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
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 changed | Where | Measured | Effect |
|---|---|---|---|
| add SAE-selected category (instead of random category) (ablation) | hybrid model on ORZ-1.5B and ORZ-32B · MATH500 + GSM8K | gap 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 + GSM8K | gap 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 + GSM8K | gap 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 + GSM8K | gap recovered: -8% (1.5B), 28% (32B) -> 77% | Improved |
| use RL-trained thinking pair instead of SFT-distilled pair at 32B | hybrid model · GSM8K | gap recovered: 31.8 -> 93.6 | Improved |
| use RL-trained thinking pair instead of SFT-distilled pair at 1.5B | hybrid model · GSM8K | gap recovered: 3.0 -> 71.3 | Improved |
| use RL-trained ORZ-32B instead of mixed QwQ-32B | hybrid model · MATH500 | gap recovered: 14.3 -> 59.8 | Improved |
| increase base model size for SFT-distilled pairs (1.5B to 32B) | hybrid model · GSM8K | gap recovered: 3.0 -> 31.8 | Improved |
| steer base model with category vectors for ORZ-32B (hybrid vs base) | Qwen2.5-32B · GSM8K | accuracy: 69.5 -> 95.5 | Improved |
| steer base model with category vectors for R1-Distill-32B (hybrid vs base) | Qwen2.5-32B · GSM8K | accuracy: 69.4 -> 78.4 | Mixed |
| increase ORZ-7B to MATH500 recovery versus GSM8K | hybrid model ORZ-7B · MATH500 | gap recovered: 87.8 -> 49.7 | Mixed |
| train SAE with different seeds | ORZ-7B SAE (layer 16, dictionary 15) · feature stability | feature-matching rate: 93.3% | No clear effect |
Closest papers by meaning
- Reasoning Scaffolding: Distilling the Flow of Thought from LLMs (ICLR 2026 · similarity 0.90)
- On the Interplay of Pre-Training, Mid-Training, and RL on Reasoning Language Models (ICML 2026 · similarity 0.89)
- Facts Don't Speak Louder than Words: The Behavioral Essence of Long-CoT Distillation (NeurIPS 2026 · similarity 0.89)
- Phenomenal Yet Puzzling: Testing Inductive Reasoning Capabilities of Language Models with Hypothesis Refinement (ICLR 2024 · similarity 0.89)
- Exploring the Limits of Sub-Billion Language Model Reasoners with Open Training Recipes (ICLR 2026 · similarity 0.89)
- Eliciting Reasoning in Language Models with Cognitive Tools (NeurIPS 2025 · similarity 0.89)
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.
