How can one training example improve reasoning across thousands of unseen problems?
This explores how a single training example can unlock reasoning that transfers to thousands of unseen problems — and why that's possible at all, which turns out to say more about what models already contain than about what training adds.
This explores how a single training example can unlock reasoning that generalizes far beyond it — and the corpus's answer is surprising: the training doesn't *teach* the reasoning, it *wakes it up*. The headline result is concrete. In reinforcement learning with verifiable rewards, one carefully chosen example lifts math accuracy from 36% to 73.6%, and test accuracy keeps climbing for 1,400 steps even after the model has perfectly memorized that single training problem Can a single training example unlock mathematical reasoning?. If the model were learning the example, performance would plateau once it mastered it. Instead it keeps generalizing — a signature that something already present is being activated, not installed.
The deeper claim that ties the corpus together is that base models already contain latent reasoning ability, and post-training merely selects it. Five independent techniques — RL steering, critique fine-tuning, decoding tweaks, sparse-autoencoder feature steering, and RLVR — all elicit reasoning that was already sitting in the base model's activations Do base models already contain hidden reasoning ability?. This reframes the whole puzzle: the bottleneck was never acquiring the capability, only triggering it. That's why a single example can ripple across thousands of unseen problems — it's a key, not a curriculum. A striking companion result shows you don't even need reinforcement learning: critique fine-tuning on *one problem*, using a teacher's critiques of right and wrong solutions, achieves comparable activation Can a single problem unlock reasoning through solution critique?. The sufficient signal is just exposure to the contrast between good and bad reasoning on a single case.
Here's where it gets genuinely strange. If activation rather than instruction is what matters, then the *content* of the training signal should matter less than we'd expect — and it does. Models trained on deliberately corrupted, semantically irrelevant reasoning traces perform about as well as those trained on correct ones, and sometimes generalize *better* out of distribution Do reasoning traces need to be semantically correct?. The traces seem to act as computational scaffolding that switches reasoning on, rather than as meaningful steps the model imitates. That fits a recurring theme in the corpus: chain-of-thought is often constrained imitation of reasoning's *form* rather than genuine inference What makes chain-of-thought reasoning actually work?, which is exactly why a minimal, even meaningless, nudge can be enough to flip the switch.
But the same lens that explains the one-example miracle also marks its ceiling — and this is the part worth sitting with. If training elicits rather than creates, it can only surface what's already latent. So the gains evaporate at the distribution's edge: chain-of-thought degrades predictably under shifts in task, length, or format, producing fluent but logically broken reasoning Does chain-of-thought reasoning actually generalize beyond training data?. Trace length tracks how close a problem sits to training schemas rather than how hard it actually is Does longer reasoning actually mean harder problems?. And on genuinely deep problems, models wander unsystematically and their success rate collapses exponentially Why do reasoning LLMs fail at deeper problem solving?. One example can unlock everything the base model already latently knows — but it cannot conjure reasoning the model never had.
If you want to pull this thread further, two doorways point in opposite directions. One asks whether we can extend this beyond verifiable math: verifier-free RL replaces answer-checking with the likelihood of a reference answer, carrying reasoning activation into general domains Can reasoning improvement work without answer verification?. The other is a caution — supervised fine-tuning can raise benchmark scores while quietly degrading the quality of reasoning steps by nearly 40%, meaning some 'improvements' are post-hoc rationalization rather than genuine inference Does supervised fine-tuning improve reasoning or just answers?. Together they sharpen the real question behind the one-example result: are we measuring reasoning, or just measuring whether the right latent capability got switched on?
Sources 10 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.
Critique Fine-Tuning achieves reasoning activation comparable to RLVR using only one problem and teacher-generated critiques of varied solutions, with no reinforcement learning. This demonstrates that exposure to correct versus incorrect reasoning on a specific problem is the sufficient activation signal.
Models trained on systematically irrelevant traces maintain solution accuracy and sometimes improve out-of-distribution generalization, suggesting traces function as computational scaffolding rather than meaningful reasoning steps.
CoT systems reproduce the form of reasoning through pattern matching rather than performing genuine logical inference. This explains why format effects dominate content, why structurally invalid prompts succeed, and why stronger reasoning models become less instruction-compliant.
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.
Controlled A* maze experiments show trace length correlates with difficulty only in-distribution but decouples entirely out-of-distribution. Trace length primarily reflects recall of training schemas, not adaptive computation.
Current reasoning models lack the three properties of systematic exploration: validity, effectiveness, and necessity. This causes success probability to drop exponentially with problem depth, making medium problems solvable but deep problems catastrophically harder.
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.
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.