arXivPaperNeeds Review

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

YOPO combines reasoning steering and answer abstention in one pass of a frozen model. Reconstructing the clean residual avoids interference and beats the paper's two-pass reference at three model sizes.

arXiv · Aug 14, 2026

Open Source Open Markdown Open JSON

Source Summary

YOPO steers a frozen Qwen2.5 model and detects insufficient evidence in **one forward pass**. A small network reconstructs the pre-steering residual so a fixed, label-free sufficiency direction can read it without the steering shift.

Practical Implication

Builders of local reasoning systems can treat answering and abstention as coupled operations sharing internal state. If one intervention corrupts another probe, reconstructing the original representation may avoid a second inference pass and its added cost.

Agent-Ready Context

YOPO steers a frozen Qwen2.5 model and detects insufficient evidence in **one forward pass**. A small network reconstructs the pre-steering residual so a fixed, label-free sufficiency direction can read it without the steering shift.

Builders of local reasoning systems can treat answering and abstention as coupled operations sharing internal state. If one intervention corrupts another probe, reconstructing the original representation may avoid a second inference pass and its added cost.

One-pass accuracy is **0.798/0.830/0.893** across 1.5B, 3B, and 7B models, versus **0.753/0.790/0.863** for the two-pass reference. An alphaNLI artifact was found, so the architectural claims rely on replications using native-label datasets.

Connected Context · Feed7 Judgment

YOPO shows that answering and evidence-based abstention need not require separate inference passes: a lightweight reconstruction can preserve access to the frozen model’s pre-steering state. This adds an efficiency-oriented reliability mechanism for local Qwen2.5 models and reveals an implementation constraint—one internal intervention can corrupt another probe. The alphaNLI artifact and limited model family make replication and configuration-specific evaluation essential.

Inducing language models to assert their own consciousness restores human beliefs and valuesBoth expose interference from model interventions: alignment can shift unrelated judgments, while YOPO finds that answer steering distorts the representation used by a sufficiency probe and introduces reconstruction to isolate that probe.Opaque Epistemic Mediation: How LLM Deployment Configurations Shape the Validation of Pseudo-ScienceThe finding that deployment configuration changes judgments makes YOPO’s abstention results a versioned dependency on checkpoint and inference setup; its one-pass scores should not be assumed to transfer across interfaces or configurations.DFM Mimir v1: An Open HRM Delivering Frontier Performance at 1B Parameters Using Only Permissible Post-Training DataMimir broadens the compact local-model selection set, while YOPO provides a reliability-and-efficiency mechanism tested only on Qwen2.5 sizes; applying it to another compact architecture would require new validation rather than assuming portability.

Context Map

modelresearch#reasoning#agent-reliability#model-selection

Uncertainty

One-pass accuracy is **0.798/0.830/0.893** across 1.5B, 3B, and 7B models, versus **0.753/0.790/0.863** for the two-pass reference. An alphaNLI artifact was found, so the architectural claims rely on replications using native-label datasets.