Can small demonstration sets unlock general reasoning without large question data?
This explores whether a handful of worked examples (or even one) can switch on broad reasoning ability, instead of needing huge labeled training sets — and what that implies about where reasoning actually lives.
This explores whether a handful of worked examples (or even a single one) can switch on broad reasoning ability without the usual mountain of question-answer data. The short version from the corpus: yes, surprisingly often — but only because the reasoning was already there, latent, waiting to be elicited rather than taught. The most striking data point is that one carefully chosen training example in an RLVR setup can lift math performance from 36% to 73.6%, and keep improving test accuracy for 1,400 steps after the model has already nailed the training set perfectly Can a single training example unlock mathematical reasoning?. That doesn't look like learning a skill; it looks like flipping a switch.
The reason small sets work is the punchline of a broader thread: base models already contain the reasoning, and post-training mostly *selects* it rather than creating it. One synthesis found five completely different mechanisms — RL steering, critique fine-tuning, decoding tweaks, sparse-feature steering, and RLVR — all eliciting the same pre-existing capability, concluding the bottleneck is elicitation, not acquisition Do base models already contain hidden reasoning ability?. You can even skip training entirely: four modular 'cognitive tools' implemented as sandboxed prompts pushed GPT-4.1 on AIME from 26.7% to 43.3% with no reinforcement learning at all Can modular cognitive tools unlock reasoning without training?. If a prompt scaffold alone unlocks that much, the capability clearly wasn't being installed by data.
Where does the latent ability come from, if not from question sets? The corpus points back to pretraining itself. An analysis of five million pretraining documents found that reasoning draws on broad, transferable *procedural* knowledge spread across many sources — the 'how to do this kind of step' — which is the opposite of factual recall, which depends on narrowly memorizing specific documents Does procedural knowledge drive reasoning more than factual retrieval?. That's why a tiny demonstration set can generalize: it's activating a procedure the model already absorbed, not teaching it from scratch. Quiet-STaR pushes this further, learning to generate rationales at every token while reading arbitrary internet text, so general reasoning emerges as a side effect of better language modeling rather than from any task-specific dataset Can models learn reasoning from predicting any text?.
The sharp caveat is that 'unlock' and 'fake it' look identical on a benchmark, so small-data methods deserve suspicion about *what* they activate. Supervised fine-tuning can raise final-answer accuracy while actually degrading the quality of the reasoning steps by ~39%, producing correct answers via post-hoc rationalization that standard metrics never catch Does supervised fine-tuning improve reasoning or just answers?. And chain-of-thought trained on narrow data degrades predictably the moment the task, length, or format shifts — fluent reasoning form without valid underlying logic Does chain-of-thought reasoning actually generalize beyond training data?. So small demonstration sets can genuinely elicit general reasoning, but they can just as easily elicit a convincing *imitation* of it; the difference only shows up off-distribution.
If you want to follow the thread further, two adjacent ideas loosen the data requirement from other angles: verifier-free RL replaces answer-checking with the likelihood the model assigns to a reference answer, extending reasoning RL into general domains without rule-based verifiers or large labeled sets Can reasoning improvement work without answer verification?, and energy-based transformers reach System-2-style deliberation from unsupervised learning alone, generalizing out-of-distribution without any domain-specific scaffolding Can energy minimization unlock reasoning without domain-specific training?. The common lesson across all of them: the expensive thing was never the question data — it was the pretraining that planted the capability the small set wakes up.
Sources 9 notes
A single example in RLVR boosts math performance from 36% to 73.6% and enables test accuracy to improve for 1,400 steps after training accuracy reaches 100%, revealing that minimal activation signals unlock latent reasoning capability.
Five independent mechanisms—RL steering, critique fine-tuning, decoding changes, SAE feature steering, and RLVR—all elicit reasoning already present in base model activations. Post-training selects rather than creates reasoning; the bottleneck is elicitation, not capability acquisition.
Four cognitive tools implemented as sandboxed LLM calls improved GPT-4.1 on AIME2024 from 26.7% to 43.3% without any RL training. Modularity enforces operation isolation that pure prompting cannot guarantee, eliciting pre-existing reasoning capability.
Analysis of 5 million pretraining documents shows reasoning relies on broad, transferable procedural knowledge from diverse sources, unlike factual recall which depends on narrow, document-specific memorization of target facts.
Quiet-STaR trains language models to generate rationales at every token position during pretraining on arbitrary internet text, enabling general reasoning without task-specific datasets. Rationale quality is judged by predictive accuracy rather than labeled correctness, allowing reasoning competence to emerge as a side effect of improved language modeling.
Supervised fine-tuning improves final-answer accuracy on benchmarks but cuts Information Gain by 38.9 percent, meaning models generate correct answers through post-hoc rationalization rather than genuine inferential steps. Standard metrics miss this degradation because they only measure final correctness.
DataAlchemy experiments show CoT fails systematically under distributional shifts in task, length, and format. Models produce fluent but logically inconsistent reasoning — imitating reasoning form without valid underlying logic.
VeriFree bypasses answer verification entirely by using the conditional probability of reference answers given generated reasoning traces as both reward signal and training weight. This approach matches or surpasses verifier-based methods on MMLU-Pro, GPQA, and SuperGPQA without rule-based or model-based verifiers.
Energy-Based Transformers assign energy values to input-prediction pairs and use gradient descent minimization for inference, yielding 35% higher training scaling rates and 29% more inference-compute gains than Transformer++, while generalizing better on out-of-distribution data without domain-specific scaffolding.