Argument Generation on APIGen sampled (test)
88.1Argument F1 (2 calls)TokMem (w/ adapt)
Evaluation results
| Method | Links | ||||
|---|---|---|---|---|---|
| TokMem (w/ adapt)Model=Llama 3.1 8B, #Params=0.20M2025.10 | 88.1 | 86.5 | 93.4 | 89.3 | |
| Fine-TuingModel=Llama 3.2 3B, #Params=2.29M2025.10 | 87.9 | 86.6 | 82.9 | 85.8 | |
| Fine-TuingModel=Llama 3.1 8B, #Params=3.41M2025.10 | 87.7 | 86.8 | 88.2 | 87.6 | |
| TokMem (w/ adapt)Model=Llama 3.2 3B, #Params=0.15M2025.10 | 85.9 | 86.7 | 88.3 | 86.3 | |
| TokMem (w/ adapt)Model=Llama 3.2 1B, #Params=0.10M2025.10 | 84.3 | 84.3 | 87.8 | 85.5 | |
| Fine-TuingModel=Llama 3.2 1B, #Params=0.85M2025.10 | 77.3 | 72.6 | 55.8 | 68.6 | |
| TokMem (w/o adapt)Model=Llama 3.2 1B, #Params=0.10M2025.10 | 68.9 | 61.1 | 73 | 67.7 | |
| TokMem (w/o adapt)Model=Llama 3.1 8B, #Params=0.20M2025.10 | 65.8 | 56.7 | 65.9 | 62.8 | |
| TokMem (w/o adapt)Model=Llama 3.2 3B, #Params=0.15M2025.10 | 65.4 | 57.2 | 50.2 | 57.6 | |
| RAGModel=Llama 3.2 3B, #Params=–2025.10 | 54.8 | 53.1 | 62.7 | 56.9 | |
| RAGModel=Llama 3.1 8B, #Params=–2025.10 | 53.3 | 57.1 | 69.2 | 59.9 | |
| ICLModel=Llama 3.1 8B, #Params=–2025.10 | 51.5 | 52.6 | 57.3 | 53.8 | |
| ICLModel=Llama 3.2 3B, #Params=–2025.10 | 42.2 | 42.3 | 38.8 | 44.1 | |
| RAGModel=Llama 3.2 1B, #Params=–2025.10 | 7.2 | 1 | 0 | 2.7 | |
| ICLModel=Llama 3.2 1B, #Params=–2025.10 | 0.6 | 0.7 | 0 | 0.4 |