Model Whisper: Steering Vectors Unlock Large Language Models' Potential in Test-time

TL;DR

The paper introduces Test-Time Steering Vectors (TTSV), a lightweight, input-side mechanism that prepends learnable vectors to input embeddings to steer frozen LLMs toward task-relevant internal states by minimizing output entropy. Through entropy-based optimization on unlabeled test data, TTSV activates latent reasoning without parameter updates, achieving substantial gains on math and reasoning benchmarks and demonstrating strong cross-distribution generalization. Theoretical analysis shows how TTSV injects an initial bias in the first attention layer that is amplified through depth, complemented by empirical evidence and ablations. The approach provides a practical, plug-and-play alternative to full fine-tuning for efficient test-time adaptation of LLMs, with broad implications for enabling robust, task-specific performance in diverse settings.

Abstract

It is a critical challenge to efficiently unlock the powerful reasoning potential of Large Language Models (LLMs) for specific tasks or new distributions. Existing test-time adaptation methods often require tuning model parameters, which is not only computationally expensive but also risks degrading the model's pre-existing abilities.To address this, we introduce a lightweight component, Test-Time Steering Vectors (TTSV), which is prepended to the input while keeping the LLM's parameters entirely frozen. By optimizing the TTSV on test data to minimize the model's output entropy, we steer the model towards an internal state of higher confidence, activating its inherent abilities most relevant to the current task. TTSV is both lightweight and highly efficient to optimize, making it a true plug-and-play enhancement. Extensive experiments validate our approach's effectiveness on both base models and reasoning-enhanced models. For instance, on the MATH500 task, TTSV achieves a 45.88% relative performance gain on the Qwen2.5-Math-7B model and a 16.22% relative gain on the Qwen3-4B model. Furthermore, our approach exhibits robust generalization, with its steering vectors proving highly transferable across diverse tasks.

Figures (7)

• Figure 1: The overall framework of our proposed Model Whisper method. The upper path (red) illustrates the optimization phase, where the TTSV is iteratively updated by minimizing the model's output entropy. The lower path (blue) shows the inference phase, where the fixed, optimized TTSV guides the model to make low-entropy, high-confidence predictions.
• Figure 2: Training dynamics of Qwen2.5-Math-1.5B on MATH500. The plot illustrates the change in training loss (left y-axis, red line) and accuracy (right y-axis, blue line) over 20 epochs.
• Figure 3: The effect of TTSV length on the accuracy of Qwen2.5-Math-1.5B on the MATH500 benchmark.
• Figure 4: Cross-distribution generalization of TTSV. The figure compares the performance of the Qwen2.5-Math-7B model with In-Distribution (ID) and Out-of-Distribution (OOD) TTSV applications across four math datasets.
• Figure 5: t-SNE visualization of the final layer activations from the Qwen2.5-math-7B model on 20 problems sampled from Minerva Math.