BlendIn replaces binary guidance acceptance with confidence-weighted distribution blending between base and guidance models, mitigating cascading failures in inference-time LLM alignment.
Manning, and Chelsea Finn
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.LG 1years
2026 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
To Intervene or Not: Guiding Inference-time Alignment with Probabilistic Model Blending
BlendIn replaces binary guidance acceptance with confidence-weighted distribution blending between base and guidance models, mitigating cascading failures in inference-time LLM alignment.