VLCP: Vision Language Control Policy Closed-Loop Code Replanning for Robot Manipulation
本文提出VLCP方法，通过冻结视觉语言模型并周期性重写控制代码来解决机器人操作中开环控制的问题，提高任务成功率。
0 citationsRead paper
Selected work
Representative Papers
Backpropagation as a Nilpotent Linear System
This work transcends the conventional computational graph perspective to uncover the global mathematical structure underlying backpropagation. It unifies the layer-wise backward recurrence of an L-layer feedforward network into a linear system $(I - \mathcal{B})X = G$, where $\mathcal{B}$ is a strictly block upper-triangular and nilpotent global backward operator. Leveraging the proposed F-adjoint framework and the notion of F-symmetry, the study establishes that backpropagation is equivalent to block back-substitution in an upper bidiagonal system and rigorously derives the gradient mechanisms for residual networks and transfer learning. By employing operator theory, nilpotent matrix analysis, and Neumann series expansions, the work precisely characterizes the single-path gradient collapse phenomenon, elucidates its failure mode in residual architectures, and achieves a unified theoretical understanding of gradient highways and gradient clipping.
0 citationsRead paper
MoxE: Mixture of xLSTM Experts with Entropy-Aware Routing for Efficient Language Modeling
To address scalability and inference efficiency bottlenecks in large language models (LLMs), this paper proposes MoxE—a novel architecture integrating extended LSTM (xLSTM) with sparse Mixture of Experts (MoE). Its core innovation is an entropy-aware dynamic routing mechanism that adaptively assigns tokens to specialized experts based on token-distribution entropy, enabling balanced modeling of both high- and low-frequency tokens. To improve rare-token representation, we introduce mLSTM, a modified LSTM variant. Additionally, we design entropy regularization and group-balanced auxiliary losses to enhance expert utilization fairness and generalization. Experiments demonstrate that, at equal parameter count, MoxE achieves a 37% speedup in inference latency, reduces memory footprint by 42%, and significantly lowers language modeling perplexity compared to strong baselines.
0 citationsRead paper
Distil-xLSTM: Learning Attention Mechanisms through Recurrent Structures
This work addresses the challenge of knowledge distillation from Transformer-based large language models (LLMs) to recurrent architectures (e.g., xLSTM, Mamba), stemming from their structural heterogeneity with self-attention mechanisms. We propose Distil-xLSTM—the first compact language model built entirely on a recurrent architecture that explicitly models and distills the parameterization of Transformer attention. Our method leverages xLSTM’s sequential mixing capability to learn a compact representation of attention behavior from an LLM via knowledge distillation, eliminating the conventional requirement for architectural homogeneity. Key contributions include the first interpretable, parameterized approximation of attention mechanisms by a purely recurrent model. Experiments demonstrate that Distil-xLSTM achieves performance competitive with larger models under minimal training overhead, while exhibiting superior computational efficiency and scalability compared to attention-based counterparts—validating the feasibility and effectiveness of recurrent models in learning attention representations.
0 citationsRead paper
Recent publications
Latest Papers
VLCP: Vision Language Control Policy Closed-Loop Code Replanning for Robot Manipulation
本文提出VLCP方法，通过冻结视觉语言模型并周期性重写控制代码来解决机器人操作中开环控制的问题，提高任务成功率。
0 citationsRead paper
Backpropagation as a Nilpotent Linear System
This work transcends the conventional computational graph perspective to uncover the global mathematical structure underlying backpropagation. It unifies the layer-wise backward recurrence of an L-layer feedforward network into a linear system $(I - \mathcal{B})X = G$, where $\mathcal{B}$ is a strictly block upper-triangular and nilpotent global backward operator. Leveraging the proposed F-adjoint framework and the notion of F-symmetry, the study establishes that backpropagation is equivalent to block back-substitution in an upper bidiagonal system and rigorously derives the gradient mechanisms for residual networks and transfer learning. By employing operator theory, nilpotent matrix analysis, and Neumann series expansions, the work precisely characterizes the single-path gradient collapse phenomenon, elucidates its failure mode in residual architectures, and achieves a unified theoretical understanding of gradient highways and gradient clipping.
0 citationsRead paper
MoxE: Mixture of xLSTM Experts with Entropy-Aware Routing for Efficient Language Modeling
To address scalability and inference efficiency bottlenecks in large language models (LLMs), this paper proposes MoxE—a novel architecture integrating extended LSTM (xLSTM) with sparse Mixture of Experts (MoE). Its core innovation is an entropy-aware dynamic routing mechanism that adaptively assigns tokens to specialized experts based on token-distribution entropy, enabling balanced modeling of both high- and low-frequency tokens. To improve rare-token representation, we introduce mLSTM, a modified LSTM variant. Additionally, we design entropy regularization and group-balanced auxiliary losses to enhance expert utilization fairness and generalization. Experiments demonstrate that, at equal parameter count, MoxE achieves a 37% speedup in inference latency, reduces memory footprint by 42%, and significantly lowers language modeling perplexity compared to strong baselines.
0 citationsRead paper
Distil-xLSTM: Learning Attention Mechanisms through Recurrent Structures
This work addresses the challenge of knowledge distillation from Transformer-based large language models (LLMs) to recurrent architectures (e.g., xLSTM, Mamba), stemming from their structural heterogeneity with self-attention mechanisms. We propose Distil-xLSTM—the first compact language model built entirely on a recurrent architecture that explicitly models and distills the parameterization of Transformer attention. Our method leverages xLSTM’s sequential mixing capability to learn a compact representation of attention behavior from an LLM via knowledge distillation, eliminating the conventional requirement for architectural homogeneity. Key contributions include the first interpretable, parameterized approximation of attention mechanisms by a purely recurrent model. Experiments demonstrate that Distil-xLSTM achieves performance competitive with larger models under minimal training overhead, while exhibiting superior computational efficiency and scalability compared to attention-based counterparts—validating the feasibility and effectiveness of recurrent models in learning attention representations.
0 citationsRead paper