Draft De-anchoring Decoding Does Not Mitigate Contextual Drag in LLM Reasoning

Abstract

Large language models in iterative reasoning workflows are susceptible to \emph{contextual drag}---erroneous information in context biases subsequent generations even when the model is instructed to verify first. We propose Draft De-anchoring Decoding (D3), a training-free method that interpolates logits from draft-present and draft-absent key-value caches to attenuate harmful anchoring. Evaluating on Game of 24 with Qwen3-8B, we find that D3 fails both pre-registered success criteria: it does not improve wrong-draft accuracy ($-$0.65pp vs.\ required $+$5pp) and loses more than allowed on correct-draft accuracy ($-$2.21pp vs.\ allowed $-$1pp). Analysis reveals a fundamental mechanism flaw: draft-absent logits are too weak ($\sim$46% accuracy) to serve as a useful reference, and the method cannot distinguish beneficial from harmful anchoring. Our negative results suggest that decoding-time logit interpolation is insufficient for mitigating contextual drag.