Unlocking reasoning in Large language models
summary
Survey of methods that reliably improve multi-step answers: chain-of-thought prompting and self-consistency boost accuracy on math and commonsense benchmarks (GSM8K, ARC) versus greedy decoding; offloading arithmetic to a Python interpreter (program-aided models) removes LLM arithmetic errors; plan-and-solve and least-to-most decomposition produce stepwise gains; model-steered distillation and STAR fine-tuning raise a 2.7B model’s math pass rates (examples show GPT-J from ~5.8% to ~10.7%); tool-enabled systems like REACT and Chameleon chain searches, detectors and calculators to reach correct decisions.
AI-written summary · about
chapters (21)
- 0:00intro
- 0:22preface
- 0:45about logesh
- 1:16agenda
- 2:14what is reasoning?
- 2:56how is reasoning measured (in the literature)?
- 5:32eliciting reasoning
- 7:47chain of thought prompting and self consistency
- 10:53program-aided language models
- 13:01plan-and-solve prompting
- 18:07star: self-taught reasoner bootstrapping reasoning with reasoning
- 19:33specializing smaller language models towards multi-step reasoning
- 20:48distilling step-by-step
- 22:34recursive and iterative prompting
- 23:16least-to-most prompting
- 25:11plan, eliminate, and track
- 27:23describe, explain, plan and select
- 30:06tool usage
- 30:43react: reason and act
- 33:02chameleon
- 37:52acknowledgement & further reading
related talks
24mThe Rise of Reasoning Models
32mScaling Test-time Inference Compute & Rise of Reasoning Models
25mWhy LLMs Still Fail: Reasoning, Robustness, and Safety
31mMaking the most of test-time compute in LLMs
27mExploring the Mechanisms of LLM Reasoning Capabilities
29mLLMs that think: Demystifying Reasoning Models
01h 22mBeyond Decoding: Meta-Generation Algorithms for Large Language Models
31mMachine Reasoning and System 2 Thinking