1 8B-Instruct2026.05 | 73.63 | — | |
| LARBackbone=Llama-3.1 8B-Instruct2026.05 | 72.46 | -23.3 | |
| COTBackbone=Qwen3-8B2026.05 | 69.43 | — | |
| TokenSkipBackbone=Llama-3.1 8B-Instruct2026.05 | 68.87 | -12 | |
| ConciseHintBackbone=Qwen3-8B2026.05 | 68.69 | -12.7 | |
| VanillaBackbone=Qwen3-8B2026.05 | 67.4 | — | |
| ConciseHintBackbone=Llama-3.1 8B-Instruct2026.05 | 67.32 | -13.5 | |
| ReActBackbone=Llama-3.1 8B-Instruct2026.05 | 59.88 | — | |
| ACONBackbone=Llama-3.1 8B-Instruct2026.05 | 57.14 | -25.4 | |
| TokenSkipBackbone=Qwen3-8B2026.05 | 57.02 | -28.7 | |
| ACONBackbone=Qwen3-8B2026.05 | 55.33 | -27.9 |