Model-capacity scaling laws for CCA benefits

Characterize how the benefit of the Context Compilation Architecture scales with the capacity of the underlying language model, particularly across models with smaller activation sizes, to establish the model-capacity boundary at which the harness benefit saturates.

Background

The experiments show that CCA substantially improves pass rates on Kimi K2.5, GLM-5, and DeepSeek-V3.2, but produces only a small and statistically insignificant improvement on Qwen3-Next-80B, which has 3B active parameters. The authors interpret this result as evidence that the harness benefit may be bounded by a model’s ability to integrate the structured intermediate representation, injected code-execution results, and original context.

The paper does not determine the scaling relationship between model capacity and CCA’s effectiveness. In particular, it leaves unresolved whether the reduced benefit on the smallest-activation model reflects a general capacity threshold, how the F2 checklist-injection and F6 correction-loop components interact across model sizes, and how CCA performance changes under broader model-capacity scaling.

References

Model-capacity scaling laws are left to future work (§\ref{sec:limitations}).

— Compile, Don't Memorize: A Context Compilation Architecture (CCA) for In-Context Learning  (2609.00759 - Qi et al., 1 Sep 2026) in Section 4.3, subsection “Answering RQ3: Where the Compilation Harness Holds”; Section 5, subsection “Moderator 2 (model capacity)”