10 results on this page
Tactile robotics is expanding from sensor hardware into datasets, foundation models, and evaluation benchmarks. Research shows touch improves physical reasoning and robot success, but inconsistent sensor formats create major cross-device transfer challenges.
FinSkillBench is a new benchmark for testing AI agents on portfolio construction, risk management, and fundamental analysis. Across nine models, curated procedural skills raised mean performance from 0.366 to 0.528, while self-generated skills delivered little benefit.
Ornith-1.5 launches open models in 397B, 35B, and 9B parameter sizes. The models feature self-improving task and scaffold generation and report strong performance on coding and reasoning benchmarks.
Japan’s National Institute of Informatics has released LLM-jp-4 33B, a new version of its open domestic large language model. The approximately 33.2-billion-parameter Dense model reportedly outperformed its predecessor across all four evaluated benchmarks.
An independent builder pretrained a 1.02-billion-parameter Kimi-K3 replica on 5 billion decontaminated tokens for roughly $250. The model achieved 33.4% on HellaSwag, exceeding the reported 28% score of GPT-2 124M despite receiving no instruction tuning.
Moonshot and ZAI are reportedly approaching the performance of models from OpenAI and Anthropic. The models can reportedly build passable websites at a cost 75% lower than Claude, highlighting potential advantages in cost-sensitive coding workflows.
Claude is expanding into protein design, adding a scientific research capability to its existing AI portfolio. The update suggests potential use in computational biology and biotech workflows, though the article excerpt does not provide implementation or benchmark details.
The paper proposes Transition Complexity Profile (TCP), a reproducible framework for measuring how difficult game-world transition prediction is at a specified interface. It evaluates branching, interaction-driven uncertainty, opponent influence, and temporal or spatial dependencies to improve comparisons across game-modeling and reinforcement-learning benchmarks.
The community project J-Space Cognition Suite claims to improve DeepSeek V4-Pro-0813 through an inference-time Agent Harness without changing model weights. Its reported gains on several benchmarks have not been independently reproduced, and the project is unrelated to Anthropic's internal J-space interpretability research.
A community benchmark reports up to 218.3 tokens per second for Qwen3.8-27B on two RTX 3090 GPUs using vLLM, INT4 quantization, and DFlash2 speculative decoding. The setup achieved a 131K context ceiling, 168–178 ms time to first token, and peak VRAM usage of 22.3 GB per card.