AI Daily — June 29, 2026
2026-06-29
TODAY'S NEWS
rss.arxiv.org
Internalizing the Future: A Unified Agentic Training Paradigm for World Model Planning
This paper proposes training LLM agents to internalize world-model-style planning by generating both prospective state rollouts and plan-conditioned success estimates (a textual Q-value analogue) within a single autoregressive model. The authors identify a 'format-capability gap' where naive fine-tuning on look-ahead traces produces superficial mimicry rather than genuine predictive grounding, and introduce a training paradigm to address this. The work targets a fundamental limitation of current agents in long-horizon sequential decision-making tasks.
LAST WEEK'S TOP STORIES
OpenAI and Broadcom unveil LLM-optimized inference chip
OpenAI publicly disclosed its first custom silicon effort with Broadcom, a strategic move to reduce dependence on NVIDIA that signals a broader industry shift toward purpose-built LLM inference hardware.
Introducing computer use in Gemini 3.5 Flash
Google DeepMind added desktop computer use capabilities to Gemini 3.5 Flash, entering direct competition with Anthropic in GUI-based agentic tasks while targeting latency-sensitive, cost-efficient deployments.
AI Decisively Outperforms Expert Humans at Persuasion — Oxford/Stanford/AISI Study
A rigorous multi-institution study found AI systems reliably outperform expert human communicators at persuasion, establishing a concrete empirical basis for concerns about AI-mediated manipulation and influence operations at scale.
OpenAI Daybreak: Codex Security and GPT-5.5-Cyber for Vulnerability Management
OpenAI launched its first purpose-built enterprise security product line, combining a specialized cybersecurity model with end-to-end vulnerability detection and remediation tooling, marking a significant expansion beyond general-purpose AI.
Perfect Detection, Failed Control: The Geometry of Knowing vs. Steering in Language Models
Researchers demonstrated that perfect linear separability in detecting a model behavior does not imply the ability to steer it, directly challenging a foundational assumption underlying much of mechanistic interpretability and activation-based alignment work.