Regularized Emphatic Temporal-Difference Learning: Stability under Constant Stepsizes
This arXiv paper shows why emphatic TD learning can look “stable” on average while the actual sampled dynamics still explode – by constructing an ergodic two-state counterexample where mean contraction and positive Lyapunov growth diverge. It then proposes Regularized Emphatic TD (RETD) – a post-shock repair mechanism that provably restores fixed points and achieves convergence guarantees under diminishing steps, plus conditional constant-stepsize moment-contraction.
Closed-World Resolution Against Tool Hallucination in LLM Agents
This study argues that tool hallucinations are a structural blind spot: a model can emit tool calls that are not decisions any “gate” can block because the call refers to a non-existent registry entry. It introduces a five-class taxonomy of hallucinated tools, proves defense must precede gating, and benchmarks a training-free “Resolution Rung” approach using registry membership and signature checks across dozens of model invocations.
California pushes a potential AI kill switch into state oversight
California Gov. Gavin Newsom’s new executive order directs work on how frontier AI companies might be required to support an “AI kill switch” alongside independent verification and audit standards. The proposal also emphasizes operational enforcement – with recommendations expected within two months on what should become state law.
Virtualize foundation models with a Self-evolving Operating System layer
This position paper makes the case that today’s agent stacks are fragmented – each protocol hides its own runtime for state, memory, budgets, and guardrails – making behavior hard to port and governance brittle. The proposed Foundation Model Operating System (FMOS) would act like a runtime virtualization layer for models, coordinating memory tiers, model selection, resource allocation, and policy enforcement under adaptive control.
Security researchers used Claude to hack into OpenAI employee accounts
A report claims three independent researchers spent less than 72 hours using Claude (Anthropic’s) to gain access to OpenAI employee accounts, ultimately reaching OpenAI’s “Monorepo” via third-party community infrastructure. While they stopped short of deeper internal code access, the episode underscores how quickly tool-use and automation can translate into real-world account compromise pathways.
Virginia orders AI and data-center task force action to restrain approvals
Virginia Gov. Abigail Spanberger signed an executive order aimed at increasing local input and slowing data-center approvals, including requirements tied to noise and backup-generation operations. The order also establishes an AI task force to assess risks to residents – reflecting mounting pressure on how states manage compute expansion and AI infrastructure impacts.