WikiPlanner: wiki-derived temporal monitors for safe LLM agent planning - a medical-domain proof of concept
AI Agent · 2026
DOI: 10.20517/aiagent.2026.22 ↗Kurz gefasst
Fachartikel in der Zeitschrift AI Agent, 2026. Verfasst von Bailing Zhang. Die Arbeit nennt 18 Literaturangaben.
Abstract (Originalsprache)
Large language model (LLM) agents in high-stakes sequential planning may produce locally reasonable steps whose overall sequence violates domain constraints; single-step filters cannot detect such cross-step temporal violations. We present WikiPlanner, which extracts temporal constraints from domain wikis, structures them via a typed Temporal Constraint Intermediate Representation (TC-IR) with seven constraint labels compiled to six Linear Temporal Logic over finite traces (LTLf) templates, and instantiates them as parallel runtime monitors. The two-stage pipeline - TC-IR extraction followed by deterministic template compilation - isolates LLM extraction uncertainty from formal compilation. We evaluate every stage on a medical planning benchmark of 45 constraints and 236 test plans (137 aligned, 99 natural). A dual-annotator audit of the 220 raw extractions (κ = 0.451, adjudicated) places semantic extraction precision at 25.9%; an automated check finds 79.5% of these records structurally well-formed (90.1% on a held-out nursing corpus); invented orderings, state-vs.-action confusion, and silent parse loss from output truncation are the dominant error classes. Plan validation achieves F1 = 0.920 (95%CI: 0.873-0.956) on the aligned set and F1 = 0.582 (95%CI: 0.447-0.701) on the natural set; the gap is attributed primarily to constraint-library coverage and symbol grounding, as wiki-derived constraint symbols and generator vocabulary form largely disjoint namespaces. Latency scales linearly with library size (0.006-0.059 ms). Single-pass template repair resolves 39.1% of violation plans (mean 1.30 word-level edits per success); local post-trigger repairs substantially outperform pre-trigger modifications. Results are a single-domain proof of concept, and symbol grounding is the main obstacle to natural-plan deployment.
Zitieren
APA
Zhang, B. (2026). WikiPlanner: wiki-derived temporal monitors for safe LLM agent planning - a medical-domain proof of concept. AI Agent. https://doi.org/10.20517/aiagent.2026.22
Vancouver
Zhang B. WikiPlanner: wiki-derived temporal monitors for safe LLM agent planning - a medical-domain proof of concept. AI Agent. 2026. doi:10.20517/aiagent.2026.22
BibTeX
@article{Zhang2026,
author = {Zhang, Bailing},
title = {WikiPlanner: wiki-derived temporal monitors for safe LLM agent planning - a medical-domain proof of concept},
journal = {AI Agent},
year = {2026},
doi = {10.20517/aiagent.2026.22}
}
EUVIMED