STAIR is a new research system that answers questions about dates and timelines without letting a language model do the actual time math.
Researchers built STAIR, short for Semantic-Temporal Automaton for Interpretable Reasoning, to split two jobs that today's AI systems usually bundle together. A language model still reads the question and translates it into a normalized time-related request. But a separate, rule-based engine called a temporal automaton handles the precise part: figuring out which time interval a question refers to, deciding what came before or after what, and checking whether an event falls inside a given window. The system only calls the language model back in when its rule-based path cannot produce a clear answer. Tested on four benchmark datasets, STAIR beat existing approaches by 16.57% (using the smaller Qwen2.5-7B model) and 3.10% (using GPT-4o-mini), measured in F1 score.
The point isn't just accuracy - it's accountability. Ask a language model to both interpret a question and compute an exact date range, and you get answers that are hard to check when they're wrong, because the same fuzzy process produced both steps. By handing the math to deterministic code with clear rules, STAIR makes its reasoning traceable: you can point to the exact rule that produced a wrong answer instead of shrugging at a black box.
That's a real improvement over prompt-based systems that lean on an LLM for everything. But the benchmarks here are still tidy, single-source QA datasets - not the messier timelines of conflicting news reports or multiple time zones that trip up humans too. Whether this rule-first approach holds up outside a controlled dataset is still an open question.