'Reasoning' in the LLM context is a contested word. Models can execute multi-step chains of thought that look like reasoning and often produce correct answers. Whether that constitutes reasoning in a philosophical sense is a separate question from whether it works in production.
- 01Distinguish process reasoning from outcome reasoning.
- 02State what chain-of-thought empirically achieves.
- 03Explain the interpretability problem with reasoning traces.
The empirical claim
Prompting a model to 'think step by step' before answering (chain-of-thought, Wei et al. 2022) improves accuracy on math, logic, and multi-step reasoning tasks by 10-40 percentage points. This is a robust, replicated result.
The interpretability problem
The trace may or may not reflect the actual computation. Recent work shows models sometimes produce correct answers with incorrect traces, and vice versa. Traces are output, not internal state.
Reasoning models
o1 and successors are trained via RL to produce long internal reasoning traces before answering. Capability gains on hard tasks are real; the resemblance to human reasoning remains a metaphor rather than a claim.
- Chain-of-thought empirically works; whether it 'reasons' is a separate question.
- Traces are output, not necessarily faithful process.