Research map / Large language models / Agents, code and math
On the Generalization Gap in Self-Evolving Language Model Reasoning
Zhenting Qi, Susanna Maria Baby, Stefanie Anna Baby, Kan Yuan, Andrew Tomkins, Tu Vu, Da-Cheng Juan, Cyrus Rashtchian
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 |
|---|---|---|---|
| apply SimpleSE self-evolution (tau=0.6) | gemma-3-4b-it · KK | accuracy (%): 31.0 -> 40.7 | Improved |
| add multi-turn feedback revision (RevisionSE) over SimpleSE | gemma-3-12b-it · KK | accuracy (%): 47.5 -> 52.8 | Improved |
| add multi-turn feedback revision (RevisionSE) (ablation) | gemma-3-1b-it · KK | accuracy (%): 23.8 -> 22.4 | Hurt |
| add iterative rounds of SimpleSE (three rounds) (ablation) | gemma-3-4b-it · KK | accuracy (%): 31.0 -> 44.1 | Improved |
| add curriculum (2-3 then 4-5 people) instead of random mixing (ablation) | gemma-3-4b-it · KK | accuracy (%): 41.2 -> 44.8 | Improved |
| add a final oracle verification round after SimpleSE iterations (ablation) | gemma-3-4b-it · KK | accuracy (%): 43.3 -> 53.2 | Improved |
| scale model size from 1B to 4B for SimpleSE (ablation) | gemma-3 SimpleSE · KK | accuracy (%): 7.8 -> 8.4 | Improved |
| raise verifier threshold tau (ablation) | gemma-3-4b-it SimpleSE · KK | accuracy (%): 41.7 | Mixed |
| apply SimpleSE trained on OpenThoughts3 | gemma-3-4b-it · MATH500 | accuracy (%): 75.8 -> 77.4 | Improved |
| apply SimpleSE trained on OpenThoughts3 | gemma-3-4b-it · GSM8K | accuracy (%): 89.2 -> 89.0 | No clear effect |
| apply SimpleSE trained on OpenThoughts3 | Qwen2.5-7B-Instruct · KK | accuracy (%): 18.1 -> 17.6 | Hurt |
| apply SimpleSE iterations (ablation) | gemma-3-4b-it · KK Pass@32 | Pass@32 (%): 78.0 -> 78.8 | No clear effect |
Closest papers by meaning
- Learning to Self-Verify Makes Language Models Better Reasoners (ICML 2026 · similarity 0.91)
- Self-Adapting Language Models (NeurIPS 2025 · similarity 0.90)
- Incentivizing LLMs to Self-Verify Their Answers (NeurIPS 2025 · similarity 0.89)
- Scalable Best-of-N Selection for Large Language Models via Self-Certainty (NeurIPS 2025 · similarity 0.89)
- PACE: Two-Timescale Self-Evolution for Small Language Model Agents (NeurIPS 2026 · similarity 0.89)
- ReGenesis: LLMs can Grow into Reasoning Generalists via Self-Improvement (ICLR 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.
