SkyRL now supports FP8-accelerated training and rollout with on-policy weight sync, reducing end-to-end step time by up to 23% while closely tracking BF16 convergence in long-run RL experiments. The s…
Self-improving AI harnesses are real and operational, with systems like Shanghai AI Lab's Self-Harness (arXiv 2606.09498) boosting held-out pass rates on Terminal-Bench-2 from 40.5% to 61.9% on MiniMa…
Researchers propose PI-Mem (Parallel-Iterative Memory), a mechanism that processes all chunks in parallel and iteratively refines a shared memory, enabling long-context reasoning up to 3.6 million tok…
Microsoft Research released Orchard, an open-source framework for training and evaluating AI agents, on August 3, featuring Orchard Env, a Kubernetes-native environment service. The framework reports …
Ramp and Prime Intellect published a case study showing a small RL-trained model, FastAsk, outperformed Claude Opus 4.6 by 4 percentage points on exact-match accuracy for financial spreadsheet retriev…
Researchers have discovered that Mixture-of-Experts (MoE) routing in large language models functions as a form of Huffman coding, governed by a new Frequency-Diversity Law. The study, published on arX…
A new method called UMoE realigns Mixture-of-Experts (MoE) models to enhance domain-specific performance without increasing expert count, parameters, or inference cost. Across Qwen3-30B-A3B and Qwen3.…
UNIBROWSE, a new unified data pipeline for multimodal browsing, achieves an average accuracy of 54.4 on five benchmarks, outperforming its predecessor Qwen3.5-35B-A3B by 10.5 points and surpassing clo…
Alibaba Qwen released Qwen 3.5 on February 13, 2026, with multiple model variants including a 27B parameter version using 64 layers. Benchmarks show strong performance in reasoning, knowledge, math, a…
Researchers introduced Self-Harness, a new paradigm enabling LLM-based agents to iteratively improve their own operating harnesses without human intervention. In tests on Terminal-Bench-2.0, Self-Harn…
A developer reports that the downloadsAllTime field for multiple public models on Hugging Face Hub is decreasing day-over-day, contrary to expectations for a cumulative count. The observed drops range…
A developer created an automated slide-making pipeline using Orxhestra, a local LLM agent framework. The pipeline uses a llama.cpp server with the Qwen3.5-35B-A3B model to plan, implement, screenshot,…
An international research team found that large language models integrated false statements into their outputs even when those claims were explicitly labeled as false in training data. The researchers…