Four new papers published on arXiv examine different aspects of how language models handle uncertainty and learn from training data.
According to arxiv.org, a study titled “Act or Escalate?” models automation decisions as choices under uncertainty, where LLMs must decide when to act versus escalate to humans. Testing across five domains including demand forecasting, content moderation, and autonomous driving, the researchers found “marked differences in the implicit thresholds models use to trade off these costs,” with thresholds varying substantially and not predicted by architecture or scale. The study reports that supervised fine-tuning on chain-of-thought targets “yields the most robust policies, which generalize across datasets, cost ratios, prompt framings, and held-out domains.”
In a separate paper on misinformation simulation, arxiv.org reports that LLM-generated survey responses “consistently overstate the association between belief and sharing” and place “disproportionate weight on attitudinal and behavioral features, while largely ignoring personal network characteristics.” The paper, accepted to ICWSM 2026, suggests LLM-based simulations are “better suited for diagnosing systematic divergences from human judgment than for substituting it.”
A third study examines preference optimization methods like DPO and KTO. According to arxiv.org, researchers investigated what drives reasoning improvements, finding that “increasing generator-level delta steadily improves performance on out-of-domain reasoning tasks” and that filtering by sample-level delta “can enable more data-efficient training.”