Multimodal real-robot dataset of 28,993 teleoperated bimanual manipulation demonstrations (~300 hours) with synchronized multi-view RGB/tactile videos, 65D robot state/action, 60D fingertip tactile signals, and temporally aligned language segments; intended for imitation, visuomotor and visual‑tactile research, CC BY‑NC 4.0 (non‑commercial).
Defines and evaluates AREX-2, an LLM agent that iteratively self-improves at test time via reflection and long-horizon execution. Trained on long-horizon improvement trajectories from ML engineering and algorithmic programming (built on Qwen3.8-27B), it scales with more rounds and achieves strong benchmark scores.
Demonstrates that pretrained transformers typically use only ~1–3 lines of depth to follow reference chains, and that a task‑trained rank‑8 LoRA applied at one early layer (with all other weights frozen) can extend reference‑following to dozens or hundreds of lines while adding only a few ten‑thousand parameters.
Groups visually grounded appearances of the same physical instance into persistent, retrievable “biographies” so agents can follow objects across hours or days for long-video question answering. Links identity-aware observations to episodic context and visual evidence; improves EgoLifeQA to 72.0% and increases evidence-window reach from 37.6% to 58.9%.
Predicts compact 'prospective tokens' that summarize upcoming information needs and uses them to select a small set of past frames for conditioning long-horizon video generation, improving long-range consistency, visual quality, and action alignment while remaining plug-and-play across diverse generators.
Constructs and continually maintains explicit belief states for long-horizon LLM agents, combining a structured world estimate with unresolved epistemic and achievement gaps. Adds consistency validation, Belief Trapping detection, and tailored recovery to improve execution and diagnosis benchmarks.
Processes English and German text with long-context reasoning and structured tool-calling. Uses a 78B mixture-of-experts architecture that activates ~3.46B parameters per token, offers native 262k-token context (validated to 1M), and is released as Apache-2.0 weights — suited for RAG, document processing and human-in-the-loop decision support.
Introduces LoHi, a training-free, single-pass method that mixes dense low-resolution video streams with sparse high-resolution frames to improve long-video vision-language model accuracy under strict token budgets while cutting front-end decoding latency.
Adapts LLM agents online by self-distilling verified execution trajectories into persistent LoRA weights during deployment to improve success and efficiency on long‑horizon tasks. Uses a frozen stable copy as a privileged teacher to predict hindsight next‑token distributions and filters invalid-action turns so experience consolidates without external solutions or memory retrieval.
Converts historical interaction traces into a reusable, queryable “worldbook” and runs a language-based world model agent (Trace2Env) as the environment for LLM agents — enabling stateful, grounded simulation with improved next-observation fidelity and long-horizon consistency.
Runs an open-source personal software agent across a user's devices to perform actions, keep readable provenance-backed memories, and coordinate via a self-hostable relay. Ships minimal end-to-end clients (mobile, desktop, web) under GPL-3.0; model-agnostic and designed for inspectability and device
Proposes TRACE, an FP4 quantization framework for RL of MoE LLMs that uses rollout-side FP4 outcomes to guide training-side rounding and caches mantissa/scale from deeper layers to limit overhead. Preserves BF16-level RL performance while enabling up to 5.4× rollout speedup.