Research map / Large language models / Reasoning and chain of thought
Provable Sample Efficiency of Curriculum Post-Training for Transformer Reasoning
Dake Bu, Wei Huang, Andi Han, Atsushi Nitanda, Hau-San Wong, Qingfu Zhang, Taiji Suzuki
Read the paperCheck your idea against 137,217 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Develops an abstract theoretical framework showing sufficient conditions under which curriculum post-training of transformers yields exponential improvements in sample efficiency for reasoning tasks.
Problem
Curriculum techniques in LLM post-training are empirically reported to beat non-curriculum training on reasoning, but a principled account of when and why they help is incomplete.
Approach
An abstract theoretical framework identifying sufficient conditions under which curriculum post-training yields exponential improvement in sample complexity for transformer reasoning, compared with non-curriculum approaches.
Key results
- Identifies sufficient conditions under which curriculum post-training yields exponential improvement in sample efficiency for transformer reasoning (per the abstract).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use depth-increasing curriculum instead of direct GRPO | Qwen2.5-1.5B-Instruct GRPO · Countdown | mean pass@K (%): 36.0 -> 40.5 | Improved |
| use hint-decreasing curriculum instead of direct GRPO | Qwen2.5-1.5B-Instruct GRPO · Countdown | mean pass@K (%): 36.0 -> 42.6 | Improved |
| use depth-increasing curriculum instead of direct GRPO | Qwen2.5-1.5B-Instruct GRPO · MATH | mean pass@K (%): 36.3 -> 38.2 | Improved |
| use hint-decreasing curriculum instead of direct GRPO | Qwen2.5-1.5B-Instruct GRPO · MATH | mean pass@K (%): 36.3 -> 33.8 | Hurt |
| use depth-increasing curriculum instead of direct GRPO | Qwen2.5-1.5B-Instruct GRPO · Blocksworld | mean pass@K (%): 8.3 -> 8.4 | No clear effect |
| use hint-decreasing curriculum instead of direct GRPO | Qwen2.5-1.5B-Instruct GRPO · Blocksworld | mean pass@K (%): 8.3 -> 9.2 (pass@256 19.7 -> 12.4) | Mixed |
| add curriculum (either type) to RL fine-tuning on parity | transformer RL fine-tuning · Parity d=8, k=3 | sample size to converge: n = 1600076 -> n = 76 | Improved |
| add curriculum oracle to test-time scaling on parity | test-time scaling (LTAR vs Alg. 4) · Parity d=8, k=3 | reward-oracle queries to converge: n = 251410 -> n = 1410 | Improved |
| add curriculum to RL fine-tuning on parity (attention structure) | transformer RL fine-tuning · Parity d=8, k=3 | attention on correct parent-child links | Improved |
Limitations
- Abstract-level framework; conditions may not hold in practice
- Full text unavailable; limitations not verified
Closest papers by meaning
- CurES: From Gradient Analysis to Efficient Curriculum Learning for Reasoning LLMs (ICLR 2026 · similarity 0.89)
- Curriculum Reinforcement Learning from Easy to Hard Tasks Improves LLM Reasoning (ICLR 2026 · similarity 0.88)
- h1: Bootstrapping LLMs to Reason over Longer Horizons via Reinforcement Learning (ICML 2026 · similarity 0.88)
- Teaching Models to Teach Themselves: Reasoning at the Edge of Learnability (ICML 2026 · similarity 0.87)
- Overthinking Reduction with Decoupled Rewards and Curriculum Data Scheduling (ICLR 2026 · similarity 0.87)
- Reshaping Reasoning in LLMs: A Theoretical Analysis of RL Training Dynamics through Pattern Selection (ICLR 2026 · 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.
