Introduces an open-source multi-agent harness that automatically constructs, composes, and evolves modular model–harness units for long‑horizon, cross‑domain workflows. Key ingredients include a Host Agent for task decomposition and orchestration, EverOS for durable memory, and a Skill Forge of reusable procedures to improve task coverage via composition.
Learns dynamics-aligned latent representations for parametric PDE forecasting using masked-latent predictive pretraining, a geometry projector that aligns latent trajectory geometry, and a physics-structured latent predictor that separates shared evolution from parameter-dependent responses.
Investigates how rollout policy, token-level KL direction, and learning rate each affect LLM distillation across Llama3 and Qwen2.5 on reasoning tasks; finds KL direction and learning rate dominate outcomes while rollout policy has a modest effect.
Uses 360° panoramic observations to improve vision-and-language navigation by predicting longer action sequences, confidence-guided execution, and combined semantic–geometric panorama features; yields large SR gains on R2R-CE and RxR-CE Val-Unseen.
Records structural priors with skill-specific policies so a runtime agent can select and compose the version of each skill best suited to new states, improving out-of-distribution and compositional generalization for robot manipulation from few demonstrations.
Teaches small reasoning models to diagnose when additional internal thinking is insufficient and to selectively query stronger models; introduces FlyBy, a supervised + cost-aware RL framework that learns when/what to ask. Improves pass@ metrics on hard benchmarks while reducing serving cost.
Normalizes each domain's teacher feedback spread during multi-teacher on-policy distillation so that no domain (e.g., instruction following) overwhelms others, improving student recovery of specialist skills and raising average scores across benchmarks.
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.
Proposes a forward-process RL method that dynamically localizes gradient updates and adapts multi-reward coordination for joint audio–video diffusion models. Uses bidirectional cross-attention responses for token/layer routing and preference-preserving reweighting to balance competing objectives, improving modality quality, alignment, and synchronization.
Compresses recurrent states in linear-attention LLMs via spatial–temporal post-training quantization, allocating bits by error lifetime and per-row impact to preserve accuracy (6-bit ≈ FP32) while cutting serving memory up to 68.7%.
Extrapolates RL-induced representation residuals into student hidden states during on-policy distillation: at each layer and token, it regresses the student beyond the teacher along the teacher’s RL-induced direction to improve stability and empirical performance.
Co-evolves candidate solutions and web-search queries to help LLM-driven evolutionary discovery, using a retrieval gate plus bilevel inner/outer loops that refine queries, rank documents by predicted solution value, and generate evaluated candidates.