Selected work

Representative Papers

Depth-Wise Emergence of Prediction-Centric Geometry in Large Language Models

Feb 04, 2026

This work investigates how decoder-only large language models transform contextual information into predictive outputs along the depth dimension. By integrating geometric representation analysis with mechanistic interventions, the study reveals—for the first time—that angular components of deep-layer representations encode similarity structures aligned with predictive distributions, while norm components carry non-predictive contextual information. This finding establishes a mechanism-geometric account of the context-to-prediction transformation process. Building upon disentangled representations and an intervenable modeling framework, the research identifies structured geometric properties underlying prediction formation and demonstrates causal, selective control over token-level predictions.

0 citationsRead paper

Redundancy Maximization as a Principle of Associative Memory Learning

Nov 04, 2025

Classical Hopfield networks exhibit associative memory capabilities, yet their local information-processing mechanisms remain poorly understood. Method: This paper introduces Partial Information Decomposition (PID) theory into associative memory modeling for the first time, proposing “redundancy maximization” as a principled learning objective. It establishes an information-theoretic, neuron-level learning rule by directly optimizing shared redundant information among inputs, enabling interpretable control over individual neuron contributions. Contribution/Results: The resulting model achieves a memory capacity of 1.59, exceeding that of the classical Hopfield network by over one order of magnitude and outperforming state-of-the-art variants. These results empirically validate redundancy maximization as a fundamental, generalizable learning principle for associative memory systems.

0 citationsRead paper

World Models as Reference Trajectories for Rapid Motor Adaptation

May 21, 2025

To address the abrupt degradation of control performance caused by sudden system dynamics changes, this paper proposes a reflective world model (RWM) dual-control framework. RWM directly utilizes the deterministic world model’s predictive output as an implicit, dynamically adaptive reference trajectory, thereby decoupling long-horizon reward optimization from millisecond-level latent-space motion execution. It introduces, for the first time, a synergistic dual-control architecture integrating latent-space model predictive control, online latent-action correction, and hierarchical reinforcement learning. Evaluated on high-dimensional continuous control tasks, RWM achieves second-scale dynamics adaptation, reduces computational overhead by 87%, and maintains policy performance above 98% of the optimal—significantly outperforming existing model-based baselines.

0 citationsRead paper

Recent publications

Latest Papers

Depth-Wise Emergence of Prediction-Centric Geometry in Large Language Models

Feb 04, 2026

This work investigates how decoder-only large language models transform contextual information into predictive outputs along the depth dimension. By integrating geometric representation analysis with mechanistic interventions, the study reveals—for the first time—that angular components of deep-layer representations encode similarity structures aligned with predictive distributions, while norm components carry non-predictive contextual information. This finding establishes a mechanism-geometric account of the context-to-prediction transformation process. Building upon disentangled representations and an intervenable modeling framework, the research identifies structured geometric properties underlying prediction formation and demonstrates causal, selective control over token-level predictions.

0 citationsRead paper

Redundancy Maximization as a Principle of Associative Memory Learning

Nov 04, 2025

Classical Hopfield networks exhibit associative memory capabilities, yet their local information-processing mechanisms remain poorly understood. Method: This paper introduces Partial Information Decomposition (PID) theory into associative memory modeling for the first time, proposing “redundancy maximization” as a principled learning objective. It establishes an information-theoretic, neuron-level learning rule by directly optimizing shared redundant information among inputs, enabling interpretable control over individual neuron contributions. Contribution/Results: The resulting model achieves a memory capacity of 1.59, exceeding that of the classical Hopfield network by over one order of magnitude and outperforming state-of-the-art variants. These results empirically validate redundancy maximization as a fundamental, generalizable learning principle for associative memory systems.

0 citationsRead paper

World Models as Reference Trajectories for Rapid Motor Adaptation

May 21, 2025

To address the abrupt degradation of control performance caused by sudden system dynamics changes, this paper proposes a reflective world model (RWM) dual-control framework. RWM directly utilizes the deterministic world model’s predictive output as an implicit, dynamically adaptive reference trajectory, thereby decoupling long-horizon reward optimization from millisecond-level latent-space motion execution. It introduces, for the first time, a synergistic dual-control architecture integrating latent-space model predictive control, online latent-action correction, and hierarchical reinforcement learning. Evaluated on high-dimensional continuous control tasks, RWM achieves second-scale dynamics adaptation, reduces computational overhead by 87%, and maintains policy performance above 98% of the optimal—significantly outperforming existing model-based baselines.

0 citationsRead paper