June 29, 2026

JetSpec introduces a causal parallel draft head that resolves the causality-efficiency dilemma in speculative decoding. Standard tree-based drafters either draft autoregressively (accurate but slow) or in one parallel pass (fast but incoherent). JetSpec trains a draft head over the target model's fused hidden states so candidate-tree token scores follow the target's autoregressive factorization, then verifies the full tree in a single forward pass. On coding and math benchmarks it achieves up to 9.64× speedup over standard autoregressive decoding on H100/B200 GPUs. Code is open-sourced.

OPID adds dense, token-level supervision to outcome-based RL for LLM agents. During training, a lightweight LLM analyzer extracts two levels of hindsight skill from completed trajectories: episode-level workflow summaries and step-level action rationales at critical decision points. A critical-first routing mechanism injects the appropriate skill into the interaction history, letting the policy contrast responses with and without skill guidance for token-level advantage estimation. On ALFWorld, WebShop, and Search-QA, OPID improves task completion, sample efficiency, and robustness over baseline outcome-only RL.

SingGuard is a guardrail model for vision-language models that accepts natural-language safety policies at runtime rather than using rules baked in at training time. It evaluates content against policy rules one-by-one with three inference speed modes (fast/hybrid/slow) to trade interpretability for latency. A new benchmark, SingGuard-Bench, contains 56,340 examples across 80+ risk categories including cross-modal joint-risk cases where neither text nor image alone is harmful but their combination implies unsafe intent. Policy-following accuracy improves from ~64.6% to ~74.1% over prior methods on runtime policy changes.