A mean-difference vector between execution and reflection/transition thoughts, added to hidden states, causally calibrates reasoning length and accuracy
measured in 1 paperChen et al. split reasoning traces into execution, reflection and transition thoughts and take the layer-i hidden state at the first "\n\n" token of each thought; execution thoughts are separable from the other two in a t-SNE of DeepSeek-R1-Distill-Qwen-1.5B (qualitative, Fig. 3) [chen-etal-2025-seal-steerable-reasoning-calibration] The steering vector S is the mean execution representation minus the mean reflection-plus-transition representation (1000 MATH training samples), added to the "\n\n" hidden states with strength alpha=1.0 at layer 20 (1.5B, 7B) or 55 (QwQ-32B-Preview) [chen-etal-2025-seal-steerable-reasoning-calibration] On Math500 steering raises accuracy and cuts tokens: 67.0 to 78.0 and 4526 to 3154 tokens (R1-Distill-Qwen-1.5B), 85.8 to 89.4 and 3389 to 2661 (7B), 90.4 to 91.0 and 2113 to 1716 (QwQ-32B-Preview) [chen-etal-2025-seal-steerable-reasoning-calibration] The same vector, extracted on MATH, transfers to GSM8K and LiveCodeBench, and steering in the opposite sense (weakening execution) drops 7B Math500 accuracy from 85.8 to 65.0 [chen-etal-2025-seal-steerable-reasoning-calibration]