Axion 1.0 simplifies Open Agent Foundation Models.
Axion 1.0 simplifies open agent models with 18,000 examples, lightweight stages, and reinforcement learning for scientific reasoning and tool use.
Axion 1.0 simplifies open agent models with 18,000 examples, lightweight stages, and reinforcement learning for scientific reasoning and tool use.
Recursa helps AI models learn new fields by generating tasks, practicing, and storing self-derived skills, improving Harvey LAB pass rates.
PrimeSim is a multiview robot world model trained on 30,110 hours across 22 platforms. It runs fast, leads EWMBench, and supports policy training.
Qwen 3.8 27 billion read model weights locally, but critics say weight-space programming claims depend on tools, code, and tiny Atari networks.
StateBench tests long-context LLM state tracking across four difficulty levels, revealing where architectures fail and aiding model selection.
Palette trains a small open model for design-language decks, splits editorial and coding roles, adds a repair harness, and tops PresentBench domains.
Collinear AI's CWE-bench v2 broadens coverage with 100 public tasks, programmatic scoring, a held-out set, and AI security results.
Embodied Agent Arena tests ten vision-language model agents, showing GeoProbe exposes failures in spatial planning, control, and completion.
Multi-teacher adaptive on-policy self-distillation lets each LLM token choose a teacher, improving Qwen3 math reasoning by up to 5.9 over base model.
Terminal task benchmarks often measure turns, not reasoning depth. DEPTH generates tasks that expose current AI model limits.
Velocity Adversarial Distillation reframes few-step diffusion and flow models as velocity-field control, reducing tracking errors with a learned critic.
NVIDIA adds four open-weight Cosmos 3 Nano checkpoints for driving, robotics, interactive streaming simulation, and 16B post-training workflows.
WeLM-Image unifies understanding, generation, editing and multi-reference synthesis via a shared sequence for single-GPU apps.
This release uses sparse autoencoders to show Qwen3 MoE experts have private, fine-grained specializations, releasing code and an expert explorer.
NPCBench tests multimodal AI on patient care trajectories, not isolated questions. Results show a composition gap; strict CarePath success is 23%.
Study tests PLatour for prefill transfer, finding gains may reflect bridge capacity, with task-dependent evidence for useful sender information.
SynthFaces is a synthetic face dataset accepted to the British Machine Vision Conference 2026. It uses ID-Flow and an identity-conditioned generator to create new identities balanced across age, race …
DiffusionGemma is Google's experimental open LLM that denoises text blocks in parallel, easing GPU bottlenecks for local Transformers and vLLM.
Google Gemini 4 Carbon is a frontier AI model designed for long-horizon work and cybersecurity defense. It ranks first on Artificial Analysis Intelligence Index v5.0 with a score of 59. The model perf…
TB-MathModel tests AI agents on real mathematical modeling via terminal tasks, formal checks, and contest data, revealing capability and cost gaps.