Research map / Large language models / Reasoning and chain of thought

Model Whisper: Steering Vectors Unlock Large Language Models’ Potential in Test-Time

AAAI 2026 · Large language models › Reasoning and chain of thought · approach: Test-time compute scaling · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes Test-Time Steering Vectors (TTSV), a short set of learnable input embeddings prepended to a frozen LLM and optimized by output-entropy minimization on unlabeled test data, yielding gains on math and reasoning benchmarks.

Problem

Test-time adaptation of LLMs usually updates model parameters, which is computationally costly and risks catastrophic forgetting; the goal is to unlock latent reasoning ability on specific tasks or distributions without modifying weights.

Approach

Prepend a small set of optimizable continuous vectors to the input embedding sequence of a frozen LLM and minimize the average token-level output entropy on unlabeled test questions, so the steering signal influences all layers rather than only the output head as in SLOT.

Key results

• On MATH500, TTSV raises Qwen2.5-Math-7B from 51.00% to 74.40% (+45.88% relative), comparable to EM's +15.00 absolute gain on the same model.
• On Qwen3-4B (thinking mode), TTSV raises MATH500 from 51.80% to 60.20% and averages 44.05 vs 38.43 across six benchmarks.
• On LLaMA-3.1-8B, TTSV yields only a marginal average gain (+0.42) over baseline, and EM's gain is +0.20.

Surprising findings

• A TTSV of length L=1 still gives a substantial improvement over the baseline on MATH500.
• An out-of-distribution TTSV (trained on MATH500) sometimes outperforms the in-distribution TTSV on the target task, e.g., 26.5% vs 22.8% on Minerva Math.

Limitations

• Gains on LLaMA-3.1-8B are negligible, and the Llama setup needed a different initialization and learning rate.
• Comparisons with SLOT and EM are limited to a few models; results are single greedy runs with no variance or significance reported.
• The theoretical analysis covers only the first attention layer and relies on empirical t-SNE evidence for deeper layers.

Experimental setup

Datasets: MATH500 Minerva Math Olympiad Bench AMC23 AIME24 GPQA Diamond

Benchmarks: MATH500 Minerva Math Olympiad Bench AMC23 AIME24 GPQA Diamond

Metrics: accuracy relative gain

Baselines: EM (entropy minimization) SLOT TTRL

Builds on: entropy minimization SLOT TENT prompt tuning

Closest papers by meaning

• Test-Time Learning for Large Language Models (ICML 2025 · similarity 0.90)
• AgentTTS: Large Language Model Agent for Test-time Compute-optimal Scaling Strategy in Complex Tasks (NeurIPS 2025 · similarity 0.89)
• Analysing the Generalisation and Reliability of Steering Vectors (NeurIPS 2024 · similarity 0.89)
• Focus On This, Not That! Steering LLMs with Adaptive Feature Specification (ICML 2025 · similarity 0.89)
• In-Place Test-Time Training (ICLR 2026 · similarity 0.89)
• Confidence-Guided Stepwise Model Routing for Cost-Efficient Reasoning (AAAI 2026 · similarity 0.88)