Large language model (LLM) news — GPT-4, Claude, Gemini, Llama, Mistral and the latest research on training, fine-tuning, RLHF, and deployment of LLMs.
Google promoted its Interactions API to general availability on June 22, 2026, making it the default interface for Gemini models and agents. The new API introduces typed steps, managed agents with Linux sandboxes, backgr…
Sakana AI released Fugu Ultra, a model that coordinates multiple frontier models to produce answers, now available on Vercel's AI Gateway. The model matches capabilities of Claude Mythos Preview and Fable 5 on reasoning …
An AI researcher gave a large language model control of a civilization in the game Civilization VI to test its decision-making capabilities. The AI built a successful trade network and alliances but, when faced with an u…
Researchers introduced CoLabScience, a proactive AI assistant that uses a reinforcement learning framework called PULI to autonomously intervene in biomedical research discussions. The system, trained on a new benchmark …
Researchers at ACL 2026 revealed PriceBlind, a visual adversarial attack that exploits a vulnerability called Visual Dominance Hallucination in multimodal large language models used as financial agents. The attack achiev…
Researchers introduced TRuST, a large-scale dataset with ~300k annotations for toxicity recognition, unifying prior resources through a synthesized definition. Benchmarking showed fine-tuned PLMs outperform LLMs on toxic…
Researchers at the Association for Computational Linguistics found that large language models' numerical accuracy drops significantly when numbers are presented in underrepresented scripts or formats, even when the under…
Researchers at ACL 2026 introduced a method to reuse LLM hidden states for classification tasks, training lightweight probes on token- and layer-selective representations to eliminate separate guard models. Their two-sta…
Researchers from multiple institutions published a comprehensive survey on Process Reward Models (PRMs) at ACL 2026, covering data generation, model construction, and usage for step-level reasoning evaluation in large la…
Researchers at Northwestern University and the University of Chicago developed a computational method to measure open codes in qualitative analysis, using LLM-enriched hierarchical clustering and four novel metrics (Cove…
Researchers introduced PARAMΔ, a method that upcycles a dense LLM into a Mixture-of-Experts (MoE) architecture to expand language capabilities without costly alignment. By grafting a MoE-expanded parameter delta onto a C…
Researchers at French institutions found that diversity-driven sampling can reduce pre-training dataset size by up to 94% and training time by 73% while maintaining performance in ModernBERT models. In some tasks, divers…
Researchers at the Association for Computational Linguistics introduced a Dual-Phase Self-Evolution (DPSE) framework for large language models that jointly optimizes user preference adaptation and domain-specific compete…
Researchers at the 64th Annual Meeting of the Association for Computational Linguistics introduced a dual-task paradigm combining arithmetic and sentence comprehension to study how language models allocate cognitive reso…
Researchers introduced DySECT, a dynamic self-evolving extraction system that uses a closed-loop cycle where an LLM extracts triples to populate a knowledge base, which then improves the LLM via prompt tuning or fine-tun…
Researchers at ACL 2026 found that entropy collapse in LLM reasoning, which undermines test-time scaling, is driven by premature overconfidence at a small set of critical tokens. They proposed SCOPE, a method that applie…
Zhipu AI's open-weight GLM 5.2 model scored 39% F1 on IDOR detection in Semgrep's benchmarks, outperforming Claude Code (32%) at $0.17 per vulnerability. The model, released June 13, 2026, with MIT-licensed weights, trai…
The U.S. Commerce Department sent an enforcement letter invoking an export-control directive that barred non-U.S. access to Anthropic's latest AI models, Fable 5 and Mythos 5, leading Anthropic to disable access to compl…
PeekAI, a new open-source tool for local-first observability and debugging of Python AI agents, launched on Hacker News. The tool stores traces in SQLite, requires no cloud accounts or API keys, and supports multi-agent …
AI agents struggle to develop expertise in new domains from text corpora alone, relying on shallow methods like RAG or long-context processing. Researchers propose 'Machine Studying' as the problem of autonomously turnin…