Model merging preserves emergent flaws but cannot create new ones

The key constraint here is directional: merging cannot create capabilities that don't exist in either parent model, only preserve or degrade what's already there. This is a ceiling, not a floor, and it reframes merging from a general assembly technique to a conservative one.

This finding sits in direct tension with the SERA work from the same day on rank-allocation for model merging. SERA optimizes how to distribute capacity across merged task vectors to preserve fidelity, but this paper establishes that no allocation strategy can synthesize novel emergent properties. The two papers together define the honest scope of merging: it's a cost-effective way to combine existing capabilities without retraining, but practitioners cannot expect it to bootstrap new behaviors. The constraint also connects to the representational bias detection work from today, which focuses on what's already latent in data; both papers are about discovering and working with what exists rather than creating something new.

If researchers publish follow-up work within the next six months showing that merging can induce emergent behaviors through specific architectural modifications (e.g., adapter layers, routing mechanisms), that would contradict this finding. Otherwise, if the constraint holds across diverse model families and scales through 2027, expect the field to shift focus from assembly-based scaling toward joint fine-tuning or retrieval-augmented approaches for capability expansion.