On October 24, 2026, a paper titled “Joint or Per-Option? Compact Encoders for Closed-Set Decisions from 17 Million to 1 Billion Parameters” was first published on the arXiv platform (arXiv:2610.30970v1). The study focuses on closed-set decision tasks at the end of language model pipelines—such as tool selection—and investigates how to build efficient, robust decision encoders with minimal parameter counts. It centers on comparing two input-processing paradigms: joint scoring of all options and per-option scoring.
In a three-random-seed experiment with 32 million parameters, on a decision‑making suite composed of 37 benchmarks, the item‑by‑item scoring method outperformed joint scoring by 5.8 points (when optimizer steps were aligned) and by 2.8 points (when training computational costs were approximately aligned), using “chance‑corrected skill” as the metric. The results from two external benchmarks followed the same trend, but the differences on tasks not seen in those benchmarks did not reach statistical significance. The study further notes that this advantage depends on the reranker’s initialization strategy: its input format (query–candidate pair) is naturally well‑suited to item‑by‑item reading; if only the raw encoder output is used, the two types of readers perform identically.
While the逐选项方案 offers superior accuracy, it requires 5 to 13 times more encoder forward passes, causing the median latency to nearly double and the average latency to increase by more than fourfold. In contrast, among the six released joint encoders—ranging in size from 17 million to 1 billion parameters—the 68-million- and 150-million-parameter models achieved accuracy scores that were 4.9 and 5.9 percentage points higher, respectively, than the 435-million-parameter DeBERTa‑v3 zero-shot classifier across 209 System1‑Arena tasks with overlapping audits; on 145 tasks unseen during training, their performance was at least on par. Meanwhile, the smallest 17-million-parameter joint model exhibited a median response latency of just 14 milliseconds on a 16‑thread processor, far lower than DeBERTa‑v3’s 728 milliseconds.
Furthermore, the joint encoder is sensitive to the order of options: reordering the options alters 12%–23% of its prediction results, whereas scoring each option individually remains unaffected. On the private test set of the public leaderboard—primarily consisting of in-distribution tasks—the joint encoder with no more than 68 million parameters outperforms others within its size range. All conclusions are based on the experimental setup and benchmarks reported in the paper and do not extend to parameter scales that were not tested or task distributions that were not disclosed.