POLAR: Ontology-Guided Risk Prevention for Tool-Calling LLM Agents
POLAR prunes risky tool calls using ontology-based reversibility scores, with mixed tau2-bench gains.
POLAR is a guardrail for small tool-calling LLM agents that assigns each action a graded reversibility score by deriving a candidate inverse sequence from a two-layer ontology, then prunes calls below a threshold before execution. On tau2-bench across six agent models, mean task reward rose by 0.11 to 0.18 points on the airline domain for four of six agents, but only eight of eighteen model-domain cells improved, and retail and stronger agents often regressed. The authors present an auditable structural check and note that reward is not a direct measure of prevented harm.
- Two-layer ontology scores action reversibility before execution.
- Airline reward rose 0.11 to 0.18 for four of six agents.
- Only eight of eighteen model-domain cells improved overall.
- Retail and stronger agents often regressed; reward is not harm prevented.
From the original article
1 min read at arxiv.orgLLM tool-use agents operate in dynamic environments where many actions carry operational risk. However, most safety mechanisms react only after errors manifest. Existing pre-emptive approaches either fine-tune the agent on chain-of-thought deliberation or compile natural-language guardrails into runtime checks, but they do so without exposing a structural, auditable verdict.
By Yunju Kang · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.