Ladders of Thought: A Self-Evolving Curriculum of Progressively Simplified Reasoning Traces

ICLR 2026 Conference Submission20266 Authors

Keywords: curriculum learning, reasoning, distillation, language models, progressive rewrites, difficulty estimation

TL;DR: We improve reasoning in small LLMs by progressively rewriting questions into easier forms, labeling difficulty via reasoning steps, and training with a self-evolving curriculum that outperforms strong distillation baselines.

Abstract: Large language models (LLMs) excel at reasoning when scaled to hundreds of billions of parameters, but small- and mid-scale models remain brittle reasoners even with knowledge distillation (KD). We present Ladders-of-Thought (LoT), a framework that improves reasoning by combining progressive question rewrites with a self-evolving curriculum. LoT automatically generates semantically faithful but easier variants of reasoning problems, organizes them into difficulty buckets using step-based measures, and employs a self-evolving bandit scheduler to allocate training adaptively. Evaluated on two reasoning domains, math and multi-hop reasoning, across OPT-1.3B/2.7B and Pythia-1.4B/2.8B, LoT consistently improves over KD. It delivers large gains on arithmetic tasks (e.g., +32 percentage points on AddSub, +25pp on SVAMP), +2–8pp improvements on in-domain test splits, and strong though dataset-dependent benefits on multi-hop reasoning (e.g., +16pp on QASC, +25pp on StrategyQA). LoT also converges faster than staged curricula, highlighting the value of adaptive progression. These results show that progressive rewrites coupled with adaptive curricula provide a simple yet effective recipe for strengthening reasoning in smaller LLMs.

Primary Area: foundation or frontier models, including LLMs

Submission Number: 20266