AI Paper Atlas

Research map / Large language models / Instruction and fine-tuning

Base Models Know How to Reason, Thinking Models Learn When

Constantin Venhoff, Iván Arcuschin, Philip Torr, Arthur Conmy, Neel Nanda

ICML 2026 · Large language models › Instruction and fine-tuning · approach: Fine-tuning for reasoning data

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
add SAE-selected category (instead of random category) (ablation)hybrid model on ORZ-1.5B and ORZ-32B · MATH500 + GSM8Kgap recovered: 20-28% -> 77%Improved
add learned category vectors (instead of norm-matched random directions) (ablation)hybrid model on ORZ-1.5B and ORZ-32B · MATH500 + GSM8Kgap recovered: -28% and -13% -> 77%Improved
add correct category selection and learned directions to MLP coefficients (ablation)hybrid model on ORZ-1.5B and ORZ-32B · MATH500 + GSM8Kgap recovered: -8% and -1% -> 77%Improved
add SAE-gated disagreement positions (instead of random positions) (ablation)hybrid model on ORZ-1.5B and ORZ-32B · MATH500 + GSM8Kgap recovered: -8% (1.5B), 28% (32B) -> 77%Improved
use RL-trained thinking pair instead of SFT-distilled pair at 32Bhybrid model · GSM8Kgap recovered: 31.8 -> 93.6Improved
use RL-trained thinking pair instead of SFT-distilled pair at 1.5Bhybrid model · GSM8Kgap recovered: 3.0 -> 71.3Improved
use RL-trained ORZ-32B instead of mixed QwQ-32Bhybrid model · MATH500gap recovered: 14.3 -> 59.8Improved
increase base model size for SFT-distilled pairs (1.5B to 32B)hybrid model · GSM8Kgap recovered: 3.0 -> 31.8Improved
steer base model with category vectors for ORZ-32B (hybrid vs base)Qwen2.5-32B · GSM8Kaccuracy: 69.5 -> 95.5Improved
steer base model with category vectors for R1-Distill-32B (hybrid vs base)Qwen2.5-32B · GSM8Kaccuracy: 69.4 -> 78.4Mixed
increase ORZ-7B to MATH500 recovery versus GSM8Khybrid model ORZ-7B · MATH500gap recovered: 87.8 -> 49.7Mixed
train SAE with different seedsORZ-7B SAE (layer 16, dictionary 15) · feature stabilityfeature-matching rate: 93.3%No clear effect

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy