Selected work

Representative Papers

Rectangular matrix multiplication from shared-leg entropy

Oct 07, 2026

This study addresses the upper bounds on the complexity of rectangular matrix multiplication and the efficiency bottlenecks in all-pairs shortest path (APSP) algorithms. The proposed method extends the OpenAI framework to rectangular multiplication, establishing key bounds such as ω(1,k,1)≤2. By integrating shared-leg entropy inequalities, polynomial degeneration, and tensor spectral duality techniques, it derives a double-exponential parameter α≥1/2, which is subsequently leveraged to optimize Zwick’s algorithm. The primary contributions of this work include novel upper bounds for rectangular multiplication complexity and a theoretical breakthrough achieving an APSP running time of O(n^2.4999).

0 citationsRead paper

BDH-CQ: In-Context Learning with Recurrent Latent Reasoning

Aug 10, 2026

This work addresses the challenge of efficiently solving abstract visual reasoning tasks, such as those in ARC-AGI, without explicitly generating intermediate reasoning steps. The authors propose an implicit iterative reasoning architecture that leverages in-context learning to drive recurrent neural memory, enabling iterative refinement of representations in a high-dimensional latent space to answer queries end-to-end—without requiring language-based intermediate rationales. Evaluated on ARC-AGI-1, their 150M-parameter model achieves a pass@2 rate of 29.5%, with a per-task inference cost of merely $0.0007. This result substantially advances the state-of-the-art by establishing a new efficiency frontier, significantly outperforming prior methods on the cost–accuracy Pareto front.

0 citationsRead paper

The Dragon Hatchling: The Missing Link between the Transformer and Models of the Brain

Sep 30, 2025

How can large language models achieve Transformer-level performance while maintaining biological interpretability and neuroscientific plausibility? This paper introduces Dragon Hatchling (BDH), a brain-inspired architecture that constructs scale-free biological networks from locally interacting neuronal particles, integrating spiking neuron dynamics, Hebbian synaptic plasticity, and attention-guided state-space modeling for GPU-efficient training. Its key contribution lies in unifying three biologically grounded properties—unambiguous neuronal activation, synapse-level dynamic traceability, and modular higher-order graph structure—within a single language modeling framework, thereby ensuring both theoretical rigor and intrinsic interpretability. Evaluated across model sizes from 10M to 1B parameters, BDH matches GPT-2’s performance on language modeling and machine translation tasks using identical training data, demonstrating scalability and biological fidelity. BDH establishes a novel paradigm bridging computational linguistics with cognitive neuroscience.

0 citationsRead paper

Recent publications

Latest Papers

Rectangular matrix multiplication from shared-leg entropy

Oct 07, 2026

This study addresses the upper bounds on the complexity of rectangular matrix multiplication and the efficiency bottlenecks in all-pairs shortest path (APSP) algorithms. The proposed method extends the OpenAI framework to rectangular multiplication, establishing key bounds such as ω(1,k,1)≤2. By integrating shared-leg entropy inequalities, polynomial degeneration, and tensor spectral duality techniques, it derives a double-exponential parameter α≥1/2, which is subsequently leveraged to optimize Zwick’s algorithm. The primary contributions of this work include novel upper bounds for rectangular multiplication complexity and a theoretical breakthrough achieving an APSP running time of O(n^2.4999).

0 citationsRead paper

BDH-CQ: In-Context Learning with Recurrent Latent Reasoning

Aug 10, 2026

This work addresses the challenge of efficiently solving abstract visual reasoning tasks, such as those in ARC-AGI, without explicitly generating intermediate reasoning steps. The authors propose an implicit iterative reasoning architecture that leverages in-context learning to drive recurrent neural memory, enabling iterative refinement of representations in a high-dimensional latent space to answer queries end-to-end—without requiring language-based intermediate rationales. Evaluated on ARC-AGI-1, their 150M-parameter model achieves a pass@2 rate of 29.5%, with a per-task inference cost of merely $0.0007. This result substantially advances the state-of-the-art by establishing a new efficiency frontier, significantly outperforming prior methods on the cost–accuracy Pareto front.

0 citationsRead paper

The Dragon Hatchling: The Missing Link between the Transformer and Models of the Brain

Sep 30, 2025

How can large language models achieve Transformer-level performance while maintaining biological interpretability and neuroscientific plausibility? This paper introduces Dragon Hatchling (BDH), a brain-inspired architecture that constructs scale-free biological networks from locally interacting neuronal particles, integrating spiking neuron dynamics, Hebbian synaptic plasticity, and attention-guided state-space modeling for GPU-efficient training. Its key contribution lies in unifying three biologically grounded properties—unambiguous neuronal activation, synapse-level dynamic traceability, and modular higher-order graph structure—within a single language modeling framework, thereby ensuring both theoretical rigor and intrinsic interpretability. Evaluated across model sizes from 10M to 1B parameters, BDH matches GPT-2’s performance on language modeling and machine translation tasks using identical training data, demonstrating scalability and biological fidelity. BDH establishes a novel paradigm bridging computational linguistics with cognitive neuroscience.

0 citationsRead paper