Large language model (LLM) news — GPT-4, Claude, Gemini, Llama, Mistral and the latest research on training, fine-tuning, RLHF, and deployment of LLMs.
Researchers introduced Indi-RomCoM, a benchmark for evaluating large language models on Romanized code-mixed Indic-English instructions. Testing across seven tasks and four languages showed that LLMs consistently underpe…
Researchers introduced GradeSQL, a framework for training Outcome Reward Models (ORMs) to improve test-time verification in Text-to-SQL tasks. ORM-based selection outperformed execution-based Best-of-N and Majority Votin…
Researchers deployed an automated pipeline to optimize skill descriptions for an enterprise AI agent, achieving 79.2% F1 accuracy versus 79.4% for manual tuning while reducing engineering effort per skill from 120 to 3.8…
Researchers at arXiv trained GPT-2 models on 'impossible' languages and found that while grammatical sensitivity degrades gradually, generative production fails significantly, suggesting transmission failures explain why…
Researchers submitted a system to SemEval-2026 Task 9 for detecting multilingual online polarization using transformer-based models with class weighting and threshold tuning. Their approach achieved competitive F1 macro …
Researchers extend Frictive Policy Optimization (FPO) to handle perceptual asymmetry in dialogue, where participants have asymmetric partial information. Cross-corpora analysis and LLM probing on the HCRC MapTask show th…
Researchers introduced Explanation Quality Markers (EQMs), a set of sixty reasoning patterns scored by large language models, to measure judgment quality in natural-language explanations. In a pre-registered analysis of …
Researchers at arXiv identify a failure mode in large language models called deductive stereotyping, where models apply population-level statistics to individuals, producing biased inferences. They propose Fair-GCG, a re…
Researchers propose CORTEX, a token-level hallucination detection method for Retrieval-Augmented Generation (RAG) that identifies ungrounded content by comparing internal representations of a large language model with an…
Researchers introduced LoFa, a benchmark for evaluating large language models' robustness against logical fallacies, using a multi-agent pipeline and a multi-round debate framework. Experiments revealed varying vulnerabi…
Researchers introduced a semantic-layer-mediated NL2SQL agent that decouples semantic intent from physical SQL execution, achieving 94.15% execution accuracy on the Spider2-snow benchmark and ranking third on the officia…
Researchers introduced a controlled student-teacher protocol to study when natural-language feedback improves multi-turn language agent performance beyond repeated attempts. Testing thirteen open-weight models across fou…
Researchers introduced BayesBench, a suite of simulation environments to evaluate how large language models update beliefs under multi-turn evidence accumulation. Testing seven LLMs from 3B to 70B parameters, they found …
Researchers introduced LearnStop, a cost-aware early-exit method for reasoning language models that learns when to stop computation based on online features like answer confidence and entropy. Across 18 task-model settin…
Researchers introduced multi-agent deliberation frameworks inspired by courtroom procedures for legal reasoning tasks using large language models. The multi-agent approaches achieved comparable overall performance to bas…
Researchers introduced HASTE, a hierarchical multi-agent system that organizes cross-competition knowledge into three scope tiers to improve transfer efficiency in ML engineering. In tests on the MLE-Bench Lite benchmark…
Researchers introduced a three-phase deep reinforcement learning system for personalized portfolio management that overcomes ticker lock-in, monolithic objectives, and static user models. Phase 1 uses a self-supervised c…
A new study on natural-language-to-Lean formalization reveals that compile-rate evaluation overestimates faithfulness, with a tool-augmented agent achieving 89.5% compilation but only 60.5% consensus faithfulness. The re…
Researchers introduced MultiUAV-Plat, a simulation platform and benchmark for evaluating large language models in multi-UAV collaborative task planning. Their proposed agent framework, Agent4Drone, achieved a 57.9% task …
The US Department of Commerce lifted export controls on Anthropic's Claude Fable 5 and Mythos 5 AI models, restoring foreign access starting July 2. Acting Commerce Secretary Howard Lutnick stated the move strengthens Am…