You Only Pass Once: Answering and Abstaining Together in a Single Forward Pass of a Frozen Language Model

TL;DR - YOPO enables frozen language models to improve reasoning and detect unanswerable inputs in a single forward pass. It reconstructs the pre-steering residual state to prevent reasoning interventions from degrading abstention reliability.

• Combines conditional residual-stream steering with zero-shot sufficiency detection.
• Uses a small label-free reconstruction network trained on paired clean and steered residuals.
• More than doubles three-way accuracy on 1.5B Qwen2.5 alphaNLI, from 0.375 to 0.798.
• Outperforms a two-pass reference across model scales and preserves abstention under domain transfer.