A Taxonomy of Hidden Capability in Frontier Models
Abstract
Frontier models can do more than their public benchmarks advertise. Capabilities that are never demonstrated during evaluation still exist, and they can surface later in unpredictable ways. We propose a taxonomy of hidden capability — undeclared, latent, and emergent — and describe a protocol for probing each class before release. The taxonomy is imperfect, but the alternative is evaluating only what a model has already shown us.
Every capability test we run is answered by the model on our terms. We decide what to measure, and a model that can do things we never thought to measure passes untouched. This is not a conspiracy; it is a blind spot. Frontier models are broad, and our evaluation vocabulary is narrower than their competence.
This paper attempts to name the blind spot. We propose that hidden capability comes in three flavors, and that each needs a different kind of probing.
Approach
We define three classes:
- Undeclared capability: something the model can do at evaluation time that we have not benchmarked, but that shows up under direct inspection.
- Latent capability: something present in the training distribution but not expressed under our prompts, which can be elicited with the right context or fine-tuning.
- Emergent capability: something that appears only at scale, when compute or context length crosses a threshold.
For each class we built a probing protocol: direct inspection for undeclared capability, elicitation sweeps across context and prompting for latent capability, and controlled scale-up runs for emergent capability. We applied all three to Corus before its release.
Results
The taxonomy held up in practice. Direct probing revealed undeclared capability in a niche but consequential area: advanced statistical reasoning in a specialized domain, present in training but absent from every public benchmark we tracked. Elicitation sweeps surfaced latent capability in long-range code synthesis — the model could do it, but only with unusually detailed prompts. The scale-up runs were the most striking: several capabilities appeared only when context length passed roughly half a million tokens, which explained why they never showed up in shorter evaluations.
None of the findings was itself alarming. The point is that all three were real, all three were findable, and none of them appeared in the standard evaluation pipeline.
Discussion
If we cannot enumerate what a model can do, we should at least be systematic about what we have not enumerated. The taxonomy is a checkable claim — for each release, we can state which classes we probed, what we found, and what we did not probe.
The limits are real. Emergent capability is the hardest class to bound, and our scale-up runs are expensive and slow. The taxonomy will not catch everything. But it converts a silent gap into a named gap with a testing protocol attached — which is, for an evaluations team, the difference between a guess and a method.
References
- Ferrand, L., et al. (2025). What benchmark coverage actually covers. Curos Internal Report 2025-11.
- Kaplan, J., et al. (2020). Scaling laws for neural language models. arXiv preprint arXiv:2001.08361.
- Kwon, S., et al. (2024). Eliciting latent skills through context engineering. ICML 2024.
- Pérez, E., et al. (2022). Discovering language model behaviors with model-written evaluations. arXiv preprint arXiv:2212.09251.
- Sato, H., & Herrera, T. (2025). Long-context probes and what they surfaced. Curos Internal Report 2025-09.
- Wei, J., et al. (2022). Emergent abilities of large language models. Transactions on Machine Learning Research, 2022(6).