NOTE-216: Chain-of-Thought Prompting

Contribution

Two lines of work existed and each had a fatal cost. Rationale-augmented training produced models that showed their working, but required a large corpus of hand-written rationales. Few-shot prompting needed no training at all, but worked poorly on anything requiring reasoning and — the damning part — did not improve much with scale. This combines them: put the rationales in the eight exemplars. The annotation cost collapses to eight examples and the scaling curve, previously flat, turns sharply upward.

Key insight

The capability was already in the model and the prompt was the bottleneck. The paper’s own summary is that “standard prompting only provides a lower bound on the capabilities of large language models” — which is a claim about measurement, not about training, and it is the reason a method this simple produced a step change. Nothing about the model was altered.

Assumptions

• Sufficiently large models. The effect is an emergent ability of scale in the authors’ framing, appearing around 100B parameters. Below that it is absent or harmful.
• Few-shot setting. Annotation cost is negligible for eight exemplars and the paper says explicitly that it would be prohibitive for fine-tuning.
• Tasks expressible in language. Arithmetic, commonsense and symbolic reasoning; the claim of general applicability is “at least in principle, to any task that humans can solve via language”, which is a hope rather than a result.
• Three model families — LaMDA, GPT-3, PaLM — with PaLM 540B carrying the headline numbers.

Key results

• GSM8K — PaLM 540B with eight chain-of-thought exemplars reaches state of the art, beating a fine-tuned GPT-3 with a verifier. Performance more than doubles for the largest GPT and PaLM models. Holds when: ≥100B parameters.
• Emergence — no positive effect for small models on the arithmetic benchmarks; gains appear only at ~100B. Small models generate fluent but illogical chains and do worse than standard prompting.
• Gain tracks difficulty — largest on the hardest dataset, negative or negligible on SingleOp, the single-step subset of MAWPS.
• Ablation: equation only — prompting for just the equation does not help on GSM8K, though it does help on one- and two-step problems where the equation follows directly from the question.
• Ablation: variable compute only — emitting a sequence of dots as long as the needed equation performs at baseline. Extra tokens alone buy nothing.
• Ablation: chain of thought after the answer — performs at baseline, so the chain is not merely activating pretraining knowledge; the answer depends on it.
• Robustness — three independent annotators all beat standard prompting; exemplars randomly sampled from the GSM8K training set do as well as hand-written ones; robust to exemplar order and count.
• Commonsense — gains across five datasets, with StrategyQA at 75.6% vs a 69.4% prior best and sports understanding beating an unaided sports enthusiast (95.4% vs 84%). Minimal on CSQA, which the paper says.
• Symbolic — chain of thought lets the model generalize out of distribution to longer sequences than the exemplars showed.
• Error analysis — of 50 correct answers from LaMDA 137B, all but two had correct chains. Of 50 wrong answers, 46% were nearly correct and 54% had major errors. Scaling 62B → 540B fixes many one-step-missing and semantic errors.

Claims

id	claim	strength	support
C1	Worked steps in exemplars substantially improve multi-step reasoning	strong	three model families, three task categories, large margins, robust to annotator and exemplar choice
C2	The benefit is not from emitting the equation	strong	direct ablation
C3	The benefit is not from spending more tokens	strong	the dots ablation — the cleanest negative result in the paper
C4	The model’s answer depends causally on the chain	moderate	the after-the-answer ablation supports it; “depends causally” is a reading of an accuracy comparison, not a mechanistic result
C5	The ability is emergent at ~100B parameters	moderate	measured on the model sizes available in 2022, and “emergent” is a description of those curves, not a mechanism
C6	Whether the model is “reasoning”	not claimed	the authors explicitly decline this and leave it open

Method

Replace each few-shot exemplar’s ⟨input, output⟩ with ⟨input, chain of thought, output⟩, where the chain is a series of intermediate natural-language steps. Eight exemplars. Prompt an off-the-shelf model. That is the entire method.

Concepts

• Chain of thought — a coherent series of intermediate natural-language reasoning steps leading to the answer. The authors prefer this to “solution” or “explanation” because those conventionally come after the answer, and the ordering is the point.
• Emergent ability of scale — a capability absent in smaller models that appears past a threshold.

Connections

Builds directly on few-shot prompting from GPT-3 (LIT-035), declared as extends. Positioned against rationale-augmented fine-tuning, whose annotation cost it avoids, and against neuro-symbolic methods that use formal languages instead of natural ones.

The relation to LIT-464 is worth stating carefully, because a fast reading makes them contradict. That paper proves a one-layer decoder with chain of thought can express a composition a constant-depth one cannot, and describes chain of thought as providing extra computation space. This paper’s dots ablation shows that extra tokens carrying no information buy nothing. Those are consistent: the theoretical benefit comes from writing intermediate results into the context and reading them back, and a row of dots holds no results. One says the room exists; the other says the room only helps if you put something in it.

Bearing on the record

• It fills the trunk named in the previous contribution. Before this the record held no chain-of-thought practice in 278, and SOTA-127 — filter such traces out of tiny models’ training data — was the entire CoT presence. That practice now has the paper that explains why small models produce bad chains sitting next to it.

• It produces SOTA-279.

• C5 sits in a live tension the record already holds. SOTA-037 says in-context learning emerges at scale; SOTA-200 says to check whether an emergent capability is a metric artefact before believing it. And this paper’s emergence claim is measured on exact-match accuracy over multi-step answers — which is, almost word for word, the case BIG-bench (LIT-077) singles out: tasks showing breakthrough behaviour “often involve multiple steps or components, or brittle metrics”. A chain-of-thought answer is scored right only when every step lands. That does not make the effect unreal — the gains at 540B are enormous and survive three ablations — but the shape of the curve is the shape SOTA-200 says to distrust, and nobody has re-run this against a continuous score. The record should not resolve that from here and should not pretend it is not there.

  Schaeffer et al.’s “Are Emergent Abilities of Large Language Models a Mirage?” (ARXIV-2304.15004) is the paper that argues this case directly, and the record does not hold it. Named here as the obvious next filing.

• It is the 2022 statement of SOTA-278. “Standard prompting only provides a lower bound on capabilities” is that practice’s claim, demonstrated rather than argued, three years before the papers the practice is sourced to.

• Much of what the record already carries assumes it. SOTA-130 skips reasoning SFT and runs RL with verifiable rewards; the reasoning models evaluated in LIT-466 are defined by an extended thinking trace. Neither had an antecedent in the record for the thing being extended.

Limitations

The authors list four and they are unusually candid.

• Whether the model is reasoning is left open, explicitly.
• Annotation cost is trivial at eight exemplars and prohibitive for fine-tuning, so the method’s cheapness is specific to the few-shot setting.
• No guarantee of correct reasoning paths. Chains can be wrong and still reach the right answer, and vice versa — their own error analysis found two of fifty correct answers arrived through incorrect chains.
• The scale requirement makes it costly to serve.

And two the record should add:

• The emergence threshold is a 2022 fact about 2022 models. Whether ~100B is still the boundary for models trained on more data with better recipes is not something this can answer, and small-model reasoning distillation has moved since.
• No cost accounting. Chains of thought are tokens, and on the easiest problems the paper reports negative gains — so the practice has a regime where it loses on both axes, which the paper notes without quantifying.

Open questions

• Is the ~100B threshold a property of scale or of the training data distribution of that generation?
• How much of the emergence is the metric? Nobody re-ran this against a continuous score.
• The dots ablation is the sharpest negative result here and has never, as far as this reading knows, been repeated against modern models that are explicitly trained to use thinking tokens — where the answer could easily be different.