Week 13: Reasoning and chain-of-thought

DATA 443/643 · Week 13

Module 4 — Reasoning and chain-of-thought

Notes

• 13.1 Chain-of-Thought(slides)
• 13.2 Self-Consistency and Tree-of-Thoughts(slides)
• 13.3 Faithfulness of Explanations(slides)

Labs

• Lab 13: Direct Answers, Traces, and a Majority Vote(PDF)

Videos

• Karpathy: Intro to Large Language Models (reasoning / System 2, 35:00–38:02)
• Karpathy: Deep Dive into LLMs like ChatGPT (tokens as compute, 1:46:56–2:01:11; SC is k of those traces)

Additional readings

• Stanford CS224N 2026 L12: Reasoning 1 (CoT, self-consistency, unfaithful traces)
• Stanford CS224N 2026 L13: Reasoning 2 (process vs outcome; test-time compute)
• Wang et al., Self-consistency improves chain of thought reasoning

Discussion

• Wang et al. cite +17.9 pp on GSM8K from majority vote. Why can that lift still leave traces unfaithful?
• Temperature 0, k=3: what do the three traces look like, and why is that not self-consistency?
• 52 in the last step, 42 in the box, gold 42: accuracy? faithfulness? lucky win?
• Why is Lightman’s process reward a different claim from scoring only the boxed answer?