After each batch, an optimizer proposes an update from scored trajectories. Instead of relying on a fixed validation set, the update is tested on new and past cases and kept only if it helps on new ...
Join the discussion on this paper page Beyond Spatio-Temporal Priors: A Generalizable Approach for Dense Correspondence Matching ...
Join the discussion on this paper page MiMo-V2.6: Scaling Reinforcement Learning Towards Self-Improvement ...
Join the discussion on this paper page Remory appends learned soft memory tokens to a compacted summary. On Qwen3.8-27B and GLM-5.3-Flash, it improves long-horizon benchmark performance while reducing ...
Join the discussion on this paper page TestPrism: Rethinking Test Evaluation Beyond a Single Reference ...
Join the discussion on this paper page Learn2Play Bench: How Well Do LLM Agents Learn from Experience in Unfamiliar Environments?
Join the discussion on this paper page DreamTrue: Action-Faithful Robot World Model with Counterfactual Post-Training ...
We tackles a key ambiguity in robot in-context learning: what should a robot actually follow from a video demonstration? We defines the learning target across action, object, composition, and ...
Join the discussion on this paper page SuperNav: An Agentic Navigation System for Any Task in Any Scene ...
Join the discussion on this paper page ...
Join the discussion on this paper page DSReg: Provably Recovering Individual World Latents without Reconstruction ...
Join the discussion on this paper page Simulated users offer a scalable alternative to costly human feedback, but they must both resemble real user behavior and provide useful learning experiences for ...
Some results have been hidden because they may be inaccessible to you
Show inaccessible results