Large Language Models (LLMs) have shown impressive reasoning capabilities in agents for complex interactive environments. However, these agents frequently propose actions that violate physical or procedural constraints. Recent neuro-symbolic guards mitigate such errors by compiling explicit admissibility rules from oracle demonstrations, yet treat them as fixed laws—a closed-world assumption that breaks down under sparse demonstrations or environment shift, leaving either missing rules that fail to block unsafe actions or stale rules that wrongly reject valid alternatives. To address this, we introduce **AdaptiveGuard** ([code](https://github.com/lpxwork666111/AdaptiveGuard)), an Adaptive framework that Guards LLM agents by treating constraints as *falsifiable hypotheses*. Each rule is represented by a Beta-distributed validity belief mapped to *tentative*, *confirmed*, and *deprecated* tiers; uncertain rule blockers are then challenged through a zero-extra-call falsification gate; and these rule beliefs are continuously updated via contrastive feedback attribution from environment evidence. Benchmark experiments on ScienceWorld and VirtualHome show aggregate gains over strong static constraint guards, especially in novel scenarios. Complementary post-hoc diagnostics under sparse supervision and controlled shift show improved adaptation, while observed failed-trial runs remain below the configured budget cap and the conservative-bias proxy declines over the measured horizon.

Large Language Models (LLMs) have shown impressive reasoning capabilities in agents for complex interactive environments. However, these agents often suffer from hallucinations and lack grounding, leading to unreliable actions that conflict with real-world constraints. Existing approaches mitigate some issues through implicit imitation or sparse reinforcement learning but rely on fitting data distributions without explicitly understanding environmental constraints, often generating actions that are behaviorally distorted or environmentally impermissible. To address this, we introduce OntoGuard, an ontological framework designed to guard LLM agents by enforcing environmental and behavioral admissibility. These constraints are constructed by extracting knowledge from oracle demonstrations, supplemented with world knowledge inherent in LLMs and general knowledge bases. During inference, OntoGuard functions as an active interceptor, using a graph-based constraint-checking mechanism to reject invalid actions and prompt self-correction before acting. Experiments on both ScienceWorld and VirtualHome demonstrate OntoGuard’s advantage over state-of-the-art methods, validating its ability to enforce physical and behavioral constraints while preventing invalid actions.