YOPO: Frozen LLMs Answer, Steer, and Abstain in a Single Forward Pass

Forum topic · ✨步子哥 · 2026-08-17

Summary

YOPO is a method from Georgia Tech and Columbia researchers that lets a frozen language model answer questions, apply steering, and abstain when input information is insufficient—all in a single forward pass. The key insight is that steering probes (which write to the residual stream to improve reasoning accuracy, e.g., +21.7 points on Qwen2.5-1.5B) interfere with sufficiency probes that read the same residual stream to decide whether to abstain. YOPO resolves this conflict by training a small reconstruction network that recovers the clean pre-steering residual state using only MSE loss, without abstention labels. Notably, the paper finds that explicitly training abstention into a model is harmful—across 10 backbone models and 6 model families, trained abstention overfits to the training distribution and degrades cross-domain generalization. On Qwen2.5-1.5B's α-NLI task, YOPO's single forward pass (0.798) outperforms both a two-pass reference baseline (0.753) and either component alone. On MuSiQue multi-hop QA, it lifts answerable-question accuracy from 0.062 to 0.313, outperforming chain-of-thought extraction. The paper also includes an unusually candid audit of dataset artifacts in α-NLI. Paper: arXiv:2608.14465.

Overview

Imagine a retired professor: the knowledge is still there, but evidence he remembers never quite reaches his lips, and he never says "I don't know"—instead he confidently makes things up. That's a frozen language model. Its residual stream encodes evidence needed for reasoning, but that evidence doesn't reach the output layer; and when input is insufficient, it doesn't abstain.

A team from Georgia Tech and Columbia unified two previously separate research lines—extracting evidence from the residual stream, and teaching models to abstain—and found they conflict on the same residual stream.

Paper: arXiv:2608.14465 · HTML full text

Key design: two operations, one vessel

YOPO applies two operations to the residual stream simultaneously:

Conditional steering probe: writes into intermediate layers, recovering reasoning accuracy (+21.7 points on Qwen2.5-1.5B, +4.96 on 7B; McNemar p=1.7×10⁻⁷).
Zero-shot sufficiency direction d: reads the residual stream to decide whether input information is sufficient; if not, the model abstains.

The problem: writing changes what is read. Steering shifts the residual stream, corrupting the sufficiency readout—cross-domain AUROC dropped 8 points on smaller models. The naive fix is two forward passes (one steered for answering, one clean for reading), doubling inference cost.

YOPO's solution: train a small network to reconstruct the pre-steering residual state from the steered one, using only MSE loss on (steered, clean) residual pairs—no abstention labels required. Because the reconstruction never sees abstention labels, it inherits the original direction's cross-task stability.

"Abstention should not be trained in"

The paper's most counterintuitive finding: training abstention into a model is harmful. Across 10 backbones and 6 model families, the authors map a full capacity-transfer frontier. Interference concentrates in small models; at 7B there's even a reversal—the steered residual stream is itself a better readout substrate (0.995 vs 0.985 in-domain AUROC), so the identity-initialized reconstruction network learns to do nothing.

This echoes the "the model already knows" line of work: the residual stream already encodes sufficiency signals—you only need to read them, not train them. Trained abstention overfits to the training distribution and collapses out of domain.

One forward pass, three tasks

Results on Qwen2.5-1.5B, α-NLI task (all-three-correct rate):

| Configuration | Score | |---|---| | Frozen baseline | 0.375 | | Steering only | 0.590 | | Gating only | 0.560 | | Two-pass reference | 0.753 | | YOPO, single pass | 0.798 |

The single-pass system beats both individual components and the two-pass baseline, across three scales (0.798/0.830/0.893 vs 0.753/0.790/0.863). With supervised gating, in-domain AUROC reaches 0.982/0.984/0.997; at 3B and 7B, cross-domain AUROC sits within 0.006/0.003 of the two-pass ceiling.

Honest self-audit

A section titled "What Failed" admits the α-NLI dataset contains surface artifacts—labels are linearly separable at early layers, indicating a leaked shortcut. Rather than hiding it, the authors use early-layer separability as a shortcut-detection signal and re-anchor all architectural claims on native-label datasets (SQuAD2, RepLiQA, MuSiQue).

Comparison with chain-of-thought

CoT is another way to extract existing knowledge, but its autoregressive drafts can never reach YOPO's prefill-time gating, which decides before any draft is generated. On MuSiQue multi-hop QA, YOPO lifts answerable-question accuracy from 0.062 to 0.313—a 5× improvement—suggesting the residual stream encodes more evidence than CoT can surface.

Conceptual lineage

YOPO joins a growing consensus (SOPHIA's residual directions, sycophancy directions, etc.): the answer is already encoded inside the model—the question is how to read it out. YOPO pushes level-shifting to the extreme: not output-layer post-processing, not fine-tuning, but simultaneous reading and writing at the residual-stream level, decoupled via a reconstruction network.

---

Paper: arXiv:2608.14465 Code: no public repository mentioned in the paper HTML full text: arxiv.org/html/2608.14465v1