A new paper on arXiv argues that large language models are still unreliable when asked to reason about the future. The authors identify a core problem: models tend to treat all future states as equally knowable, ignoring the fact that some events are far more predictable than others.
The paper points out that existing models often commit to specific answers even when the queried facts are inherently volatile. That overconfidence, the authors suggest, is a key source of errors in temporal reasoning tasks.
To address this, they propose explicitly modeling future predictability — that is, having the model account for how uncertain or changeable a future state is before giving an answer. The approach is framed as a step toward more reliable temporal reasoning in LLMs.
Since only one source was provided, there are no contrasting findings to compare; the summary reflects the paper’s own claims. The full methodology and experimental results are not detailed in the abstract.}