Commonsense Reasoning on HellaSwag (Accuracy, Exit Position)
59.1AccuracyBackbone
Evaluation results
| Method | Links | ||
|---|---|---|---|
| BackboneBackbone Model=Llama3.1 8B (32 layers)2026.04 | 59.1 | — | |
| River-LLMBackbone Model=Llama3.1 8B (32 layers), Threshold (τ)=0.72026.04 | 58.6 | 8.8 | |
| River-LLMBackbone Model=Llama3.1 8B (32 layers), Threshold (τ)=0.52026.04 | 58.5 | 3.09 | |
| Ministral3 8BModel Architecture=Ministral3 8B (34 layers), Evaluation Protocol=Backbone2026.04 | 58.5 | — | |
| River-LLMModel Architecture=Ministral3 8B (34 layers), Early-exit threshold (tau)=0.8, Evaluation Protocol=Adaptive2026.04 | 58.2 | 2 | |
| River-LLMModel Architecture=Ministral3 8B (34 layers), Early-exit threshold (tau)=0.9, Evaluation Protocol=Adaptive2026.04 | 58.2 | 2.38 | |
| Phi4-miniModel Architecture=Phi4-mini (32 layers), Evaluation Protocol=Backbone2026.04 | 54.4 | — | |
| River-LLMModel Architecture=Phi4-mini (32 layers), Early-exit threshold (tau)=0.9, Evaluation Protocol=Adaptive2026.04 | 53.9 | 4.14 | |
| River-LLMModel Architecture=Phi4-mini (32 layers), Early-exit threshold (tau)=0.8, Evaluation Protocol=Adaptive2026.04 | 53.7 | 2.05 | |
| BackboneBackbone Model=Llama3.2 1B (16 layers)2026.04 | 45.1 | — | |
| River-LLMBackbone Model=Llama3.2 1B (16 layers), Threshold (τ)=0.72026.04 | 44.9 | 10.94 | |
| River-LLMBackbone Model=Llama3.2 1B (16 layers), Threshold (τ)=0.52026.04 | 44.3 | 3.71 |