Towards Reasoning in Large Language Models: A Survey

Proposes an overview of what is reasoning (deductive - based on premises, inductive - conclusion based on observations, abductive - conclusion based on explanation of observations/requires working knowledge of observation, formal - structured, and informal - unstructured/open ended), LLM reasoning methods, evaluations and analysis, and some future direction for LLM reasoning research. Fully supervised fine-tuning: You could fine-tune on a dataset containing explicit reasoning. LLM reasoning methods by prompting (ICL - in-context learning): chain-of-though (CoT), zero-shot CoT (prefix “think step-by-step”), intermediate scratchpads; rationale exploration and verification (consistency checks, add scratchpad - tree-of-thought came later) augmented CoT (give rationale in each example); problem decomposition using divide and conquer techniques (least-to-most, decomposition, and successive prompting); others have cascade of LLMs for breaking query, handling sub-tasks, verification, and union. End task performance can be measured through arithmetic reasoning (MathQA, GSM8K, etc.), common sense reasoning - everyday knowledge (CSQA, StrategyQA), symbolic reasoning - form rules by symbols (Coin Flip), task-based (BIG-bench), generalization (SCAN); reasoning can be analyzed through step-by-step metrics (ROSCOE), correctness of conclusions based on premise (FOLIO). Reasoning could be an emergent ability of LLMs, CoT could elicit it, LLMs unskilled at complex reasoning. Some emergent trends in reasoning could just be because of training data and memorization (they still hallucinate). From UIUC.

Links: PapersWithCode, GitHub