Semantic Steering Alignment on LitCovid 5,000-article subset
0.302∆SilOurs (hybrid)
Evaluation results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| Ours (hybrid)Steering Method Type=hybrid, Mixing weight (γ)=0.5, LLM Model=gpt-5.4, α_ours=0.9, Abstention mechanism=tnone=0.202026.07 | 0.302 | 0.023 | 70 | 79 | 83 | 94 | |
| Per-item LLMLLM Model=gpt-5.4, Blending coefficient (λ_eff)=0.49 ± 0.02, Abstention mechanism=tnone=0.202026.07 | 0.259 | 0.071 | 80 | 90 | — | — | |
| Ours (seed-only)Steering Method Type=seed-only, Mixing weight (γ)=1, LLM Model=gpt-5.4, α_ours=0.9, Abstention mechanism=tnone=0.202026.07 | 0.248 | 0.01 | 64 | 78 | 77 | 94 | |
| Ours (profile-only)Steering Method Type=profile-only, Mixing weight (γ)=0, LLM Model=gpt-5.4, α_ours=0.9, Abstention mechanism=tnone=0.202026.07 | 0.19 | 0.003 | 71 | 74 | 87 | 91 |