Unverified50% confidenceBenchmarkExact time
In the German-English translation task, a word-by-word autoregressive model as a control scored 2.6 points higher in quality than NCM, which caught up through minimum Bayes risk decoding but consumed about twice the compute.
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
10/31/2026
First Seen
Valid until: 1/29/2027
Sources
Related Entities
Related Claims
UnverifiedThe largest NCM model the authors trained is under 1 billion parameters, and whether it holds at 27B or even larger scales has not been verified.66% similarUnverifiedBLEU measures translation quality by counting the overlap rate of n-grams (usually 1- to 4-grams) between the model output and the reference translation, and imposes a length penalty on overly short outputs; its value ranges from 0 to 100, but it is not a linear scale.65% similarUnverifiedThe generator LLM, probe model, and prompt verbosity together contribute less than 10% of the variance.62% similarUnverifiedThe 406M model's peak inference memory is less than half that of the 1.2B system, its parameter count is approximately one-third, and the two are comparable in performance.62% similarUnverifiedEven well-prompted, schema-constrained models with 8 billion to 14 billion parameters still have an error rate of a few percent on real-world inputs—this is a rate to design around, not a bug to eliminate.62% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1121437API
curl https://kongchang.com/api/v1/knowledge/claims/1121437MCP
get_claim(id=1121437)