A Formalization of the Mean-Field Derivation of the Vlasov Equation – AI-Assisted Lean Formalization as a Strategy Game
A mathematician “plays” a strategy game in Lean 4: direct an AI to turn a LaTeX proof into machine-checked Lean code with no sorrys, where the build itself certifies correctness. The case study fully formalizes well-posedness for the nonlinear Vlasov equation using Dobrushin’s mean-field approach, including existence, uniqueness, stability, and the mean-field limit, with an axiom-clean optimal-transport layer that compiles against Mathlib alone.
What Anthropic’s latest AI discovery does – and doesn’t – show
The report breaks down Anthropic’s new research claim about AI capability and interpretability, highlighting what the experiments support and what they cannot conclude. It frames the finding as a step toward understanding model behavior, while cautioning against over-reading results when evidence is indirect or under-specified.
Interval Certifications for Multilayered Perceptrons via Lattice Traversal
This work recasts adversarial robustness for multilayer perceptrons as a lattice traversal problem over interval “certification boxes” – defining both sound certifications (prediction is invariant inside the box) and complete certifications (prediction must change outside). It introduces lattice traversal operators with a refine-and-verify scheme that guarantees maximal soundness and minimal completeness, then shows strong intractability for sound certification optimization while finding efficient logarithmic methods for symmetric \ell_\infty cases.
KV-PRM: Efficient Process Reward Modeling via KV-Cache Transfer for Multi-Agent Test-Time Scaling
KV-PRM tackles a key bottleneck in process reward models by avoiding quadratic re-encoding of long trajectories: it reads directly from the KV cache produced during generation. The paper proves KV cache carries strictly greater information capacity than text for reward modeling, and reports major efficiency wins – up to a 5,000x reduction in scoring FLOPs – while matching or beating text-based PRMs on MATH, GSM8K, and AIME under multiple test-time scaling strategies.
Long-Horizon-Terminal-Bench: Testing the Limits of Agents on Long-Horizon Terminal Tasks with Dense Reward-Based Grading
Instead of judging agents only by final success, this benchmark adds dense, graded subtasks inside 46 long-horizon terminal tasks so you can measure partial progress. Across evaluations with 15 frontier models, the strongest system reaches only 15.2% pass@1 at a near-perfect threshold, underlining how hard long-horizon planning and iterative debugging remain compared with short terminal challenges.
CogniConsole: Externalizing Inference-Time Control as a Formal Abstraction for Reliable LLM Interactions
CogniConsole argues that reliability failures often come from weak “inference-time control” – the logic that frames tasks and selects context – not just from model quality. By externalizing control into a structured interface that mixes programmatic coordination with bounded prompt reasoning, the authors show that increasing scaffolding systematically reduces output variance and failure rates in multi-step interactive settings.
GATS: Graph-Augmented Tree Search with Layered World Models for Efficient Agent Planning
GATS aims to make planning cheaper and more deterministic by removing LLM calls during inference-time search. It combines UCB1-style tree search with a layered world model that uses symbolic action matching, execution-log statistics, and an LLM predictor only for unknown actions, reaching 100% success on synthetic stress tests while maintaining 100% across a broader suite where LATS and ReAct degrade.
MedRealMM: A Real-World Multimodal Benchmark for Chinese Online Medical Consultation
MedRealMM builds a multimodal benchmark directly from de-identified patient-doctor interactions at a nationwide Chinese internet hospital, not from simulators or synthetic dialogs. Using a physician-refined rubric on 5,620 cases with text-image context, the study finds that images are critical for reliable clinical performance and that frontier models still underperform physicians on safety-sensitive error avoidance.
ARCANA: A Reflective Multi-Agent Program Synthesis Framework for ARC-AGI-2 Reasoning
ARCANA tackles ARC-AGI-2-style tasks by orchestrating multiple agents that alternate between perceptual grounding, DSL program hypothesis generation, symbolic execution, and reflection-driven refinement. With a shared differentiable blackboard and a learned meta controller, the framework iteratively corrects failures under tight test-time and hardware constraints to improve solution quality on abstract transformation problems.
Lorde says Ray-Ban Meta AI glasses are ‘not sexy’
During a festival set, Lorde criticized Meta’s Ray-Ban AI glasses, saying it’s increasingly hard to tell what’s real and calling the product “fucked up” in a wider rant about AI wearables. While the comments are more culture than engineering, they signal growing skepticism around AI gadgets that blend into everyday life.
Waze is getting a bunch of new AI-powered features
Waze is integrating Google’s Gemini to make the navigation app more conversational and customizable, including improvements to incident reporting via voice. The update also adds Destination Search with natural voice requests, aiming to reduce friction for drivers while keeping answers actionable inside the driving context.
The 6 wildest claims in Apple’s lawsuit against OpenAI
Apple’s lawsuit against OpenAI alleges broad misconduct tied to trade secrets and hardware collaboration, including claims that OpenAI employees requested unusual prototype materials during interviews. The report pulls out the most striking allegations – the kind that, if proven, would reshape how AI companies treat partnerships, prototypes, and proprietary development artifacts.
Neuro-Agentic Control: A Deep Learning-based LLM-Powered Agentic AI Framework for Controlling Security Controls
This paper proposes an LLM-driven defense loop for industrial IoT security that’s designed to avoid hallucinated actions – a major safety risk in closed-loop control. It uses a foundation time-series model as a deterministic “sentinel” and a counterfactual physics injection mechanism to simulate proposed interventions before actuation, reporting fewer breaches and zero physically invalid actions on an industrial dataset under stochastic attacks.
Interval Certifications for Multilayered Perceptrons via Lattice Traversal
This article/abstract overlaps with a separate arXiv entry in the provided set and focuses on turning robustness certification into a structured lattice search over input intervals for MLPs. It formalizes optimization trade-offs between sound and complete certification and shows how refine-and-verify with lattice operators can guarantee specific certification properties.