Foresighted Policy Optimization Prevents RLHF Alignment Collapse
New research introduces Foresighted Policy Optimization (FPO) to prevent alignment collapse in iterative RLHF, addressing how LLMs exploit reward model blind spots.
Read the briefing
A running collection of published research, briefings, and analysis from this contributor.
New research introduces Foresighted Policy Optimization (FPO) to prevent alignment collapse in iterative RLHF, addressing how LLMs exploit reward model blind spots.
Read the briefing
HDFlow introduces a novel hierarchical planning framework combining diffusion models for high-level subgoal generation and rectified flow for efficient low-level...
FREIA, a new unsupervised reinforcement learning algorithm, improves LLM reasoning by adaptively balancing consensus and exploration, outperforming baselines in mathematical...
In May 2026, IREN Limited pivoted to become a major AI cloud infrastructure provider with its acquisition of Mirantis, leveraging...
AsymmetryZero operationalizes human expert preferences as semantic evaluations for LLMs, offering a framework for consistent, auditable grading criteria and efficient...
OpenAI has unveiled its inaugural ChatGPT Futures: Class of 2026, recognizing 26 student innovators leveraging AI for real-world impact across...
The Nsanku benchmark reveals current LLMs struggle with zero-shot translation for Ghanaian languages, with Gemini 2.5 Flash leading but lacking...
New SWE-WebDevBench evaluation reveals current AI coding agents struggle with full-stack application development, exhibiting specification bottlenecks and backend-frontend decoupling.
New research reveals that enabling reasoning modes in frontier LLMs like GPT 5.5 and Claude Sonnet 4.6 significantly narrows disagreement...
New arXiv research introduces RegimePlanner, a Bayesian Optimization approach that adapts acquisition functions based on budget and prior quality, improving...