Tag
This paper presents a taxonomy of LLM reasoning strategies along two orthogonal axes: fast vs. slow thinking and internal vs. external knowledge, and surveys recent adaptive reasoning methods.
This paper investigates whether reinforcement learning post-training can compose primitive skills into higher-level compositional strategies, using a fully observable rewrite-grammar environment. The authors find that RL reorganizes primitive competence through phased compositional mechanisms, while rejection fine-tuning plateaus due to producing many invalid shortcut-like rewrites.
Proposes Strategy-Induct, a framework that induces task-level instructions from only example questions without requiring labeled answers, by first generating explicit reasoning strategies for each question and then using question-strategy pairs to induce an instruction. The method outperforms existing approaches in question-only settings across multiple tasks and model scales.