Can abstention behavior transfer from small models to frontier models?

This explores whether abstention — a model declining to answer when uncertain rather than confidently hallucinating — is something small models can teach or hand up to frontier ones. The corpus doesn't study that exact 'transfer' path, but several notes together suggest the question may be framed backwards: abstention looks less like a size-dependent capability that flows upward and more like a learnable behavior that standard training leaves underdeveloped at every scale. The most direct evidence is that small models trained with uncertainty-aware objectives and an explicit abstention option already match models ten times their size on conversation forecasting Can models learn to abstain when uncertain about predictions?. The lesson there isn't that big models lack the latent ability — it's that calibration 'exists but remains undertrained.' So the interesting transfer isn't small→frontier weights; it's the training recipe.

What makes abstention learnable at all is reward design. TruthRL shows that a binary right/wrong reward actively punishes honesty, because saying 'I don't know' scores the same as being wrong — so models learn to guess. Splitting the reward three ways (correct, hallucination, abstention-in-between) makes restraint something the model can optimize toward, cutting hallucinations by ~29% while preserving accuracy Can three-way rewards fix the accuracy versus abstention problem?. This reframes your question: the thing that 'transfers' is the reward structure, not behavior copied from a smaller network. A frontier model given the same ternary signal would learn to abstain regardless of where a small model stands.

The deeper reason size isn't the lever comes from work showing that frontier-grade reasoning is a property of the post-training pipeline, not the parameter count — a 3B model with curriculum SFT and multi-domain RL reaches scores that rival far larger systems on verifiable tasks Can small models match frontier reasoning without massive scale?. If reasoning itself is recipe-bound rather than scale-bound, abstention (a close cousin — knowing the boundary of what you can verify) almost certainly is too. There's even a mechanistic hint: RL tends to rewrite only a small, structured 5–30% subnetwork rather than the whole model Does reinforcement learning update only a small fraction of parameters?, suggesting behaviors like abstention may live in compact, targetable regions — which is friendlier to instilling-via-training than to mysterious emergence-at-scale.

One caution the corpus raises against expecting a model to bootstrap abstention on its own: pure self-improvement stalls because of the generation–verification gap and reward hacking, and reliable gains always smuggle in an external anchor — a judge, a tool, a human correction Can models reliably improve themselves without external feedback?. Abstention is fundamentally about recognizing the limit of your own verification, which is exactly where self-generated signal is weakest. That argues the honest calibration signal has to come from outside the model — and notably, suppressing wrong answers (rather than only rewarding right ones) is itself a strong training lever that preserves diversity Does negative reinforcement alone outperform full reinforcement learning?, which is structurally what abstention does: prune confident-but-wrong outputs.

The thing you didn't know you wanted to know: the win here isn't a small model donating its caution to a big one — it's that a small calibrated model is a working proof that the reward recipe makes abstention learnable, and that recipe is what scales. The frontier model's advantage was never that it abstains better; it's that, left in standard binary-reward training, it was taught not to.