13 articles — 🟡 8 important , 🟢 5 interesting
Ollama 0.30 brings integration with llama.cpp for better performance and GGUF model compatibility, with up to 20% faster throughput on NVIDIA GPUs. It expands hardware support with Vulkan on AMD and Intel devices and adds support for tool-calling. It complements the existing MLX engine for Apple silicon.
The vLLM team has released v0.3 'Themis' of its Semantic Router, the first production-ready version for routing queries between models. It brings canonical configuration, an inspectable decision flow, and reproducible routing behavior for Kubernetes deployments.
Engram is an open-source memory system that shows that smartly retrieved 'lean' context outperforms loading the entire conversation history. On the LongMemEval_S benchmark it achieved 83.6% versus 73.2% for full-context, using about 8× fewer tokens.
SWE-Marathon is a new benchmark for evaluating agents on ultra-long-horizon software engineering tasks. Frontier coding agents solve less than 30% of the 20 tasks, with reward-hacking behavior in 13.8% of rollouts. The benchmark, eval code, and trajectories have been made public.
Google Research and Google Cloud have introduced a multi-agent RAG framework with a 'Sufficient Context Agent' that assesses whether the retrieved context is sufficient. On the FramesQA benchmark it achieved 90.1% accuracy in cross-corpus scenarios and up to 34% better factuality with minimal latency increase.
LangChain has announced the general availability of LangSmith Sandboxes — isolated, hardware-virtualized microVMs that give an agent its own file system, shell, and persistent state. Capabilities include snapshots, forks, and blueprints. It integrates with the existing LangSmith SDK and API.
Anthropic has published a white paper on Claude's capabilities in chemistry, particularly NMR spectroscopy. Claude Opus 4.7 achieved an average error of ±0.079 ppm in hydrogen prediction and demonstrated the ability to reason bidirectionally between spectrum and structure.
A new paper addresses pre-training language models when compute grows faster than available text data. It introduces masked-input regularization (MIR) and the SoftQ scaling law, estimating that MIR's gains correspond to about 1.3 times more unique data. The code has been released.
Instead of generating criteria per prompt, the Eval-Skill method synthesizes reusable domain evaluation skills through exploration. On the RewardBench 2 benchmark it delivers +13.44% for Qwen3-8B and +18.51% for DeepSeek-V4-Flash over baseline scoring. The code is open source.
The compact multilingual reasoning benchmark mmPISA-bench is derived from OECD PISA testing and covers 43 languages, for a total of 2,150 data points. Modern LLMs reason effectively across all languages, and machine translations perform comparably to human ones. Certain languages simultaneously show higher costs and lower accuracy.
