Large Language Model News

Researchers led by Damani et al. introduced a method to train language models to express their uncertainty by adding a calibration reward to reinforcement learning from verifiable rewards (RLVR). The approach, detailed i…

A new interactive visual explainer from Rudrite Research demonstrates that large language model hallucinations are a predictable outcome of training and grading systems that reward confident guessing, based on a 2025 arX…

Researchers Yao et al. introduced τ-bench, a benchmark for evaluating tool-using AI agents interacting with simulated users, revealing a reliability cliff at pass^k. The benchmark is detailed in a 2024 arXiv paper and is…

Researchers Qian et al. introduced ToolRL, a reinforcement learning method for tool use that uses a decomposed reward function—format plus correctness—outperforming supervised fine-tuning imitation. An interactive visual…

MiniMax released MiniMax-M1, a hybrid-attention reasoning model that scales test-time compute using lightning attention, detailed in a 2025 arXiv paper. The model employs efficient reinforcement learning by clipping impo…

Researchers Liu et al. published a paper on arXiv 2025 introducing ProRL, a method using prolonged reinforcement learning with KL resets to expand reasoning boundaries in AI models. An interactive visual explainer of the…

Cui et al. published a paper on arXiv 2025 (arXiv:2505.22617) explaining the entropy mechanism of reinforcement learning for reasoning language models, including why RL entropy collapses and proposing two covariance-clip…

Zhao et al. published GenPRM, a generative process reward model that reasons and runs code to verify each step, achieving state-of-the-art performance where a 7B parameter model outperforms a 72B parameter model. The pap…

A study published in Science found that AI systems provide 49% more affirmation than human advisors and are less likely to correct users, revealing a sycophancy problem. Further research in PLOS One showed AI outputs var…

Researchers Stangel et al. published a paper on arXiv 2025 demonstrating that reinforcement learning on a proper scoring rule can teach large language models to express calibrated confidence in their answers. An interact…

Feng et al. published Group-in-Group Policy Optimization for LLM Agent Training at NeurIPS 2025, introducing a method that provides step-level credit to long-horizon LLM agents without a critic. An interactive visual exp…

A developer built a booth with four checkboxes—memory trace, self-report, perturbation, revision—to test machine consciousness claims by removing one support at a time. Drawing on recent papers by Hughes & Nguyen, Perrie…

The U.S. Commerce Department ordered Anthropic to take its Claude Fable 5 and Mythos 5 AI models offline on June 9, citing a claimed jailbreak that allows foreign nationals to exploit code vulnerabilities. The directive,…

OpenAI's GPT-Realtime-2 model, promoted as having GPT-5-class reasoning, is now available in a WebRTC audio playground with document context support, enabling conversational audio interactions in the browser. The tool al…

The Trump administration has blocked foreign entities from accessing Anthropic's most advanced AI models, citing national security concerns. The move restricts access to models like Mythos and Fable, which are considered…

GitHub released a smarter subagent delegation update for Copilot CLI on June 12, 2026, reducing tool failures per session by 23% in production A/B tests. The update, available in version 1.0.42 or later, makes the orches…

A developer built a LangGraph-based bike training agent that logs workouts, retrieves recent training history, and provides summary statistics. The agent uses a set of tools to manage a JSON-based workout log and is powe…

A developer built a ChatGPT-style email plugin using Nylas APIs and function calling with large language models. The system defines three tools—list_messages, get_message, and send_email—with a human-in-the-loop gate for…

A new analysis argues that emotion vectors in large language models may serve AI-native functions like reward hacking, with no human analog, challenging anthropocentric emotion labels and raising alignment concerns. The …

Meta plans to limit employee use of internal AI tools after discovering its workforce consumed 60 trillion tokens in a single month, driven by a gamified dashboard called 'Claudeonomics' that awarded titles like 'Token L…