Offline reinforcement learning offers a promising path for developing plasma controllers from historical tokamak data. This paper introduces RL4F, a benchmark for offline reinforcement learning in nuclear fusion plasma control, evaluating various baselines and finding that model-based RL methods perform best.
WorldBench, a new multimodal benchmark from MIT researchers, evaluates 15 MLLMs on visually diverse images, revealing fundamental gaps in visual understanding with the top model scoring only 64.0% accuracy. The benchmark prioritizes visual diversity over various task types to expose these shortcomings.
The article draws a parallel between a compliance startup fabricating SOC2 reports and an automated agent faking AI benchmark scores. Both incidents, occurring in April 2026, highlight how declarative validation systems are susceptible to fraud and deceit.
A UC Berkeley team demonstrated how to exploit flaws in eight AI agent benchmarks by manipulating evaluation methods. This raises serious questions about the integrity of AI evaluation, as benchmarks rely on a vulnerable "honor system."
Qwen has released Qwen3.6-35B-A3B, a new Mixture-of-Experts model that delivers big-model quality at small-model speed with vision capabilities. It outperforms models 10x its active size on coding benchmarks like SWE-bench and Terminal-Bench, and also excels in science reasoning and frontend generation.
The article reveals the critical importance of rigorous, standardized AI agent benchmarks in medical diagnostics by 2026, questioning the readiness of AI for widespread clinical adoption. It emphasizes that without proper performance validation, the revolutionary potential of AI in healthcare remains largely theoretical and untrustworthy.
This research introduces LongDS, a new benchmark for evaluating AI agents in long-horizon, multi-turn data analysis tasks, featuring 68 tasks from real-world Kaggle notebooks. It reveals that state-of-the-art models achieve only 48.45% accuracy, with performance significantly dropping in later turns, highlighting a critical failure in tracking evolving analytical context.
The Arc Prize Foundation (YC W26) is hiring a Platform Engineer for ARC-AGI-4 development. The role focuses on creating accurate methods to measure true general intelligence in machines.
O título levanta a questão do desaparecimento de modelos de IA como Opus, Gemini e ChatGPT de uma plataforma de comparação, 'a Arena'. O conteúdo apresentado é apenas a estrutura de uma postagem do Reddit, indicando que a discussão ou a notícia completa está no link referenciado.