Controllable LLM Reasoning via Sparse Autoencoder-Based Steering

AI summary

Overview

Research area: Mechanistic interpretability and controllable generation for Large Reasoning Models (LRMs), combining Sparse Autoencoders (SAEs) with activation steering.

Technical level: Advanced. The paper assumes familiarity with transformer residual streams, sparse autoencoders, logit lens, and activation-based control vectors.

Scope: The paper proposes SAE-Steering, a two-stage pipeline that identifies sparse, strategy-specific SAE features and injects them into the residual stream to control five fine-grained reasoning strategies in two 8B-parameter LRMs, and shows the same intervention can correct erroneous reasoning paths.

What This Paper Is About

Large Reasoning Models choose their own reasoning strategies as they generate long chains of thought, and that autonomous choice sometimes produces inefficient or plainly wrong reasoning paths. Existing ways to intervene — extra instructions in the prompt or control vectors built from contrastive examples — are either ignored by the model or capture several entangled concepts at once, so they cannot steer a single strategy precisely. The paper's goal is to decompose the model's hidden states into disentangled features with a Sparse Autoencoder, find the few features that correspond to each reasoning strategy, and use those features as control vectors.

Key Contributions

1. The authors use SAEs to disentangle strategy-entangled hidden states and identify strategy-specific features, which they argue overcomes the concept entanglement problem that limits prior control-vector methods.
2. They propose SAE-Steering, a two-stage feature identification pipeline that first recalls candidate features by their logit contribution to strategy-specific keywords (filtering out over 99% of features) and then ranks the survivors by measured control effectiveness on a small validation set.
3. They validate the method on five reasoning strategies — Problem Understanding, Procedural Planning, Backtracking, Multi-Perspective Verification, and Hypothesis Reasoning — across two LRM architectures and two evaluation datasets, reporting that SAE-Steering outperforms baselines by over 15% in control effectiveness on average.
4. They demonstrate that controlling reasoning strategies can redirect an LRM from an erroneous path to a correct one, reporting a 7% absolute accuracy improvement over Budget Forcing, plus a preliminary negative-steering result that suppresses a strategy and shortens reasoning length.

Main Findings

• Superior control effectiveness: On AIME, SAE-Steering reaches an average control effectiveness of 0.76 versus 0.62 for Vector Steering, 0.43 for Think Intervention, and 0.41 for Logit Boosting. On GPQA the averages are 0.87, 0.72, 0.51, and 0.49 respectively. The paper reports an average improvement of 15% over Vector Steering.
• Activation-based methods beat prompt-based ones: Vector Steering and SAE-Steering consistently outperform Think Intervention except in some cases within Hypothesis Reasoning.
• Features encode more than keywords: SAE-Steering achieves over 35% better control effectiveness than Logit Boosting, and a case study shows that boosting the logits of the keyword "another" makes the model emit the word without changing its verification behavior, whereas the SAE feature does change the strategy.
• Cross-domain and cross-model transfer: Features identified in the math domain remain effective on GPQA, a science reasoning benchmark spanning biology, physics, and chemistry.
• Longer reasoning is harder to control: All methods score better on GPQA than on AIME, which the authors attribute to GPQA responses being roughly 7k tokens versus roughly 15k tokens for AIME.
• Logit-based recall is more precise than activation-based: Using the same number of recalled features (143 for R1-Llama-8B, 357 for Qwen3-8B), SAE-Steering reaches a precision of 0.61 ± 0.08 on R1-Llama-8B and 0.52 ± 0.05 on Qwen3-8B, against ReasonScore's 0.33 ± 0.08 and 0.27 ± 0.04 — a 28% improvement.
• Strategy features live in deeper layers: Strategy-specific features are rare in shallow layers (0, 3, 7, 11) and prevalent in deeper layers (23, 27, 31, 35) of Qwen3-8B; controls applied beyond layer 20 are strong and relatively stable, while shallow layers show poor control effectiveness.
• Negative steering works: Subtracting the feature direction instead of adding it reduces the frequency of the corresponding strategy by 30% and shortens reasoning length by 14%.
• Error correction is difficult but improved: The highest correction rate across all settings is 33% (R1-Llama-8B on MATH500 with SAE-Steering). SAE-Steering beats Budget Forcing on every model and dataset with an average absolute accuracy improvement of 7%, and beats Vector Steering with an average absolute correction-rate improvement of 5%. AIME25 is the hardest: SAE-Steering reaches 0.06 ± 0.03 for R1-Llama-8B and 0.14 ± 0.08 for Qwen3-8B.

Methodology in Plain English

The authors first train a Sparse Autoencoder on the residual stream activations of the last layer of two models — DeepSeek-R1-Distill-Llama-8B (R1-Llama-8B) and Qwen3-8B. The SAE is a TopK-SAEs configuration with a latent dimension of 65,536, an expansion factor of 16 over the model's activation size of 4,096, and K = 50. Training data is a mix of LMSYS-Chat-1M and OpenThoughts-114K.

The SAE turns each hidden state into a small set of active, human-interpretable directions. Because SAEs encourage each direction to capture one concept, these directions are better candidates for control than differences between hand-picked contrastive examples.

Finding the handful of useful directions among 65,536 is the hard part, so they do it in two stages. In Stage 1 they multiply the SAE decoder matrix by the model's unembedding matrix in a single matrix multiplication, giving each feature a score for every token in the vocabulary. They keep only features whose top-10 logit contributions include at least n = 2 strategy keywords above a threshold of τ = 0.1. This drops the candidate pool from tens of thousands to several tens. In Stage 2 they test each surviving feature on a validation set of 50 past AIME problems (1983–2023), generating a 512-token continuation with and without steering, and use GPT-4o as a judge to decide whether steering made the target strategy more explicit. The control effectiveness of a feature is the fraction of validation problems where the judge says yes, and the top-ranked feature is used at test time.

At generation time they add a scaled version of the chosen feature direction to the residual stream at the SAE-trained layer for T = 512 consecutive tokens. The steering strength starts at 15 and is decreased by one until repetitive output disappears, averaged over the validation set. For error correction they extend an already-wrong response with a "wait" token following Budget Forcing, train a strategy router to pick which strategy to apply, and run the extended reasoning at temperature 0.6 with a maximum length of 32,768 tokens. Control effectiveness evaluations use temperature 0. Test evaluation uses majority voting across GPT-4o, Gemini-2.5-flash, and Deepseek-V3.2, with a reported agreement rate of 0.82 between LLM judges and human annotators.

Why This Matters

Impact on research. The paper connects two lines of work that are usually separate: SAE-based interpretability and activation steering for reasoning control. It argues that activation-strength-based feature selection (exemplified by ReasonScore) identifies features that correlate with a behavior rather than cause it, and shows that a logit-contribution criterion is measurably more precise. It also provides layer-wise evidence that reasoning strategy features concentrate in middle-to-late layers, which is useful guidance for anyone building steering interventions.

Potential real-world applications.

• Debugging and repairing model outputs: flagging a flawed reasoning path and steering the model toward a different strategy mid-generation so it can reach a correct answer.
• Controlling reasoning cost: negative steering reduced strategy frequency by 30% and reasoning length by 14%, which points toward suppressing redundant verification and reducing overthinking.
• Domain adaptation without retraining: a strategy feature identified on math data still worked on science questions (GPQA), and the authors also report a test on the financial reasoning benchmark FinQA in their appendix.
• Compliance and safety-oriented behavior shaping: deliberately inducing or suppressing a named cognitive behavior at inference time rather than through prompt engineering.

Industry relevance. The intervention operates at inference time on frozen weights, requires only a single matrix multiplication for the recall stage, and is more reliable than prompt-based instructions that fail when reasoning contexts are long or conflict with pre-trained behavior. That makes it a candidate for deployment in serving stacks where controlled reasoning style, latency, and token budget matter. The paper reports inference latency on GPQA in its appendix but does not give a latency figure in the main text.

Future Directions

1. Scale beyond five strategies and automate keyword discovery. The authors state that only five representative strategies were evaluated and that the pipeline relies on manually specified strategy keywords, with keyword extraction taking roughly 3–4 hours of manual annotation.
2. Broaden the applications of negative steering. Only a preliminary study of suppression is reported; the authors suggest using it to alleviate overthinking and improve reasoning efficiency.
3. Intervene earlier in the reasoning trajectory. Current error correction forces the model to continue and steers only the subsequent strategy; the authors propose guiding the model at the beginning or at intermediate steps instead.
4. Handle the hardest tasks better. AIME25 correction rates remain low (0.06 ± 0.03 for R1-Llama-8B, 0.14 ± 0.08 for Qwen3-8B), leaving open how to correct errors when the task itself is very hard.

Target Audience

Researchers and engineers working on mechanistic interpretability, sparse autoencoders, and inference-time control of LLM reasoning. It is also relevant to practitioners building reasoning systems that need predictable, steerable behavior, and to readers already familiar with activation steering who want to see how SAE features compare against contrastive control vectors for fine-grained, strategy-level control. Readers without background in transformer internals or SAE training will find the method sections dense.

Authors’ abstract

Large Reasoning Models (LRMs) exhibit human-like cognitive reasoning strategies (\eg backtracking, cross-verification) during the reasoning process, which improves their performance on complex tasks. Currently, reasoning strategies are autonomously selected by LRMs themselves. However, such autonomous selection often produces inefficient or even erroneous reasoning paths. To make reasoning more reliable and flexible, it is important to develop methods for controlling reasoning strategies. Existing methods struggle to control fine-grained reasoning strategies due to conceptual entanglement in LRMs' hidden states. To address this, we leverage Sparse Autoencoders (SAEs) to decompose strategy-entangled hidden states into a disentangled feature space. To identify the few strategy-specific features from the vast pool of SAE features, we propose SAE-Steering, an efficient two-stage feature identification pipeline. SAE-Steering first recalls features that amplify the logits of strategy-specific keywords, filtering out over 99\% of features, and then ranks the remaining features by their control effectiveness. Using the identified strategy-specific features as control vectors, SAE-Steering outperforms existing methods by over 15\% in control effectiveness. Furthermore, controlling reasoning strategies can redirect LRMs from erroneous paths to correct ones, achieving a 7\% absolute accuracy improvement. Our code and data are available at https://github.com/Peter-Fy/SAE-Steering.