Base Models Know How to Reason, Thinking Models Learn When

TL;DR

The paper investigates why reasoning-focused thinking models outperform base language models and whether base models already possess latent reasoning capabilities. It introduces a hybrid approach that steers a base model to perform reasoning steps by activating learned steering vectors at times determined by a thinking-model-like controller. An unsupervised taxonomy of reasoning mechanisms is derived using Top-K Sparse Autoencoders, enabling interpretable, independent reasoning categories. Across GSM8K and MATH500 benchmarks, the method recovers up to 91% of the thinking-model gap without weight updates, suggesting RLVR primarily teaches when to deploy existing skills rather than acquiring new ones.

Abstract

Why do thinking language models like DeepSeek R1 outperform their base counterparts? Despite consistent performance gains, it remains unclear to what extent thinking models learn entirely new reasoning capabilities or repurpose pre-existing base model ones. In this work, we propose a hybrid model where we activate reasoning mechanisms in base models at the right time to elicit thinking-model-level reasoning chains, implying that thinking models exploit already existing capabilities. To ground our analysis, we introduce an unsupervised, bottom-up approach for uncovering human-interpretable reasoning behaviors in thinking models. This approach provides an unbiased method to discover reasoning behaviors without imposing manual or LLM-derived assumptions. Across three base and four thinking models, using GSM8K and MATH500, our hybrid model recovers up to 91% of the performance gap to thinking models without any weight updates while steering only 12% of tokens. Concretely, our empirical setup provides a simple, causal way to test the effectiveness of existing reasoning mechanisms in base models by invoking them directly and measuring the resulting task performance. More broadly, these results reframe our understanding of how thinking models are trained: pre-training is when models acquire most of their reasoning mechanisms, and post-training teaches efficient deployment of these mechanisms at the right time, enabling efficient use of their inference-time compute.

Figures (3)

• Figure 1: Hybrid Models Unlock Reasoning Model-Level Behavior with Minimal Intervention. Overview of our approach for steering base language models to reason like thinking models. (A) We use the base model as the primary generator of tokens in the output sequence. (B) For each token position, we evaluate the current rollout in a target thinking model and use a "thinking model activation classifier" to detect the reasoning mechanism that should be applied next. (C) When the classifier detects a reasoning step, we apply a corresponding steering vector to the base model's activations, triggering structured reasoning behavior. This approach shows that base models already possess latent reasoning abilities, and that these can be reliably activated without any parameter updates, bridging much of the gap to full reasoning models with minimal extra machinery.
• Figure 2: Grid search results comparing the performance of Sparse Autoencoder taxonomies across the five thinking models in our taxonomy evaluation. Each heatmap shows the combined score (average of completeness, independence, and consistency) for different combinations of layer locations (x-axis) and cluster sizes ranging from $5$ to $50$ with increments of $5$ (y-axis). Values highlighted in blue indicate particularly strong performing configurations, while red ones indicate poor-performing configurations. Optimal configurations are typically found in the upper portions of the grid ($0.70$-$0.80$). Despite the common pattern of high scores on large cluster sizes, we find "elbow" scores at cluster sizes between $10$ and $20$, suggesting that reasoning mechanisms are reasonably well represented using $\mathbf{10}$ to $\mathbf{20}$ categories.
• Figure 3: Hybrid model in action. Example of a hybrid model (Qwen2.5-32B as the base model with steering vectors trained on QwQ-32B thinking model) solving a MATH500 problem, showing how steering vectors are dynamically applied based on SAE activations to guide the base model's reasoning process. The model successfully identifies and applies appropriate reasoning mechanisms at each step, demonstrating the effectiveness of our approach in practice.