Large language models exhibit non-negligible risks of generating policy-violating content, underscoring the need for effective and timely guardrail mechanisms. Existing streaming safety monitoring approaches enable detection during decoding, but still lack a reliable mechanism for deciding when to terminate generation, limiting their ability to act on early risk signals that are often sparse and unstable under partial observability. In response, we propose LTS, a token-level framework that explicitly learns a stopping criterion for preemptive early termination. LTS introduces a risk-aware STOP head for token-level risk assessment and leverages temporally structured supervision with multiscale evidence to model the evolution of safety risks during generation. By casting streaming safety monitoring as a constrained online stopping task and incorporating exponential moving average temporal aggregation, LTS enables reliable translation of emerging risk evidence into stable and timely termination decisions. LTS achieves an F1 score of 95.5% and is able to terminate harmful content at 13% of the generated sequence, substantially outperforming existing methods during early generation. In realistic streaming settings, it achieves around 90% token reduction.

Evidence-Based Medical Question Answering (EB-MQA) aims to provide accurate and verifiable responses to support clinical decisions. To retrieve evidence from general medical knowledge graphs (KGs), existing knowledge augmentation methods use semantic similarity or dynamic sampling for entity matching and agentic expansion or subgraph scoring for path retrieval. These approaches can introduce entity-level noise from irrelevant or generic entities and path-level noise from redundant or spurious reasoning chains. We propose ToToFilter, a train-free topic- and topology-aware noise filtering framework with Topic-Aware Entity Localization for query-relevant entity anchoring and Topology-Aware Path Filtering for path selection based on semantic consistency, structural compactness, and information specificity. Across four MQA benchmarks, ToToFilter achieves average accuracy gains of 2.30 and 1.22 percentage points over DALK on Llama3.1-8B and Qwen2.5-7B, with maximum gains of 2.60 and 2.78 points. On PrimeKGQA, it improves answer recall and triple recall by up to 8.81 and 19.70 percentage points. Further analyses verify the effectiveness of the proposed two-level filtering framework.