General AI Assistant Tasks on GAIA (Internalized and In-context inference metrics)
33.3Internalized Inference RateExperience Internalization
Evaluation results
| Method | Links | ||
|---|---|---|---|
| Experience InternalizationBase Model=Qwen3-4B-Instruct-2507, Experience Source=DeepSeek-generated, Injection Pattern=Step-wise injection, Distillation Regime=Off-policy distillation, Iteration=iter 32026.06 | 33.3 | — | |
| Experience InternalizationBase Model=Qwen3-4B-Instruct-2507, Experience Source=DeepSeek-generated, Injection Pattern=Step-wise injection, Distillation Regime=Off-policy distillation, Iteration=iter 22026.06 | 30.1 | 24.7 | |
| Experience InternalizationBase Model=Qwen3-8B-Instruct, Experience Source=DeepSeek-generated, Injection Pattern=Step-wise injection, Distillation Regime=Off-policy distillation, Iteration=iter 12026.06 | 30.1 | — | |
| Experience InternalizationBase Model=Qwen3-4B-Instruct-2507, Experience Source=DeepSeek-generated, Injection Pattern=Step-wise injection, Distillation Regime=Off-policy distillation, Iteration=iter 12026.06 | 29.8 | 22.6 | |
| Experience InternalizationBase Model=Qwen3-8B-Instruct, Experience Source=DeepSeek-generated, Injection Pattern=Step-wise injection, Distillation Regime=Off-policy distillation, Iteration=iter 32026.06 | 29.8 | — | |
| Experience InternalizationBase Model=Qwen3-4B-Instruct-2507, Experience Source=DeepSeek-generated, Injection Pattern=Step-wise injection, Distillation Regime=On-policy distillation, Iteration=iter 12026.06 | 28.8 | — | |
| Experience InternalizationBase Model=Qwen3-8B-Instruct, Experience Source=DeepSeek-generated, Injection Pattern=Step-wise injection, Distillation Regime=Off-policy distillation, Iteration=iter 22026.06 | 28.8 | — | |
| Experience InternalizationBase Model=Qwen3-8B-Instruct, Experience Source=DeepSeek-generated, Injection Pattern=Step-wise injection, Distillation Regime=On-policy distillation, Iteration=iter 12026.06 | 28.5 | — | |
| Experience InternalizationBase Model=Qwen3-8B-Instruct, Experience Source=DeepSeek-generated, Injection Pattern=Step-wise injection, Distillation Regime=On-policy distillation, Iteration=iter 22026.06 | 27.8 | — | |
| Experience InternalizationBase Model=Qwen3-4B-Instruct-2507, Experience Source=DeepSeek-generated, Injection Pattern=Step-wise injection, Distillation Regime=On-policy distillation, Iteration=iter 22026.06 | 27.2 | — | |
| Experience InternalizationBase Model=Qwen3-4B-Instruct-2507, Experience Source=DeepSeek-generated, Injection Pattern=Global injection, Distillation Regime=Off-policy distillation, Iteration=iter 12026.06 | 25.9 | 20.4 | |
| Experience InternalizationBase Model=Qwen3-4B-Instruct-2507, Experience Source=DeepSeek-generated, Injection Pattern=Step-wise injection, Distillation Regime=On-policy distillation, Iteration=iter 32026.06 | 25.6 | — | |
| Experience InternalizationBase Model=Qwen3-4B-Instruct-2507, Experience Source=Qwen-generated, Injection Pattern=Step-wise injection, Distillation Regime=Off-policy distillation, Iteration=iter 32026.06 | 24.6 | — | |
| Experience InternalizationBase Model=Qwen3-4B-Instruct-2507, Experience Source=Qwen-generated, Injection Pattern=Step-wise injection, Distillation Regime=Off-policy distillation, Iteration=iter 22026.06 | 24 | 23.3 | |
| Experience InternalizationBase Model=Qwen3-8B-Instruct, Experience Source=DeepSeek-generated, Injection Pattern=Step-wise injection, Distillation Regime=On-policy distillation, Iteration=iter 32026.06 | 23.9 | — | |
| Experience InternalizationBase Model=Qwen3-4B-Instruct-2507, Experience Source=DeepSeek-generated, Injection Pattern=Global injection, Distillation Regime=On-policy distillation, Iteration=iter 12026.06 | 22.8 | — | |
| Experience InternalizationBase Model=Qwen3-4B-Instruct-2507, Experience Source=Qwen-generated, Injection Pattern=Step-wise injection, Distillation Regime=Off-policy distillation, Iteration=iter 12026.06 | 22.7 | 24.9 | |
| Experience InternalizationBase Model=Qwen3-4B-Instruct-2507, Experience Source=DeepSeek-generated, Injection Pattern=Global injection, Distillation Regime=Off-policy distillation, Iteration=iter 22026.06 | 21.4 | 12 | |
| Experience InternalizationBase Model=Qwen3-4B-Instruct-2507, Experience Source=DeepSeek-generated, Injection Pattern=Global injection, Distillation Regime=On-policy distillation, Iteration=iter 22026.06 | 19.4 | — | |
| Qwen3-8B-InstructExperience Status=w/o experience2026.06 | 18.5 | — | |
| Experience InternalizationBase Model=Qwen3-4B-Instruct-2507, Experience Source=DeepSeek-generated, Injection Pattern=Global injection, Distillation Regime=On-policy distillation, Iteration=iter 32026.06 | 18.1 | — | |
| Experience InternalizationBase Model=Qwen3-4B-Instruct-2507, Experience Source=Qwen-generated, Injection Pattern=Global injection, Distillation Regime=Off-policy distillation, Iteration=iter 12026.06 | 15.9 | 10 | |
| Qwen3-4B-Instruct-2507Experience Status=w/o experience2026.06 | 13.6 | — | |
| Experience InternalizationBase Model=Qwen3-4B-Instruct-2507, Experience Source=DeepSeek-generated, Injection Pattern=Global injection, Distillation Regime=Off-policy distillation, Iteration=iter 32026.06 | 13.6 | — | |
| Experience InternalizationBase Model=Qwen3-4B-Instruct-2507, Experience Source=Qwen-generated, Injection Pattern=Global injection, Distillation Regime=Off-policy distillation, Iteration=iter 22026.06 | 11.3 | 6.2 | |
| Experience InternalizationBase Model=Qwen3-4B-Instruct-2507, Experience Source=Qwen-generated, Injection Pattern=Global injection, Distillation Regime=Off-policy distillation, Iteration=iter 32026.06 | 6.5 | — | |
| Qwen3-4B-Instruct-2507Experience Source=Qwen-generated experience2026.06 | — | 11.7 | |
| Qwen3-4B-Instruct-2507Experience Source=DeepSeek-generated experience2026.06 | — | 19.7 | |
| Qwen3-8B-InstructExperience Source=DeepSeek-generated experience2026.06 | — | 26.21 |