Selected work

Representative Papers

Digital Red Queen: Adversarial Program Evolution in Core War with LLMs

Jan 06, 2026arXiv.org

This work proposes the Digital Red Queen (DRQ) algorithm, which introduces the Red Queen hypothesis into large language model (LLM)-driven program evolution for the first time. Unlike conventional approaches that rely on static objectives, DRQ simulates the dynamic co-evolutionary arms race observed in nature by embedding LLMs in a Core War environment. There, the model generates assembly-level adversarial programs (warriors) through self-play, with each new generation required to outperform all historical versions, thereby forming a continuously evolving adversarial sequence. This behavior-driven feedback loop enables automatic evolution and evaluation, yielding programs that demonstrate significantly improved generalization against both historical variants and human-designed warriors. Notably, independent runs exhibit convergent evolution, revealing underlying mechanisms of strategic convergence and enhanced generalization in open-ended adversarial settings.

1 citationsRead paper

Fast-weight Product Key Memory

Jan 02, 2026arXiv.org

This work addresses the trade-off between memory capacity and computational efficiency in sequence modeling layers of language models by proposing a dynamic fast-weight Product Key Memory (PKM) mechanism. Building upon the fast weight paradigm, the method transforms the conventional static PKM into a module with dynamic contextual memory capabilities, enabling real-time parameter updates during both training and inference via localized block-wise gradient descent. This facilitates efficient writing and retrieval of new key-value pairs. By integrating sparse PKM architecture with dynamic parameter adaptation, the approach substantially reduces perplexity on long-context tasks. Notably, it achieves strong generalization on a 128K-token “needle-in-a-haystack” retrieval task after training on only 4K tokens, thereby overcoming the limitations inherent in static memory modules.

1 citationsRead paper

The AI Scientist-v2: Workshop-Level Automated Scientific Discovery via Agentic Tree Search

Apr 10, 2025

Achieving a fully autonomous, end-to-end AI-driven scientific research loop—including hypothesis generation, experimental design and execution, data analysis/visualization, and scholarly writing—remains an open challenge due to reliance on manual code templates and limited cross-domain generalization. Method: We propose a progressive agent-tree search framework that eliminates template-based coding and enables generalization across machine learning domains; integrate a vision-language model (VLM)-driven chart review feedback loop to enhance result interpretability and credibility; and unify multi-agent collaboration, automated code generation & execution, and scientific writing & typesetting. Contribution/Results: Three fully AI-generated papers were submitted to ICLR workshops; one was accepted after formal peer review with a score exceeding the average acceptance threshold for human-authored submissions—the first such publication globally validated through rigorous academic peer review.

1 citationsRead paper

CEDAR: Agent-Orchestrated Tree Search for Goal-Directed Optimization of Complex Systems

Aug 07, 2026

This work addresses the challenge of goal-directed design in complex systems, whose emergent behaviors—arising from nonlinear feedback—are notoriously difficult to predict, while existing modeling approaches remain cumbersome and inefficient. To overcome these limitations, the authors propose a novel framework that deeply integrates large language models (LLMs) with Monte Carlo Tree Search (MCTS). In this approach, the LLM serves dual roles: as a structural editor implementing mutation operators and as a behavioral evaluator approximating the fitness function. Operating within a constrained search space composed of executable Python subsets and domain-specific primitives, the method enables efficient and interpretable evolution of system architectures. It substantially reduces manual intervention, automatically discovers structures that fulfill specified behavioral objectives, and uncovers causal mechanisms linking system structure to emergent dynamics, thereby advancing the application of complex systems across diverse domains.

0 citationsRead paper

Recursive Harness Self-Improvement

Jul 16, 2026

This work addresses the challenge of continuously optimizing user-defined reasoning scaffolds at low cost within a framework of co-evolution between models and inference scaffolds, thereby enhancing the quality of agent execution trajectories. It proposes a recursive scaffold self-improvement mechanism that unifies scaffold optimization with model training data generation, representing scaffolds via prompt-level specifications and iteratively refining them through pairwise preference feedback derived from their own revision history. Rather than extending reasoning chains, the method emphasizes task context management to achieve efficient information flow control with minimal inference overhead. Evaluated on 30 cross-domain synthetic tasks, the approach significantly outperforms high-overhead baselines within just a few iterations, reducing inference costs by up to 60%, thus demonstrating the critical role of effective context management in performance improvement.

0 citationsRead paper

Recent publications

Latest Papers

CEDAR: Agent-Orchestrated Tree Search for Goal-Directed Optimization of Complex Systems

Aug 07, 2026

This work addresses the challenge of goal-directed design in complex systems, whose emergent behaviors—arising from nonlinear feedback—are notoriously difficult to predict, while existing modeling approaches remain cumbersome and inefficient. To overcome these limitations, the authors propose a novel framework that deeply integrates large language models (LLMs) with Monte Carlo Tree Search (MCTS). In this approach, the LLM serves dual roles: as a structural editor implementing mutation operators and as a behavioral evaluator approximating the fitness function. Operating within a constrained search space composed of executable Python subsets and domain-specific primitives, the method enables efficient and interpretable evolution of system architectures. It substantially reduces manual intervention, automatically discovers structures that fulfill specified behavioral objectives, and uncovers causal mechanisms linking system structure to emergent dynamics, thereby advancing the application of complex systems across diverse domains.

0 citationsRead paper

Recursive Harness Self-Improvement

Jul 16, 2026

This work addresses the challenge of continuously optimizing user-defined reasoning scaffolds at low cost within a framework of co-evolution between models and inference scaffolds, thereby enhancing the quality of agent execution trajectories. It proposes a recursive scaffold self-improvement mechanism that unifies scaffold optimization with model training data generation, representing scaffolds via prompt-level specifications and iteratively refining them through pairwise preference feedback derived from their own revision history. Rather than extending reasoning chains, the method emphasizes task context management to achieve efficient information flow control with minimal inference overhead. Evaluated on 30 cross-domain synthetic tasks, the approach significantly outperforms high-overhead baselines within just a few iterations, reducing inference costs by up to 60%, thus demonstrating the critical role of effective context management in performance improvement.

0 citationsRead paper

Microcosmos: Reimagining Artificial Life for the GPU Era

Jul 03, 2026

Existing approaches struggle to support large-scale artificial life evolution while preserving physical realism. This work proposes a GPU-optimized, differentiable simulation engine that embeds elastic filament chains within a two-dimensional viscous fluid, uniquely integrating end-to-end differentiable fluid dynamics, neuroevolution, and quality-diversity search for the first time. The resulting framework ensures physical consistency, full differentiability, and linear scalability, enabling efficient generation of diverse swimming and chemotactic behaviors. By combining these capabilities, the system establishes a novel paradigm for conducting large-scale open-ended evolutionary experiments with strong grounding in real-world physics.

0 citationsRead paper

Conversable Complexity: Agentic LLM Collectives as Interpretable Substrates

Jul 01, 2026

This work addresses a central challenge in artificial life research: constructing computational substrates that simultaneously exhibit complexity and interpretability. The paper proposes a novel interpretable substrate based on populations of agentified large language models (LLMs), integrating persistent memory, tool use, skill sharing, and autonomous action. By leveraging natural language interaction, this framework enables direct interrogation and comprehension of emergent collective behaviors. The authors develop a systematic interpretability framework tailored to LLM-agent collectives and validate it through representative case studies spanning controlled experiments to real-world deployments. Their results demonstrate that this paradigm effectively overcomes the inherent opacity of traditional complex systems, offering a promising new pathway toward interpretable artificial life.

0 citationsRead paper

Diffusing Blame: Task-Dependent Credit Assignment in Biologically Plausible Dual-Stream Networks

Jun 30, 2026

This work proposes a biologically plausible learning framework based on a dual-stream excitatory/inhibitory neural network that strictly adheres to Dale’s law—where each neuron’s synapses are exclusively excitatory or inhibitory—and achieves credit assignment without weight transposition through an error diffusion mechanism. The method incorporates modular error routing for multi-class classification, combined with layer-specific Sigmoid widths, batch-centered class-wise error signals, and asymmetric initialization to effectively alleviate credit assignment bottlenecks across diverse tasks. In supervised learning, the model attains 96.7% and 61.7% accuracy on MNIST and CIFAR-10, respectively. Furthermore, the proposed ED-PPO algorithm demonstrates competitive performance against current backpropagation-free baselines in Brax and Craftax reinforcement learning environments, confirming its scalability and generality.

0 citationsRead paper