Research map / Label-efficient and robust learning / Recognition and long-tail
The Generalization Spectrum: A Chromatographic Approach to Evaluating Learning Algorithms
Jinghan Zhang, Zerui Cheng, Shiqi Chen, Ge Zhang, Wenhao Huang, Jiashuo Liu, Junxian He, Tianle Cai
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 |
|---|---|---|---|
| train with RL (GRPO) instead of reference-solution SFT at matched memorization (ablation) | Qwen3-4B-Thinking · Generalization Spectrum, D0≈0.6 | near-transfer normalized gain (D1): 21.9% -> 39.0% | Improved |
| train with RL (GRPO) instead of SFT at matched memorization (ablation) | Qwen3-4B-Thinking · Generalization Spectrum D3/D4 | far-transfer normalized gain: -7.1%/-0.5% -> 12.3%/16.8% | Improved |
| use paired ICL demonstration instead of random demonstration (ablation) | Qwen3-4B-Thinking · Generalization Spectrum D0-D2 | pass@1: D2 0.21 -> 0.63 | Improved |
| paired ICL at category-matched and unpaired levels | Qwen3-4B-Thinking · Generalization Spectrum D3/D4 | pass@1: D3 0.24 -> 0.17; D4 0.22 -> 0.10 | Hurt |
| add pseudocode or key insight to ICL code demonstration (ablation) | paired ICL · Generalization Spectrum D0-D2 | pass@1: D2 64.4% -> 71.1% (pseudocode) | Improved |
| add chain-of-thought trace to ICL code demonstration (ablation) | paired ICL · Generalization Spectrum D0-D2 | pass@1: D2 64.4% -> 63.0% | No clear effect |
| add oracle correspondence hint plus pseudocode to ICL (ablation) | paired ICL · Generalization Spectrum D2 | pass@1: 64.4% -> 80.1% | Improved |
| add oracle hint to ICL at category-matched and unpaired levels (ablation) | paired ICL · Generalization Spectrum D3/D4 | pass@1 | No clear effect |
| use RFT with self-generated correct targets instead of demonstration SFT (ablation) | SFT on Qwen3-4B-Thinking · Generalization Spectrum D2-D4 | far-transfer pass@1: compile error 17.6% -> 6.8% | Improved |
| use GPT-OSS-20B off-policy targets instead of Qwen3-4B or QwQ targets for SFT (ablation) | SFT · Generalization Spectrum D1-D4 | pass@1: target perplexity 4.29 | Hurt |
| use self-distillation fine-tuning (SDFT) instead of vanilla SFT (ablation) | SFT at D0≈0.6 · Generalization Spectrum | pass@1 at D1-D2 vs D3-D4 | Mixed |
| add training-time hints (pseudocode or key idea) to GRPO (ablation) | GRPO · Generalization Spectrum, D0≈0.6 | steps to reach D0≈0.6: step 170 -> step 60 (KI), step 70 (Pseudo) | Improved |
Closest papers by meaning
- Slower Generalization, Faster Memorization: A Sweet Spot in Algorithmic Learning (NeurIPS 2026 · similarity 0.88)
- What Do Learning Dynamics Reveal About Generalization in LLM Mathematical Reasoning? (ICML 2025 · similarity 0.88)
- Swing-by Dynamics in Concept Learning and Compositional Generalization (ICLR 2025 · similarity 0.87)
- Memorize Theorems, Not Instances: Probing SFT Generalization through Mathematical Reasoning (NeurIPS 2026 · similarity 0.87)
- Training on the Test Task Confounds Evaluation and Emergence (ICLR 2025 · similarity 0.87)
- Quantifying Generalization Complexity for Large Language Models (ICLR 2025 · similarity 0.87)
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.
