Task Representation Accuracy on Task Vector Evaluation Suite Llama2-13B (test)
87.69AccuracyLTV
Evaluation Results
| Method | Links | |
|---|---|---|
| LTVInjection Scenario=2) More Pos., Injection Position (P)={-5,...,-1}2025.09 | 87.69 | |
| LTVInjection Scenario=5) ICL prompts2025.09 | 84.99 | |
| LTVInjection Scenario=3) More layers, Injection Layer (L)={0,4,8,...}2025.09 | 82.25 | |
| LTVInjection Scenario=Baseline, Injection Position (P)={-1}, Injection Layer (L)={20}2025.09 | 80.33 | |
| FVInjection Scenario=5) ICL prompts2025.09 | 77.51 | |
| LTVInjection Scenario=1) Diff. Pos., Injection Position (P)={4}2025.09 | 71.53 | |
| LTVInjection Scenario=4) More layers & Pos., Injection Position (P)={-5,...}, Injection Layer (L)={0,4,...}2025.09 | 51.46 | |
| Vanilla TVInjection Scenario=5) ICL prompts2025.09 | 43.84 | |
| FVInjection Scenario=2) More Pos., Injection Position (P)={-5,...,-1}2025.09 | 42.25 | |
| FVInjection Scenario=Baseline, Injection Position (P)={-1}, Injection Layer (L)={20}2025.09 | 41.59 | |
| FVInjection Scenario=3) More layers, Injection Layer (L)={0,4,8,...}2025.09 | 36.97 | |
| Vanilla TVInjection Scenario=Baseline, Injection Position (P)={-1}, Injection Layer (L)={20}2025.09 | 27.67 | |
| FVInjection Scenario=4) More layers & Pos., Injection Position (P)={-5,...}, Injection Layer (L)={0,4,...}2025.09 | 24.74 | |
| Vanilla TVInjection Scenario=3) More layers, Injection Layer (L)={0,4,8,...}2025.09 | 20.46 | |
| Vanilla TVInjection Scenario=2) More Pos., Injection Position (P)={-5,...,-1}2025.09 | 16.42 | |
| Vanilla TVInjection Scenario=4) More layers & Pos., Injection Position (P)={-5,...}, Injection Layer (L)={0,4,...}2025.09 | 16.07 | |
| Vanilla TVInjection Scenario=1) Diff. Pos., Injection Position (P)={4}2025.09 | 1.84 | |
| FVInjection Scenario=1) Diff. Pos., Injection Position (P)={4}2025.09 | 1.22 |