A parameter-efficient AI system called UO-FIE just won a competition that tests whether machines can tell how strongly a sentence commits to something being fact.
The competition, FIE2026, sorts Chinese context-hypothesis pairs into nine ordered "factivity" intervals, and its scoring rewards both exact answers and near-misses. The training data is lopsided: 64.1% of the 566 examples fall into a single class. That skew broke earlier attempts - mDeBERTa classifiers mostly just guessed the dominant class, while a simpler Huber-regression baseline landed closer to correct answers more often but nailed exact matches less often. UO-FIE, built on Qwen3.5-9B with LoRA fine-tuning, split the difference by combining hard-label training with a graded utility score, scheduled class weighting, and an ordinal loss, then applying expected-utility decoding and calibration. It won the fine-tuning track with a macro utility of 0.8316.
Factivity inference underpins anything that has to treat text as an assertion rather than a hedge - search, summarization, fact-checking tools. But the results also expose a wrinkle in how these systems get judged: a separate prompt-based ensemble with no fine-tuning at all scored a higher macro utility, 0.8450, yet placed third because rival non-fine-tuned entries scored even higher.
A benchmark that rewards proximity as well as precision is a reasonable fix for skewed data, but it also means the leaderboard order and the raw utility scores can tell different stories - worth remembering before crowning any single approach the winner.