{"slug": "uniguard-multi-view-guardrail-learning-for-cross-interface-safety-generalization", "title": "UniGuard: Multi-View Guardrail Learning for Cross-Interface Safety Generalization", "summary": "A team of researchers has proposed UniGuard, a training framework that learns interface-invariant safety representations by showing a guardrail model multiple semantically aligned views of the same safety case, and used it to build UniGuard-0.6B. Trained on the UniGuard-FT dataset, which transforms 1,000 public safety cases into prompt, response, multi-turn conversation, agent-trace, and multi-agent-trace views, the 0.6B model improves over its pretrained guardrail by 45.12 F1 points on average across the unseen-interface benchmarks ATBench, PolicyGuardBench, and DynaBench and reduces expected calibration error by 35.41 points. The researchers report UniGuard-0.6B remains competitive with much larger guardrails such as Claude-Sonnet-4.5 and GPT-5.2 despite its size and training only on synthetic data.", "body_md": "arXiv:2610.30742v1 Announce Type: new \nAbstract: Guardrail models are increasingly used as safety layers for LLMs and agentic systems, but most are trained on fixed prompt or response interfaces and become unreliable when the same safety case appears as a conversation, agent trajectory, multi-agent trace, or policy-conditioned input. We propose UniGuard, a training framework that learns interface-invariant safety representations from multiple semantically aligned views of the same safety case. We also construct the UniGuard-FT dataset by transforming 1,000 public safety cases into prompt, response, multi-turn conversation, agent-trace, and multi-agent-trace views, and train UniGuard-0.6B with classification, supervised contrastive learning, and asymmetric confidence calibration losses. Across unseen-interface benchmarks ATBench, PolicyGuardBench, and DynaBench, UniGuard-0.6B improves over its pretrained guardrail by 45.12 F1 points on average and reduces expected calibration error by 35.41 points, while remaining competitive with much larger guardrails such as Claude-Sonnet-4.5 and GPT-5.2 - despite being a 0.6B model and trained only with synthetic data.", "url": "https://wpnews.pro/news/uniguard-multi-view-guardrail-learning-for-cross-interface-safety-generalization", "canonical_source": "https://www.machinebrief.com/news/uniguard-multi-view-guardrail-learning-for-cross-interface-s-zjfj", "published_at": "2026-10-23 04:00:00+00:00", "updated_at": "2026-10-23 04:42:35.024006+00:00", "lang": "en", "topics": ["ai-safety", "ai-agents", "large-language-models", "ai-research", "machine-learning"], "entities": ["UniGuard", "UniGuard-0.6B", "UniGuard-FT", "ATBench", "PolicyGuardBench", "DynaBench", "Claude-Sonnet-4.5", "GPT-5.2"], "also_reported_by": [], "alternates": {"html": "https://wpnews.pro/news/uniguard-multi-view-guardrail-learning-for-cross-interface-safety-generalization", "markdown": "https://wpnews.pro/news/uniguard-multi-view-guardrail-learning-for-cross-interface-safety-generalization.md", "text": "https://wpnews.pro/news/uniguard-multi-view-guardrail-learning-for-cross-interface-safety-generalization.txt", "jsonld": "https://wpnews.pro/news/uniguard-multi-view-guardrail-learning-for-cross-interface-safety-generalization.jsonld"}}