Model techniques map
Techniques

specific method · not yet filed

Non-uniform embedding-slot allocation across N-gram orders

source
1
model
1
labs adopt it
0
strongest
evaluated

How sources treat it

One count per evidence span, weakest treatment to strongest.

evaluated 1

Documented in

Evidence

1 span quoted from the sources, strongest treatment first.

including but not limited to token normalization for vocabulary compression ..., non-uniform allocation across N-gram orders, and frequency-based partitioning of embedding slots. Despite these efforts, we observed no consistent performance gains

evaluatedmodel architecturein Qwen3.8-NextQwen