Research Briefing — June 3, 2026

Building on yesterday's News announcement, here's the full Cosmos 3 technical paper, Cosmos 3 from NVIDIA is a family of omnimodal world models that jointly process and generate language, image, video, audio, and action within a unified mixture-of-transformers architecture, subsuming vision-language models, video generators, world simulators, and world-action models. It claims state-of-the-art across understanding and generation tasks for Physical AI.

NVIDIA's OmniDreams is a foundation generative world model, post-trained from the Cosmos diffusion model, for real-time closed-loop autonomous vehicle simulation that autoregressively generates action-conditioned sensor observations. Addresses long-tail scenario evaluation beyond reconstruction-based simulators.

LEAP is an agentic framework enabling general-purpose foundation models to achieve state-of-the-art automated formal theorem proving in Lean by decomposing problems and iterating with the Lean compiler. It introduces Lean-IMO-Bench for rigorous evaluation.

Derives closed-form near-optimal linear blockwise transforms for joint weight-activation LLM quantization, called WUSH, combining a Hadamard backbone with a data-dependent second-moment component. It provides provable near-optimality for both integer and floating-point quantizers.

This paper presents the first framework for establishing refusal boundaries for AI agents in offensive cybersecurity contexts, defining principled refusal criteria, task categories warranting refusal, and an evaluation methodology under benign and adversarial conditions. It complements proficiency-focused cyber benchmarks with a safety-refusal dimension.

Reasoning Primitive Induction mines successful ReAct agent traces, clusters recurrent reasoning moves, and converts them into a library of reusable typed pseudo-tools. The induced libraries notably outperform the agents that generated their traces.

A position paper arguing that superintelligence developed via solipsistic unilateral optimization is unlikely to be cooperative, because deploying AI induces endogenous non-stationarity and a self-undermining train-test-deploy gap. It advocates for AI designed to participate in cooperative equilibrium selection.

This paper introduces a benchmark of logic puzzles plus a pipeline that converts reasoning traces into verifiable graphs of claims and dependencies, allowing reasoning to be measured structurally rather than only by accuracy or token count. It defines a reasoning efficiency metric capturing how concentrated a model's logical flow is, revealing behaviors hidden by standard metrics.

Economy of Minds studies how a population of LLM agents can self-organize into stronger collective intelligence through economic interactions—auctions for the right to act, payments, and wealth accumulation—inspired by Hayekian decentralized coordination. Economic signals induce decentralized credit assignment and evolutionary selection without central control.

Shows that backdoor neutralization via unlearning generalizes, where training a model to ignore one trigger can also suppress other never-targeted backdoors across three model families. This offers a path to removing unknown backdoors without knowing triggers.

Introduces a nested geometric decomposition framework analyzing how prompting reshapes internal representations of LLMs and VLMs, using increasingly expressive stimulus-invariant alignment maps and causal layer interventions. It seeks to explain mechanistically how instructions steer behavior.

TRAP demonstrates the first targeted behavior-hijacking attack on Vision-Language-Action robotic models by exploiting their Chain-of-Thought reasoning via adversarial patches, e.g., causing a robot to deliver a knife instead of an apple. It shows CoT strongly governs action even when misaligned with instructions.