Selected work

Representative Papers

HAPMoE: Heterogeneity-Aware Automatic Parallelism Planning for Mixture-of-Experts Models Training

Sep 30, 2026

This study addresses the challenge of parallelizing Mixture-of-Experts (MoE) models across heterogeneous clusters, where architectural complexity and hardware disparities are difficult to reconcile simultaneously. We propose the first unified planning framework that jointly incorporates MoE awareness and cluster heterogeneity. The method constructs a lightweight cost model and employs a pruning-enhanced dynamic programming algorithm to efficiently search a six-dimensional parallelism space. It further supports non-uniform pipeline partitioning to accommodate complex hardware environments, generating training schedules directly deployable in Megatron-LM. Experimental results demonstrate up to a 3.2× improvement in training throughput, with non-uniform partitioning contributing an additional 78% gain. The search completes in under one minute, highlighting both the efficiency and practical value of the proposed approach.

0 citationsRead paper

Harness VLA: Steering Frozen VLAs into Reliable Manipulation Primitives via Memory-Guided Agents

Jul 09, 2026

Existing vision-language-action (VLA) models exhibit limited generalization under perturbations such as semantic re-targeting, object re-binding, layout changes, and unstable contacts, while purely analytical primitives struggle with irregular grasps and complex interactions. This work proposes a memory-augmented agent framework that decouples a frozen VLA—used as a retry-capable primitive for contact-intensive manipulation—from a small set of fixed analytical primitives. A memory-guided planner handles non-contact phases and semantic re-localization, invoking the VLA only during localized contact stages. The system learns the applicability boundaries of each primitive through execution trajectories, success heuristics, and failure models, thereby extending the VLA’s capabilities without fine-tuning. The approach outperforms the strongest baseline by 38.6 and 25.4 percentage points on LIBERO-Pro and RoboCasa365, respectively, and achieves a 58.4% success rate on RoboTwin C2R.

0 citationsRead paper

STEAM: Self-Supervised Temporal Ensemble Advantage Modeling for Real-World Robot Learning

Jun 29, 2026

This work addresses the challenge in real-world robot learning where demonstration and interaction data often contain stalls, errors, and suboptimal behaviors, lacking a reliable frame-level signal to discern task progress. The authors propose STEAM, the first method to enable label-free, frame-level advantage modeling by training an ensemble of temporal offset predictors on expert trajectories. Normalized temporal offsets serve as a self-supervised signal, which is transformed into scalar advantage estimates via the ensemble’s predictive distribution, with conservative evaluation achieved through the ensemble minimum. Integrated with CFGRL policy optimization, STEAM significantly improves policy success rates by 59%, 54.3%, 23%, and 16.2% on real-world tasks including bimanual towel folding, chip checkout, cola restocking, and single-arm pick-and-place, effectively identifying and suppressing degenerate behaviors.

0 citationsRead paper

ReMP: Low-Downtime Runtime Model-Parallelism Reconfiguration for LLM Serving

Jun 17, 2026

Current large language model inference systems lack the capability to dynamically adjust model parallelism topologies at runtime, necessitating service restarts under varying workloads and causing multi-minute disruptions, loss of KV cache, and substantial recomputation overhead. This work proposes ReMP, the first framework enabling online elastic reconfiguration of combined tensor and pipeline parallelism. By decoupling topology from execution state, designing a two-dimensional KV cache migration mechanism, and orchestrating an end-to-end reconfiguration pipeline, ReMP reduces topology switching latency to 1–7 seconds across 7B–70B models—orders of magnitude faster than restarting. This significantly improves time-to-first-token (TTFT), time per output token (TPOT), and throughput under dynamic workloads.

0 citationsRead paper

LaWAM: Latent World Action Models for Efficient Dynamics-Aware Robot Policies

Jun 14, 2026

This work addresses the challenge that existing vision-language-action models struggle to explicitly predict scene dynamics induced by actions, while conventional world models rely on computationally expensive pixel-level video generation. To overcome this limitation, the authors propose LaWAM (Latent World Action Model), which for the first time integrates the latent space of pretrained vision foundation models with action-conditioned dynamics modeling. By leveraging lightweight latent dynamic prediction and compact latent visual subgoals to guide policy learning, LaWAM enables efficient and dynamics-aware robotic control. The method achieves state-of-the-art performance on the LIBERO (98.6%) and RoboTwin (91.22%) benchmarks as well as in real-world tasks, with a single action prediction requiring only 187 milliseconds—reducing latency by up to 24× compared to pixel-level world models.

0 citationsRead paper

Recent publications

Latest Papers

HAPMoE: Heterogeneity-Aware Automatic Parallelism Planning for Mixture-of-Experts Models Training

Sep 30, 2026

This study addresses the challenge of parallelizing Mixture-of-Experts (MoE) models across heterogeneous clusters, where architectural complexity and hardware disparities are difficult to reconcile simultaneously. We propose the first unified planning framework that jointly incorporates MoE awareness and cluster heterogeneity. The method constructs a lightweight cost model and employs a pruning-enhanced dynamic programming algorithm to efficiently search a six-dimensional parallelism space. It further supports non-uniform pipeline partitioning to accommodate complex hardware environments, generating training schedules directly deployable in Megatron-LM. Experimental results demonstrate up to a 3.2× improvement in training throughput, with non-uniform partitioning contributing an additional 78% gain. The search completes in under one minute, highlighting both the efficiency and practical value of the proposed approach.

0 citationsRead paper

Harness VLA: Steering Frozen VLAs into Reliable Manipulation Primitives via Memory-Guided Agents

Jul 09, 2026

Existing vision-language-action (VLA) models exhibit limited generalization under perturbations such as semantic re-targeting, object re-binding, layout changes, and unstable contacts, while purely analytical primitives struggle with irregular grasps and complex interactions. This work proposes a memory-augmented agent framework that decouples a frozen VLA—used as a retry-capable primitive for contact-intensive manipulation—from a small set of fixed analytical primitives. A memory-guided planner handles non-contact phases and semantic re-localization, invoking the VLA only during localized contact stages. The system learns the applicability boundaries of each primitive through execution trajectories, success heuristics, and failure models, thereby extending the VLA’s capabilities without fine-tuning. The approach outperforms the strongest baseline by 38.6 and 25.4 percentage points on LIBERO-Pro and RoboCasa365, respectively, and achieves a 58.4% success rate on RoboTwin C2R.

0 citationsRead paper

STEAM: Self-Supervised Temporal Ensemble Advantage Modeling for Real-World Robot Learning

Jun 29, 2026

This work addresses the challenge in real-world robot learning where demonstration and interaction data often contain stalls, errors, and suboptimal behaviors, lacking a reliable frame-level signal to discern task progress. The authors propose STEAM, the first method to enable label-free, frame-level advantage modeling by training an ensemble of temporal offset predictors on expert trajectories. Normalized temporal offsets serve as a self-supervised signal, which is transformed into scalar advantage estimates via the ensemble’s predictive distribution, with conservative evaluation achieved through the ensemble minimum. Integrated with CFGRL policy optimization, STEAM significantly improves policy success rates by 59%, 54.3%, 23%, and 16.2% on real-world tasks including bimanual towel folding, chip checkout, cola restocking, and single-arm pick-and-place, effectively identifying and suppressing degenerate behaviors.

0 citationsRead paper

ReMP: Low-Downtime Runtime Model-Parallelism Reconfiguration for LLM Serving

Jun 17, 2026

Current large language model inference systems lack the capability to dynamically adjust model parallelism topologies at runtime, necessitating service restarts under varying workloads and causing multi-minute disruptions, loss of KV cache, and substantial recomputation overhead. This work proposes ReMP, the first framework enabling online elastic reconfiguration of combined tensor and pipeline parallelism. By decoupling topology from execution state, designing a two-dimensional KV cache migration mechanism, and orchestrating an end-to-end reconfiguration pipeline, ReMP reduces topology switching latency to 1–7 seconds across 7B–70B models—orders of magnitude faster than restarting. This significantly improves time-to-first-token (TTFT), time per output token (TPOT), and throughput under dynamic workloads.

0 citationsRead paper

LaWAM: Latent World Action Models for Efficient Dynamics-Aware Robot Policies

Jun 14, 2026

This work addresses the challenge that existing vision-language-action models struggle to explicitly predict scene dynamics induced by actions, while conventional world models rely on computationally expensive pixel-level video generation. To overcome this limitation, the authors propose LaWAM (Latent World Action Model), which for the first time integrates the latent space of pretrained vision foundation models with action-conditioned dynamics modeling. By leveraging lightweight latent dynamic prediction and compact latent visual subgoals to guide policy learning, LaWAM enables efficient and dynamics-aware robotic control. The method achieves state-of-the-art performance on the LIBERO (98.6%) and RoboTwin (91.22%) benchmarks as well as in real-world tasks, with a single action prediction requiring only 187 milliseconds—reducing latency by up to 24× compared to pixel-level world models.

0 citationsRead paper