YOPO: Frozen LMs 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, steer its reasoning, and decide when to abstain—all within a single forward pass. The approach combines two operations on the residual stream: a conditional steering probe that writes corrections into intermediate layers (boosting Qwen2.5-1.5B accuracy by +21.7 points), and a zero-shot sufficiency direction that reads the residual stream to detect insufficient evidence. Because steering distorts the readout signal, YOPO trains a small reconstruction network to recover the pre-steering state using only MSE on (steered, clean) pairs, without abstention labels. A key counterintuitive finding across 10 backbone models is that abstention should not be trained into a model: learned abstention overfits to the training distribution and collapses out of domain, whereas reading an inherent sufficiency signal generalizes better. On the α-NLI task, YOPO's single forward pass (0.798 three-way accuracy) beats both individual components and even a two-pass reference baseline. On MuSiQue multi-hop QA it raises answerable-question accuracy from 0.062 to 0.313, outperforming chain-of-thought extraction. The paper also openly audits a surface artifact in its α-NLI dataset and re-anchors claims on cleaner benchmarks.

Paper: arXiv:2608.14465

Imagine asking a retired professor to answer questions. The knowledge is still in his head, but two problems remain: evidence he clearly remembers won't make it to his mouth, and he doesn't know when to say "I don't know"—asked whether Mars has water, he might invent an answer rather than admit uncertainty.

A frozen language model is that professor. Its residual stream already encodes the evidence needed for reasoning, but that evidence never reaches the output layer; and when input information is insufficient, instead of abstaining, the model confidently makes things up. These two problems have traditionally been studied as independent research lines—one tries to "pry out" evidence from the residual stream, the other teaches the model to say "I don't know."

A team from Georgia Tech and Columbia twisted the two lines together and found an unexpected complication: on the same residual stream, they conflict.

Two Surgeries, One Blood Vessel

YOPO applies two operations to the residual stream simultaneously:

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

The problem: writing changes what reading sees. Steering pushes the residual stream off its original state, so the sufficiency direction no longer reads the original signal. On small models this interference cost 8 AUROC points cross-domain. The naive fix is two forward passes—one steered (for answering), one clean (for reading)—but that doubles inference cost.

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

"Abstention Should Not Be Trained In"

The paper's most counterintuitive finding: training abstention capability into the model is actively harmful.

The authors map a complete "capacity–transfer" frontier across 10 backbone models from 6 families. Interference concentrates in small models, but at 7B the effect inverts—the steered residual stream is itself a better readout substrate (0.995 vs 0.985 in-domain AUROC), and the identity-initialized reconstruction network learns to "do nothing."

This design principle—"abstention should not be trained in"—echoes the "the model already knows" line of work. The residual stream already encodes a sufficiency signal; you only need to read it, not train it. Once abstention is trained into the model, it overfits to the training distribution and collapses out of domain.

One Forward Pass, Three Tasks

The final system performs three things in a single forward pass: answering, steering, and abstaining.

On the α-NLI task with Qwen2.5-1.5B:

| Configuration | Three-way accuracy | |------|-----------| | Frozen baseline | 0.375 | | Steering only | 0.590 | | Gating only | 0.560 | | Two-pass reference | 0.753 | | YOPO single pass | 0.798 |

The single forward pass beats either component alone and even the two-pass reference baseline, at all three scales: 0.798/0.830/0.893 vs 0.753/0.790/0.863.

Finer numbers: on 1.5B, cross-domain AUROC improves from 0.836 to 0.888 (two-pass ceiling 0.918); in-domain from 0.913 to 0.959. With supervised gating, in-domain reaches 0.982/0.984/0.997, and on 3B and 7B the cross-domain gap to the two-pass ceiling shrinks to 0.006/0.003.

An Honest Self-Audit

The paper includes a section titled "What Failed"—rare candor. The authors admit their α-NLI dataset has surface artifacts: answer labels are linearly separable from early layers onward, meaning shortcut signals leaked into the data construction. Rather than hiding this, they used source-side auditing (early-layer separability as a shortcut-detection signal) to find it, then re-anchored all architectural claims on natively labeled datasets (SQuAD2, RepLiQA, MuSiQue).

This self-audit connects directly to the "evaluation blind spot" problem: you think you're measuring reasoning, but the dataset leaked a surface feature the model exploits. Without auditing, your conclusions rest on sand.

Comparison with CoT

Chain-of-thought is another way to extract a model's existing knowledge. But CoT's generated drafts can never reach YOPO's prefill-time gating—the gate decides during the prefill phase, before any CoT draft is finished.

On MuSiQue multi-hop QA, YOPO raises answerable multi-hop accuracy from 0.062 to 0.313—a fivefold improvement. This suggests the residual stream encodes more evidence than CoT can extract: CoT is constrained by autoregressive path dependence, while YOPO reads the residual stream's global state directly.

Conceptual Lineage

YOPO is another member of the "the model already knows" series. From SOPHIA's residual-stream directions, SWE-Pruner Pro's AUC 0.83, and the sycophancy-bias direction, to YOPO—the cross-paper consensus is strengthening: the model's internals already encode the answer; the question is how to read it out.

More specifically, YOPO pushes "solve the problem at a different layer" to a new extreme: no post-processing at the output layer, no fine-tuning at the training layer—reading and writing happen simultaneously at the intermediate residual-stream level, where the two operations interfere and require a reconstruction network to decouple.

The "abstention should not be trained in" principle resonates with i-have-adhd's "alignment doesn't necessarily require retraining." Alignment interventions are getting finer-grained: from RLHF (training layer), to Epanorthosis (inference-layer prompting), to Beyond Sycophancy (interaction structure), to i-have-adhd (output-layer skill). YOPO now says: abstention intervention belongs at the residual-stream readout layer, not in model parameters.

---

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