AI Paper Atlas

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

ICML 2026 · Large language models › Reasoning and chain of thought · approach: Efficient and adaptive thinking

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

Measured effects reported in the paper

What was changedWhereMeasuredEffect
use depth-increasing curriculum instead of direct GRPOQwen2.5-1.5B-Instruct GRPO · Countdownmean pass@K (%): 36.0 -> 40.5Improved
use hint-decreasing curriculum instead of direct GRPOQwen2.5-1.5B-Instruct GRPO · Countdownmean pass@K (%): 36.0 -> 42.6Improved
use depth-increasing curriculum instead of direct GRPOQwen2.5-1.5B-Instruct GRPO · MATHmean pass@K (%): 36.3 -> 38.2Improved
use hint-decreasing curriculum instead of direct GRPOQwen2.5-1.5B-Instruct GRPO · MATHmean pass@K (%): 36.3 -> 33.8Hurt
use depth-increasing curriculum instead of direct GRPOQwen2.5-1.5B-Instruct GRPO · Blocksworldmean pass@K (%): 8.3 -> 8.4No clear effect
use hint-decreasing curriculum instead of direct GRPOQwen2.5-1.5B-Instruct GRPO · Blocksworldmean pass@K (%): 8.3 -> 9.2 (pass@256 19.7 -> 12.4)Mixed
add curriculum (either type) to RL fine-tuning on paritytransformer RL fine-tuning · Parity d=8, k=3sample size to converge: n = 1600076 -> n = 76Improved
add curriculum oracle to test-time scaling on paritytest-time scaling (LTAR vs Alg. 4) · Parity d=8, k=3reward-oracle queries to converge: n = 251410 -> n = 1410Improved
add curriculum to RL fine-tuning on parity (attention structure)transformer RL fine-tuning · Parity d=8, k=3attention on correct parent-child linksImproved

Limitations

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