Researchers have introduced several new approaches to improve how large language models are aligned with safety and policy requirements, according to papers published on arxiv.org.
SpecAlign presents a framework that synthesizes alignment data directly from specification documents, treating provider-authored model specifications as the primary alignment target. According to arxiv.org, the system “combines structured rule annotation, controllable specification instantiation, and multi-agent adversarial data synthesis to generate fine-grained, boundary-aware preference pairs.” Experiments demonstrated that training with SpecAlign “consistently improves rule compliance while preserving general capabilities and avoiding over-conservative behavior.”
DOG-DPO takes a geometric approach to data selection for safety alignment. According to arxiv.org, the framework “treats preference pairs as structured geometric signals” and decomposes multi-dataset preference geometry into global and dataset-specific components. Testing across six safety benchmarks showed DOG-DPO “achieves a strong utility-robustness trade-off using only 11% of the preference pairs,” recovering most safety gains while remaining “entirely teacher-free, training-free.”
CHILLGuard addresses Chinese-language LLM safety specifically. According to arxiv.org, the system introduces “a 5-macro, 31-micro category fine-grained risk taxonomy for Chinese scenarios” to address limitations in existing guardrails that “lack adaptation to Chinese-specific regulatory policies, cultural context and linguistic nuances.”
All three papers were published on June 16, 2026.