Proposes ZPPO, a distillation method that keeps the teacher inside prompts rather than injecting teacher gradients, using binary- and negative-candidate prompts plus a prompt replay buffer to recover learning signal on hard examples; shows gains for small Qwen3.5 students across 31 multimodal benchmarks.
Provides a dual-path approach for spatial vision-language models: a Language-Only Reasoning (LOR) path for stepwise linguistic deduction and a Detect-Then-Reason (DTR) path that detects 3D cues via region tokens before numerical inference. Trains with chain-of-thought cold-start supervision and reinforcement learning to improve 3D grounding and multi-step spatial reasoning.
Proposes Monotonic Inference Policy Improvement (MIPI) and a two-step Monotonic Inference Policy Update (MIPU) to address training–inference probability mismatch in LLM reinforcement learning by constructing sampler-referenced candidate updates and accepting synchronized updates using an inference-gap proxy; shows improved reasoning accuracy and stability under FP8-quantized rollouts.
Adapts pretrained Vision-Language-Action (VLA) models to new camera poses and robot embodiments from a single demonstration by performing weight-vector arithmetic that injects domain-specific information. Filters noise via subspace alignment of singular components; designed for one-shot adaptation under visual and embodiment shifts.
Zhilin Wang, Han Song +14University of Science and Technology of China, The Chinese University of Hong Kong +6
Provides a benchmark and protocol to evaluate agents that iteratively edit executable policies under a fixed interaction budget, recording full execution–feedback–revise trajectories. Built from compact RL environments with trajectory-level diagnostics and hidden held-out validation.
Embodied AI2026
Provides a systematic benchmark and design roadmap for video-based world models to evaluate robot policies, introducing WMBench and GigaWorld-1 optimized for long-horizon, action-faithful rollouts. Offers controlled comparisons across model families, action encodings, and 324k+ simulated vs real rollouts, with code, models, and datasets released for reproducible evaluation.
Embodied AI2026
Detects when an action-chunked VLA policy drifts from expected visual dynamics and triggers lightweight corrective replanning via a latent-space vision monitor and online gradient guidance; creates an event-driven adaptive action horizon without retraining the backbone.
AI Agent Papers2026
Trains cross-platform GUI agents by combining a Uni-GUI cross-platform dataset with platform-conditioned multi-teacher on-policy distillation, enabling a shared policy to adapt to new platforms while retaining platform-specific behaviors; suitable for research on continual GUI agent learning and cross-platform adaptation.
Stabilizes on-policy policy distillation by dynamically constructing a proximal teacher that controls gradient variance. Provides theoretical global convergence and monotonic improvement bounds, and shows improved training stability, sample efficiency, and final performance on mathematical reasoning tasks with zero extra compute overhead.
Transfers RL-induced policy shifts from a smaller 'weak' teacher to a stronger target by using the teacher's post-/pre-RL log-ratio as a dense implicit reward applied on the student's on-policy states. Enables reuse of RL supervision without running RL rollouts on the target, improving sample/time efficiency.
Reconstructs historical experience into latent memory tokens and weaves short- and long-term latent memories directly into vision-language-action reasoning to improve long-horizon robotic manipulation. Uses a four-part pipeline (curator, seeker, condenser, weaver) so memory participates natively in multimodal action formation.
AI Agent Papers2026
Provides a terminal-style benchmark of 46 long-horizon tasks decomposed into fine-grained graded subtasks to produce dense intermediate rewards and partial credit, enabling evaluation of long-horizon planning, long-context management, and iterative debugging. Tasks typically require hundreds of episodes and minutes-to-hours of execution; baseline evaluations report high token and episode consumption with low pass rates, highlighting evaluation headroom.