Selected work

Representative Papers

Enhancing PIBT via Multi-Action Operations

Nov 12, 2025

Myopic, rule-based MAPF solvers such as PIBT suffer significant performance degradation in orientation-aware scenarios due to their neglect of rotational motion costs. Method: We propose Enhanced PIBT—the first PIBT variant incorporating multi-action rollout, which extends the single-step decision horizon to multiple consecutive actions, thereby alleviating myopia. We further integrate graph-guided pre-filtering of feasible orientations and large-neighborhood search for online path refinement. Contribution/Results: The method maintains millisecond-level response times—enabling real-time coordination of thousands of agents—while achieving state-of-the-art performance on the LMAPF-T benchmark: a 12.7% improvement in path efficiency, a 9.3% increase in task success rate, and substantially enhanced robustness and scalability under orientation constraints.

2 citationsRead paper

GradMem: Learning to Write Context into Memory with Test-Time Gradient Descent

Mar 14, 2026

This work addresses the high memory overhead imposed by KV caching in large language models during long-context reasoning. To mitigate this, the authors propose GradMem, a compressed memory mechanism that optimizes a small set of prefix memory tokens via sample-level gradient descent at test time—while keeping model weights frozen—using a self-supervised context reconstruction loss to iteratively refine memory contents. This approach pioneers the integration of test-time gradient-based optimization into memory encoding, enabling loss-guided error correction and substantially enhancing both memory capacity and generalization. Experiments demonstrate that GradMem outperforms forward-pass memory methods under equivalent memory budgets on associative key-value retrieval tasks and achieves competitive performance on natural language benchmarks such as bAbI and SQuAD variants, successfully answering multi-turn queries using only compressed memory representations.

0 citationsRead paper

Memory Retention Is Not Enough to Master Memory Tasks in Reinforcement Learning

Jan 21, 2026

This study addresses a critical limitation in existing reinforcement learning approaches: their excessive emphasis on memory retention at the expense of dynamically updating or overwriting memory contents in partially observable environments. The work explicitly focuses on memory rewriting capability for the first time, introducing a dedicated benchmark environment to evaluate this ability and systematically comparing recurrent neural networks, Transformers, and structured memory architectures in partially observable reinforcement learning settings. Experimental results demonstrate that classical recurrent models exhibit superior robustness and performance in memory rewriting tasks, whereas structured memory and Transformer-based approaches are only effective under specific conditions and tend to fail in more complex scenarios. These findings reveal fundamental limitations of current mainstream memory mechanisms and argue that future designs must balance stable retention with flexible forgetting.

0 citationsRead paper

KAGE-Bench: Fast Known-Axis Visual Generalization Evaluation for Reinforcement Learning

Jan 20, 2026

This work addresses the limited visual generalization of existing pixel-based reinforcement learning agents under distribution shifts, a challenge exacerbated by the inability of current benchmarks to disentangle and systematically evaluate individual visual factors. To this end, we introduce KAGE-Env and KAGE-Bench, which for the first time enable factorized disentanglement of visual observations by modeling background, illumination, and agent appearance as independent, controllable axes. This framework provides an isolatable and reproducible evaluation setting for visual generalization while preserving the underlying control task. Implemented in JAX, our efficient 2D platform supports training at 33 million environment steps per second on a single GPU. Experiments reveal that background and photometric variations significantly degrade policy performance, whereas appearance changes have minimal impact; notably, relying solely on reward metrics may obscure certain generalization failures.

0 citationsRead paper

HELP: Hierarchical Embodied Language Planner for Household Tasks

Dec 25, 2025

Embodied agents exhibit insufficient planning capability in complex domestic environments due to linguistic ambiguity, environmental dynamism, and skill constraints. Method: This paper proposes a hierarchical language planning architecture that orchestrates multiple lightweight open-source large language models (e.g., Phi-3, Qwen2) to explicitly decouple semantic parsing, environment perception, and skill scheduling—enabling on-device deployment. Leveraging task decomposition, embodied environment interaction interfaces, and state feedback mechanisms, the architecture robustly maps natural-language instructions to executable action sequences. Contribution/Results: Evaluated in real-world home settings, the approach achieves a 42% higher success rate than single-layer LLM baselines on multi-step tasks (e.g., “brew coffee and deliver it to the living room”) and reduces inference latency by 58%. It establishes the first efficient, hierarchical embodied planning paradigm powered by small-parameter models.

0 citationsRead paper

Recent publications

Latest Papers

GradMem: Learning to Write Context into Memory with Test-Time Gradient Descent

Mar 14, 2026

This work addresses the high memory overhead imposed by KV caching in large language models during long-context reasoning. To mitigate this, the authors propose GradMem, a compressed memory mechanism that optimizes a small set of prefix memory tokens via sample-level gradient descent at test time—while keeping model weights frozen—using a self-supervised context reconstruction loss to iteratively refine memory contents. This approach pioneers the integration of test-time gradient-based optimization into memory encoding, enabling loss-guided error correction and substantially enhancing both memory capacity and generalization. Experiments demonstrate that GradMem outperforms forward-pass memory methods under equivalent memory budgets on associative key-value retrieval tasks and achieves competitive performance on natural language benchmarks such as bAbI and SQuAD variants, successfully answering multi-turn queries using only compressed memory representations.

0 citationsRead paper

Memory Retention Is Not Enough to Master Memory Tasks in Reinforcement Learning

Jan 21, 2026

This study addresses a critical limitation in existing reinforcement learning approaches: their excessive emphasis on memory retention at the expense of dynamically updating or overwriting memory contents in partially observable environments. The work explicitly focuses on memory rewriting capability for the first time, introducing a dedicated benchmark environment to evaluate this ability and systematically comparing recurrent neural networks, Transformers, and structured memory architectures in partially observable reinforcement learning settings. Experimental results demonstrate that classical recurrent models exhibit superior robustness and performance in memory rewriting tasks, whereas structured memory and Transformer-based approaches are only effective under specific conditions and tend to fail in more complex scenarios. These findings reveal fundamental limitations of current mainstream memory mechanisms and argue that future designs must balance stable retention with flexible forgetting.

0 citationsRead paper

KAGE-Bench: Fast Known-Axis Visual Generalization Evaluation for Reinforcement Learning

Jan 20, 2026

This work addresses the limited visual generalization of existing pixel-based reinforcement learning agents under distribution shifts, a challenge exacerbated by the inability of current benchmarks to disentangle and systematically evaluate individual visual factors. To this end, we introduce KAGE-Env and KAGE-Bench, which for the first time enable factorized disentanglement of visual observations by modeling background, illumination, and agent appearance as independent, controllable axes. This framework provides an isolatable and reproducible evaluation setting for visual generalization while preserving the underlying control task. Implemented in JAX, our efficient 2D platform supports training at 33 million environment steps per second on a single GPU. Experiments reveal that background and photometric variations significantly degrade policy performance, whereas appearance changes have minimal impact; notably, relying solely on reward metrics may obscure certain generalization failures.

0 citationsRead paper

HELP: Hierarchical Embodied Language Planner for Household Tasks

Dec 25, 2025

Embodied agents exhibit insufficient planning capability in complex domestic environments due to linguistic ambiguity, environmental dynamism, and skill constraints. Method: This paper proposes a hierarchical language planning architecture that orchestrates multiple lightweight open-source large language models (e.g., Phi-3, Qwen2) to explicitly decouple semantic parsing, environment perception, and skill scheduling—enabling on-device deployment. Leveraging task decomposition, embodied environment interaction interfaces, and state feedback mechanisms, the architecture robustly maps natural-language instructions to executable action sequences. Contribution/Results: Evaluated in real-world home settings, the approach achieves a 42% higher success rate than single-layer LLM baselines on multi-step tasks (e.g., “brew coffee and deliver it to the living room”) and reduces inference latency by 58%. It establishes the first efficient, hierarchical embodied planning paradigm powered by small-parameter models.

0 citationsRead paper

LookPlanGraph: Embodied Instruction Following Method with VLM Graph Augmentation

Dec 24, 2025

Static scene graphs fail to adapt to environmental dynamics in embodied instruction following. Method: This paper proposes a dynamic scene graph modeling approach based on an online-evolving vision-language model (VLM), which parses the agent’s egocentric visual input in real time and closed-loop integrates static assets with object priors to jointly update perception, planning, and execution. Contributions/Results: (1) We introduce the first online-evolving VLM-augmented graph structure, enabling both prior validation and novel entity discovery; (2) We construct GraSIF—the first Graph Scenes for Instruction Following dataset—equipped with an automated verification framework. Evaluated on VirtualHome, OmniGibson, and a real robot platform, our method achieves a 23.6% higher task success rate than static-graph baselines under object displacement perturbations, demonstrating significantly improved robustness and generalization.

0 citationsRead paper