Mapping how language models learned to think

What it does This repository is a curated bibliography of papers and resources on reasoning in large language and multimodal models. It organizes surveys, critical analyses, and techniques into a single reading list that runs from early chain-of-thought work through recent reasoning architectures like OpenAI o1 and DeepSeek-R1. If you need a map of the field, this is the map.

The interesting bit The list does not just collect successes; it catalogs failures and open questions. Papers arguing that LLMs are not yet genuine reasoners, that chain-of-thought can be unfaithful, and that models cannot self-correct share the same sections as the latest prompting breakthroughs. That balance turns a simple index into a useful reality check.

Key highlights

• Chronological coverage from 2022 emergent-abilities work to 2025 reasoning-model evaluations
• Dedicated sections for multimodal reasoning and squeezing reasoning into smaller language models
• Includes skeptical analyses such as “Large Language Models Cannot Self-Correct Reasoning Yet” and “Language Models Don’t Always Say What They Think”
• Points to a companion benchmark repo, LLMSymbolicReasoningBench, for testing symbolic reasoning
• Over 3,600 GitHub stars

Caveats

• It is a reading list, not a framework; expect citations, not code or reproducible experiments
• The README is extensive and was truncated in the provided sources, so the full depth of later technique sections is unclear

Verdict Reach for this if you are writing a literature review, designing a reasoning benchmark, or just tired of missing citations. Skip it if you are hunting for a drop-in reasoning library.