← News

ResearchSources: arXiv · 16 Nov 2026

Active learning framework infers LTLf temporal specifications from trace labels and hints

Researchers describe an active-learning method for LTLf in which a teacher only labels traces and answers natural-language questions, avoiding equivalence queries and counterexample construction.

In plain words

Formulas that describe how systems behave over time are hard to write by hand, so the authors propose learning them by asking questions. Their learner keeps a pool of candidate formulas, generates traces that separate them, and uses a teacher's labels and hints to narrow the pool. In simulated tests it identified every target formula.

What was announced

The framework targets LTLf, a temporal logic used in verification and monitoring, where capturing intended behaviour in formulas is difficult. Unlike earlier active-learning approaches that require the teacher to answer equivalence queries and build counterexamples, often with SAT-based procedures that restart after each new example, this learner only needs trace labels plus natural-language answers. It maintains an enumerated version space of candidate formulas, generates traces that distinguish between them, and uses teacher hints to rank candidates or restrict the search.

The authors report consistency proofs, identification within bounded hypothesis classes with a query bound, and a PAC guarantee for the final hypothesis. In experiments with a simulated teacher across 36 targets, including four standard environments, the learner identified every target when comparing all rivals, needing fewer than seven queries beyond an initial ten-trace sample on average. Correct hints cut queries by up to 16% and candidates by about a fifth; ranking hints never blocked identification, even when wrong. The enumeration-based method was faster than a SAT-based predecessor at every target size.

What it does not show

  • All experiments used a simulated teacher on 36 targets; no evaluation with human teachers or real systems is reported in the abstract.
  • The abstract reports the authors' own guarantees and comparisons; no independent confirmation is included.
  • The abstract does not state whether an implementation is publicly available.

Sources

Source-support review completed 17 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.

Coverage: arXiv: abstract.

Updates and corrections

No corrections so far.

First published 17 Nov 2026

New to the subject? AI from the Ground Up starts from the beginning.