Small-model behavior does not predict large-model behavior
Some capabilities appear discontinuously as model scale crosses a threshold. Below the onset point, prompt optimization is wasted work: you are in the wrong regime.
Teams build a prototype with a small model because it's cheap, then watch it fail in messy ways: ignoring tools, breaking output format, losing the thread on multi-step constraints. Weeks go into prompt hacks and brittle routing logic to paper over it. Then someone swaps in a larger model, and half of those hacks turn out to have been unnecessary the whole time.
What happened is not mystical. Some capabilities appear discontinuously as model scale crosses a threshold. Below that threshold, small-model behavior is not predictive of large-model behavior. You are in the wrong regime, and optimizing there is wasted work.
Why loss can be smooth while capability is not
Training loss averages over a massive corpus and tends to improve smoothly with scale. But production capabilities are measured as task-level outcomes: did it call the right tool, did it follow the schema, did it complete all steps without a catastrophic error. Those metrics are nonlinear functions of token-level quality.
Slight improvements in average token prediction can push a model across a threshold where multi-step trajectories stop collapsing. Smooth microscopic improvements produce sharp macroscopic behavior changes. That is the pattern of a phase transition.
Wei et al. (2022) documented a concrete instance: on 3-digit addition and subtraction, GPT-3 stayed at near-zero accuracy for several orders of magnitude of training compute, then jumped sharply above random performance at roughly 13B parameters (~2×10²² training FLOPs). Nothing about the smaller models predicted that jump; the task simply looked unsolved until it was solved.
Same model, three tasks of increasing difficulty. The onset point — where the curve actually bends — moves right as the task gets harder.
Architecture can shift the threshold
You do not always need a bigger model. You sometimes need a simpler problem. RAG removes the need for parametric recall. Tools remove the need for in-context arithmetic. Constrained decoding removes format failure. Reranking removes multi-hop reasoning burden. Each of these moves the task into a regime where a smaller model becomes viable, by changing the problem, not by changing the model.
| Signal | What it means |
|---|---|
| Outputs are either correct or unusable, no middle ground | Bimodal distribution typical of near-threshold behavior |
| High prompt sensitivity at small scale that vanishes at larger | The model is not failing the task; it is failing the regime |
| Rapid jump in quality over a small size increase | You crossed the onset; behavior is now above the phase transition |
| One capability unlocking makes adjacent ones easier | Compositional capabilities have a shared threshold |
Evaluate stability, not a single run
Below threshold, lucky single-run passes are common. The model memorized a surface pattern or got a favorable sample. Evaluating stability means running multiple paraphrases, multiple seeds, and measuring pass@k rather than pass@1. What you are measuring is whether the capability is in the right regime, not whether it was lucky once.
That brittleness shows up the moment you push past a single test case:
I have been suspicious that smaller LLMs seem to have good performance, but are actually rather brittle. This paper bears that out. When given single math problems, small models do well. When given a 2nd problem that relates to the first, they fail to generalize. Big models can.
References: Wei et al., Emergent Abilities of Large Language Models (2022), Scaling Laws (Kaplan et al., 2020)