{"slug": "uo-fie-combining-exact-label-supervision-with-graded-utility-for-factivity", "title": "UO-FIE: Combining Exact-Label Supervision with Graded Utility for Factivity Inference", "summary": "UO-FIE, a parameter-efficient system built on Qwen3.5-9B with LoRA, ranked first in the fine-tuning track of the Factivity Inference Evaluation 2026 (FIE2026) with a macro utility of 0.8316, according to the paper's abstract. The system combines hard-label supervision, utility-based soft targets, scheduled class weights, and an ordinal loss to predict a distribution over nine ordered factivity intervals for Chinese context-hypothesis pairs, addressing a training set where 64.1% of 566 examples fall in a single class. A separate prompt-based ensemble ranked third in the non-fine-tuning track with a macro utility of 0.8450.", "body_md": "arXiv:2609.28605v1 Announce Type: new \nAbstract: The Factivity Inference Evaluation 2026 (FIE2026) classifies Chinese context-hypothesis pairs into nine ordered factivity intervals. Its evaluation metric rewards both exact predictions and proximity to the correct interval, while 64.1% of the 566 training examples belong to a single class. In preliminary experiments, several mDeBERTa classification models predominantly predict the dominant class, whereas a Huber-regression baseline produces more predictions near the correct interval but fewer exact matches.\n  We introduce Utility-Oriented Factivity Inference (UO-FIE), a parameter-efficient system that combines exact-label supervision with graded utility. UO-FIE predicts a distribution over the nine classes and combines hard-label supervision, utility-based soft targets, scheduled class weights, and an ordinal loss. We evaluate expected-utility decoding in controlled comparisons and use ordinal calibration selected on out-of-fold predictions for the submitted system.\n  Based on Qwen3.5-9B with LoRA, UO-FIE ranks first in the fine-tuning track with a macro utility of 0.8316. A separate prompt-based ensemble ranks third in the non-fine-tuning track with a macro utility of 0.8450.", "url": "https://wpnews.pro/news/uo-fie-combining-exact-label-supervision-with-graded-utility-for-factivity", "canonical_source": "https://arxiv.org/abs/2609.28605", "published_at": "2026-09-25 04:00:00+00:00", "updated_at": "2026-09-25 04:30:15.355469+00:00", "lang": "en", "topics": ["natural-language-processing", "machine-learning", "ai-research"], "entities": ["UO-FIE", "FIE2026", "Qwen3.5-9B", "LoRA", "mDeBERTa"], "also_reported_by": [], "alternates": {"html": "https://wpnews.pro/news/uo-fie-combining-exact-label-supervision-with-graded-utility-for-factivity", "markdown": "https://wpnews.pro/news/uo-fie-combining-exact-label-supervision-with-graded-utility-for-factivity.md", "text": "https://wpnews.pro/news/uo-fie-combining-exact-label-supervision-with-graded-utility-for-factivity.txt", "jsonld": "https://wpnews.pro/news/uo-fie-combining-exact-label-supervision-with-graded-utility-for-factivity.jsonld"}}