YOPO: Enabling Reasoning and Refusal in Frozen LLMs via Single Forward Pass

Background and Context

Frozen large language models (LLMs) face two critical, coupled limitations in reasoning tasks. First, these models fail to fully utilize the evidence already encoded within their residual streams, which restricts their reasoning accuracy. Second, they lack the ability to detect when input information is insufficient, often leading to hallucinations or nonsensical outputs when data is missing. The YOPO (You Only Pass Once) framework addresses these issues by enabling a model to answer, guide, and refuse simultaneously within a single forward pass. This approach integrates two distinct research lines: conditional guiding probes, which write into the residual stream at intermediate layers to restore reasoning precision, and zero-shot sufficiency directions, which read the residual stream to determine if the information is adequate for a reliable response.

However, deploying these mechanisms directly causes interference. The writing process of guidance alters the state that the sufficiency direction reads, resulting in degraded cross-domain transfer performance. To resolve this, YOPO trains a lightweight network to reconstruct the residual stream before guidance is applied. This reconstruction allows the model to maintain high-precision reasoning while ensuring that the refusal judgment is based on the original, undisturbed state. By solving this state interference, YOPO significantly enhances the practicality and reliability of frozen models in complex reasoning tasks without increasing inference costs.

Deep Analysis

The technical core of YOPO lies in decoupling the state interference between guidance writing and direction reading. The research team keeps the sufficiency direction fixed and trains a small network using mean squared error as the loss function. This network reconstructs the pre-guidance residual stream using pairs of (post-guidance, clean) residual streams. Notably, this process requires no sufficiency labels, relying solely on the paired data of the residual streams themselves. The reconstructed stream is then used to read the sufficiency direction, ensuring that the refusal decision is grounded in the original state, free from the negative effects of the guidance mechanism.

The system is deployed on frozen Qwen2.5 backbones of 1.5B, 3B, and 7B parameters. By completing all operations in a single forward pass, YOPO avoids the doubled inference costs associated with traditional dual-forward approaches. The design leverages the linear separability of the residual stream, using a simple reconstruction task to achieve complex state decoupling. This demonstrates a sophisticated understanding of the internal representation structure of language models, offering a novel technical pathway for enhancing the functionality of frozen models without retraining the entire backbone.

Industry Impact

Experimental results demonstrate that YOPO significantly outperforms both frozen baselines and dual-forward reference benchmarks. On the alphaNLI dataset, the three-class accuracy for the 1.5B model improved from 0.375 to 0.798, more than doubling the performance. Across all scales, YOPO achieved accuracies of 0.798, 0.830, and 0.893 for the 1.5B, 3B, and 7B models, respectively, consistently surpassing the dual-forward baselines of 0.753, 0.790, and 0.863. The method’s generalizability was further validated across ten backbones and six model families. Ablation studies revealed a capacity transfer frontier, quantifying the principle that refusal capabilities should not be trained into the model but rather achieved through architectural design.

The research team also conducted a source audit, identifying surface artifact leakage in the construction of alphaNLI. Consequently, architectural claims were anchored to native label replication using datasets such as SQuAD2, RepLiQA, and MuSiQue. In the standard four-domain suite, YOPO’s gateway performed best on all in-domain datasets. Furthermore, the no-label direction was the only gateway family that maintained performance during domain transfer, proving the method’s robustness and cross-domain applicability. This reliability is crucial for industrial scenarios where consistent performance across varied data distributions is required.

Outlook

YOPO offers profound implications for the open-source community and industrial deployment. It provides an efficient method for frozen LLMs to possess both high-precision reasoning and reliable refusal capabilities without increasing inference costs, which is particularly valuable for resource-constrained environments. By contributing the first answer-or-refuse benchmark, the work fills a significant gap in the field, providing standardized evaluation tools for future research. The finding that refusal should be an architectural feature rather than a trained behavior offers new guidance for model design and training strategies, potentially reducing the complexity of model development.

For industrial applications, the single-forward-pass nature of YOPO makes it suitable for real-time systems such as customer service bots and intelligent assistants, where low latency and cost efficiency are paramount. The residual stream reconstruction method and the analysis of the capacity transfer frontier provide new perspectives on understanding internal representations and functional decoupling in language models. This work may inspire further research into enhancing frozen models, driving the natural language processing field toward more efficient and reliable solutions for complex reasoning tasks.