A team of LinkedIn researchers describes replacing generated YES/NO lists with semantic scoring for profile qualification labels. Instead of generating a list of decisions, the system scores each candidate label using a model pass that avoids generating a full answer. The paper says its approach changed at most 0.01% of decisions when candidate order changed, and matched packed generation’s agreement with reference labels using about 4.7 times fewer training pairs.
A batching method called Multiple-Item Scoring raised throughput by 1.24 times offline and 1.9 times in model serving, under a one-second 99th-percentile latency target. The paper says the offline system uses a reusable Ray Data API and serving uses support in SGLang.
Why it matters
For teams labeling LinkedIn profiles, the reported gains could mean processing more candidates within the same serving-time target; the paper does not say whether the approach is available to other organizations.
