Resolving Steering and Abstention Conflicts in Frozen LLMs

The Problem: Accuracy vs. Efficiency Dilemma in Frozen LLMs

When utilizing frozen language models for reasoning tasks, developers encounter two coupled weaknesses. First, these models under-use the evidence already encoded within their own residual streams. Second, they fail to detect when the input is insufficient to answer, so it confabulates.

To address these issues, two research lines are commonly applied separately. A conditional steering probe writes to the residual stream at mid-stack layers to recover reasoning accuracy from a frozen backbone. Meanwhile, a zero-shot sufficiency direction reads the residual stream to detect when information is insufficient and the model should abstain. However, when both techniques are deployed together in a single forward pass, interference occurs that degrades the performance of both.

Why Interference Occurs in a Single Forward Pass

The core issue lies in the manipulation of the residual stream. A steering probe writes information into mid-stack layers to enhance reasoning, but this action inadvertently shifts the state that the sufficiency direction needs to read for determining input sufficiency.

If you attempt to run both processes separately to avoid interference, inference costs double because a second forward pass is required. Conversely, running them simultaneously without mitigation causes a loss of up to 8 AUROC points of cross-domain transfer on small models. This creates a practical dilemma: high accuracy with doubled cost, or high efficiency with degraded accuracy.

This dilemma is particularly relevant for developers deploying models in production. Doubling inference cost is impractical at scale, while a significant AUROC drop undermines the very abstention system designed to prevent unreliable answers.

The Solution: Implementing the YOPO Architecture

YOPO (You Only Pass Once) combines three functions—answering, steering, and abstaining—into a single forward pass of a frozen model. The approach is tested on Qwen2.5 backbones at 1.5B, 3B, and 7B scales.

1. Keeping the Abstention Direction Fixed

Instead of retraining the sufficiency direction, YOPO keeps the detection direction in its original state. This ensures the direction remains capable of reading the information required to determine whether the input is sufficient, without being affected by changes introduced by steering.

2. Residual Stream Reconstruction via a Small Network

The key component of YOPO is training a small network to reconstruct the original (pre-steering) residual stream from the steered residual stream.

Implementation steps:

3. Single-Pass Execution

Once the reconstruction network is trained, the system operates as follows:

This approach allows all three operations to run without interference because the sufficiency direction reads the pre-steering reconstruction, not the manipulated state.

Results and Evaluation Considerations

According to the paper, YOPO reports improved three-way accuracy compared to the frozen baseline. However, it is important to note that the paper itself identifies that their alphaNLI construction was leaking a surface artifact. Therefore, the main architectural claims are anchored on native-label replications using SQuAD2, RepLiQA, and MuSiQue datasets.

YOPO reports that a single pass outperforms the two-pass reference at every model scale tested and across ten backbones from six different model families. The paper also states that the label-free direction is the only gate family to survive domain transfer, and contributes an answer-or-abstain benchmark across four domains.

For developers looking to adopt this approach, note that results on alphaNLI may not fully reflect the architecture's capabilities due to the data leakage issue identified by the authors themselves. Evaluation on native-label datasets such as SQuAD2, RepLiQA, and MuSiQue provides a more reliable picture of the system's actual performance.