LeRobot v0.6.0 adds world model policies, new VLAs, and unified evaluation
LeRobot v0.6.0 introduces world model policies (VLA-JEPA, FastWAM, LingBot-VA) that learn to imagine future states, along with new VLAs (GR00T N1.7, MolmoAct2, EO-1, EVO1, Multitask DiT). It ships six new simulation benchmarks unified under lerobot-eval, a lerobot-rollout CLI with DAgger-style human-in-the-loop corrections, FSDP training, and cloud training on HF Jobs. Datasets gain depth support, automatic language annotation, custom video encoding, and up to 2x faster data loading.
Entities
Related
UniClawBench benchmark proposed for evaluating proactive AI agents in real-world tasks
Researchers introduced UniClawBench, a universal benchmark for evaluating proactive agents that operate everyday tools in real-world environments. Unlike existing benchmarks that rely on sandboxed settings and single-turn evaluations, UniClawBench aims to isolate specific model capabilities to identify root causes of agent failures.
RightNow AI releases 'Auto' compiler that records LLM agent behavior into verified WASM binaries
RightNow AI introduced Auto, a compiler that records LLM agent behavior, identifies repeatable patterns, and compiles them into verified, sandboxed WebAssembly binaries that run for microdollars. The system includes a tiered runtime that falls back to frontier models for novelty and recompiles results, aiming to eliminate re-solving the same tasks. A paper detailing the approach is available on arXiv.
Community compares local LLMs for agentic workflows using tool-eval-bench
A GitHub user published an interactive comparison report evaluating local LLMs for agentic workflows, using the tool-eval-bench benchmark (84 scenarios, 16 categories, 8 trials). The report targets single DGX Spark or other 96-128GB rigs and covers multi-turn tool orchestration, function calling, and autonomous planning as exercised by Hermes Agent.
vLLM v0.25.0 released with Model Runner V2 as default and PagedAttention removed
vLLM v0.25.0 is now available, featuring 558 commits from 232 contributors. Model Runner V2 becomes the default execution path for all dense models, with new support for EVS, realtime embeddings, prefix caching for Mamba hybrid models, multimodal-prefix bidirectional attention, and dynamic speculative decoding. The legacy PagedAttention implementation has been removed.
LingBot World open-weights game engine released, runs on consumer GPU
LingBot World, an open-weights game engine with WASD movement and hotkeys for storms, fire, and lightning, was released with a variant targeting a single consumer GPU. The city renders frame by frame as the player moves, but areas are not remembered when revisited. Weights are publicly available.
