cs.AINov 12, 2026

How many LLM-labelled rows does a decision encoder need, and how wrong can the labels be?

Authors: Mohammed Nayeem Khan, Aliyas Ahmad Misri

Organizations: Intel Corporation

Abstract

A decision encoder such as Laya reads a question with named options (for example, which department should handle a ticket) and a text, and picks an option. It reads both in one forward pass and returns a probability for every option, without writing any text. We ask how many LLM-written training rows such an encoder needs and how wrong those labels may be. We ran 19 fine-tunes of Laya (ModernBERT-large) on nested subsets of a 2,818-row pool and scored them on 1,065 held-out items from the same generator, so every test item is in-distribution and no out-of-family test was run. Accuracy rises from 69.4% before fine-tuning to 74.3% with 5% of the rows, 84.8% with 50% and 86.9% with all rows. The curve flattens, and a saturating fit with bootstrap intervals cannot pin down its plateau. Seed spread at 20% of the rows (s.d. 1.6 points) exceeds some of the steps between sizes. Giving small subsets as many updates as the full pool gains 1.9 points at 5% and 1.9 points at 20% of the rows (one seed; p=0.14 and 0.13), and at 20% that is below the 2.3-point difference between two ordinary seeds. A single run with 10% symmetric label noise shows no detectable change (-0.2 points, 95% interval -1.7 to +1.4); 20% costs 3.4 points and 40% costs 8.7 points. Confusable-pair noise at 20% costs 7.4 points, and under noise the model moves from over-confident to under-confident. An independent labeller disagrees with the gold labels on 3.5% of held-out items, and about half of its choice disagreements (16 of 33, leaving each item out of the partner map) name the partner option that the confusable-pair noise uses. Laya is +0.9 points ahead of a TF-IDF classifier trained on the same rows at full size (the interval includes zero) and +21.1 points ahead at 5% of the rows. Everything is synthetic, LLM-labelled and from one backbone, one trainer and mostly one seed.

Explore similar work

Nov 12, 2026

Where in the Input Does a One-Pass Decision Encoder Look?

Laya is a small model that answers a multiple-choice question about a text in one pass: the question, the options (each with a name and a description) and the text form one sequence of at most 512 tokens, the window, in the English release. A longer text is silently cut at the end, and we measure what that costs. Test documents place a held-out row's text, the evidence, at a chosen token offset among distractor text; 177 items, five tasks, three checkpoints (released, short-row fine-tune, long-document fine-tune), four window sizes and 93,275 scored requests. When the evidence lies beyond the window, accuracy falls to 26.4% for the released checkpoint (27.2% for both fine-tunes), near the 27.0% of guessing and below the 35.6% of always giving the commonest label. Confidence gives no warning (AUROC 0.50 to 0.55 for telling cut from read requests; 0.5 = chance), and the runtime's truncation flag also fires on requests whose evidence was visible. The released checkpoint reads visible evidence badly once text precedes it: 38.5% pooled against 60% when the evidence comes first, so a larger window alone gains +6.2 points. Fine-tuning on long documents adds +27.6 points on visible evidence over the short-row fine-tune (95% interval +22.3 to +33.1) and nothing on cut evidence. That is an in-family result: training rows and test documents share generator and question types, and the main filler is of the training kind. With encyclopedic filler never seen in training the gain is +13.2 (+9.9 to +16.6), with rows of other tasks as filler -2.0. Among training-free rearrangements a 312-token window scan helps most (+11.9 points over the default for the released checkpoint) at six to eleven forward passes. Documents are synthetic, labels are LLM-made and one backbone is studied, so these diagnose this model, not long documents in general.
Nov 12, 2026

Recalibrating a Typed-Decision Model in an Afternoon: A Hands-On Recipe for Laya's Confidence Scores

A typed-decision model such as Laya reads a text and a list of options that the caller defines, for example ticket categories, and returns a probability for each option. Code around it compares the top probability with a threshold to answer, abstain or escalate to a person. Earlier work reports, for a Turkish model, that temperature scaling can raise held-out calibration error (ECE). This tutorial gives a five-step recipe (measure, diagnose, fit, validate, deploy) in 130 lines of numpy with standard methods, none new. On 1,065 synthetic, LLM-labelled items the released Laya checkpoint has a pooled ECE of 0.073 but a bucket-macro ECE of 0.143 (0.127-0.184), the mean over buckets (questions grouped by type and option count): the 12-option bucket is over-confident by +0.22, the three-to-five-option bucket under-confident by 0.05, and they cancel. One global temperature makes pooled ECE worse (0.173); a cross-fitted temperature per bucket brings macro-ECE to 0.073 (0.078 over 10 fold assignments) and reaches 0.089 with 10 labelled items per bucket, but holding out one of seven evaluation sets gives 0.099. After the refit, a gate aimed at 10% error answers 46.3% of items (10.5% wrong) against 6.9% on the published confidences, but one cut-off per bucket on the published confidences answers 45.4%, so most of the gain comes from separating buckets. Confidence-only maps (Platt, histogram, isotonic) need 20 or more labelled items per bucket to approach the temperature; with 320 per bucket the lowest macro-ECE belongs to Platt. In these files the retrained checkpoints store no per-bucket table and the stored 12-option value 0.1006 is clamped to 0.5 at runtime; the refit helps 31 of 33 checkpoints on random folds and 21 with a set held out. Fit and test share one generator and LLM labels, so these are findings about these checkpoints and this data.
Oct 22, 2026

Joint or Per-Option? Compact Encoders for Closed-Set Decisions from 17 Million to 1 Billion Parameters

Many language-model pipelines end in a closed-set decision, such as selecting a tool. Prompting a large language model for it is slow and order-sensitive. We ask how small an encoder for such decisions can be and whether it should read the options jointly or separately. In a three-seed study at 32 million parameters, per-option scoring beats joint scoring on a held-out sample of a 37-benchmark decision suite by 5.8 points of chance-corrected skill at matched optimizer steps and by 2.8 at approximately matched training compute; two external benchmarks agree, though not significantly on their unseen tasks. Crossing reader with initialisation shows that the advantage requires a reranker initialisation, whose query-candidate pair input matches per-option reading; from the raw encoder the two readers are indistinguishable. Per-option scoring needs 5 to 13 times more encoder passes: at most twice the median latency, but more than four times the mean. Of six released joint encoders (17 million to 1 billion parameters), the 68- and 150-million-parameter models exceed a 435-million-parameter DeBERTa-v3 zero-shot classifier by 4.9 and 5.9 accuracy points on 209 overlap-audited System1-Arena tasks and are at least on par with it on the 145 tasks unseen in training. The smallest answers a median request in 14 milliseconds on 16 processor threads, against 728 for the classifier. Reordering options changes 12 to 23% of the joint models' answers (per-option scoring: none). Up to 68 million parameters, they lead their size range on a public leaderboard's private, mainly in-distribution tests.