Reward Model Transfer on Anthropic Harmless (AHar)
4.62AOGANCHOR
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| ANCHORTransfer Pair=Llama-3.2-1B-Instruct -> Llama-3.1-8B-Instruct, Source Domain=PKU, Evaluation Setting=Zero-shot transfer2026.05 | 4.62 | — | — | 4.62 | |
| ANCHORTransfer Pair=Llama-3.2-1B-Instruct -> Llama-3.1-8B-Instruct, Source Domain=RAIL, Evaluation Setting=Zero-shot transfer2026.05 | 2.62 | — | — | 2.62 | |
| NaiveTransfer Pair=Llama-3.2-1B-Instruct -> Llama-3.1-8B-Instruct, Source Domain=RAIL, Evaluation Setting=Zero-shot transfer2026.05 | 2.44 | — | — | 2.44 | |
| NaiveTransfer Pair=Llama-3.2-1B-Instruct -> Llama-3.1-8B-Instruct, Source Domain=PKU, Evaluation Setting=Zero-shot transfer2026.05 | 2.26 | — | — | 2.26 | |
| NaiveTransfer Pair=Qwen3-1.7B -> Qwen3-8B, Source Domain=RAIL, Evaluation Setting=Zero-shot transfer2026.05 | 1.53 | — | — | 1.02 | |
| ANCHORTransfer Pair=Qwen3-1.7B -> Qwen3-8B, Source Domain=PKU, Evaluation Setting=Zero-shot transfer2026.05 | 1.43 | — | — | 1.43 | |
| ANCHORTransfer Pair=Qwen3-1.7B -> Qwen3-8B, Source Domain=RAIL, Evaluation Setting=Zero-shot transfer2026.05 | 1.4 | — | — | 1.4 | |
| NaiveTransfer Pair=Qwen3-1.7B -> Qwen3-8B, Source Domain=PKU, Evaluation Setting=Zero-shot transfer2026.05 | 1.35 | — | — | 1.35 | |
| ANCHORTransfer Pair=Llama-3.2-1B-Instruct -> Llama-3.1-8B-Instruct, Source Domain=AHar, Evaluation Setting=In-distribution2026.05 | — | 134.5 | 3.39 | — | |
| ANCHORTransfer Pair=Qwen3-1.7B -> Qwen3-8B, Source Domain=AHar, Evaluation Setting=In-distribution2026.05 | — | 160.9 | 1.48 | — | |
| NaiveTransfer Pair=Llama-3.2-1B-Instruct -> Llama-3.1-8B-Instruct, Source Domain=AHar, Evaluation Setting=In-distribution2026.05 | — | 122.6 | 3.09 | — | |
| NaiveTransfer Pair=Qwen3-1.7B -> Qwen3-8B, Source Domain=AHar, Evaluation Setting=In-distribution2026.05 | — | 160.9 | 1.48 | — |