Claude Skills by hiyenwong

Framework for understanding when architectural modularity matters in continual learning based on representational dimensionality. Shows that modular networks only outperform monolithic ones in low-dimensional regimes where representational geometry is constrained. Triggers: continual learning dimensionality, modular vs monolithic networks, representational geometry, stability-plasticity tradeoff, structure matters continual learning.

Discounted MPC under plant-model mismatch - stability and suboptimality analysis for infinite-horizon optimal control with surrogate models. Activation: MPC, model predictive control, plant-model mismatch, robustness, stability, discounted control.

Discrete signaling mediates chaotic regularization in recurrent neural networks - linking microscopic chaos to macroscopic neural representation geometry

Product-of-Experts (PoE) defense against adaptive distillation attacks — a minimax game framework between a utility-constrained teacher and an adaptive student that reweights high-value examples. PoE is a simple forward-pass-only defense combining teacher with proxy student during generation (arXiv: 2605.22737).

Distributed AI agent orchestration methodology for large-scale multi-agent systems. Covers architecture patterns for orchestrated multi-agent collaboration, distributed training infrastructure for agentic AI, and agentic federated learning frameworks. Use when: (1) designing multi-agent system architectures, (2) building distributed training infrastructure for AI agents, (3) implementing federated learning with agentic coordination, (4) scaling agent systems to thousands of concurrent tasks, ...

DoLQ: Discovering Ordinary Differential Equations with LLM-based qualitative and quantitative evaluation. Multi-agent architecture for symbolic regression of governing ODEs from data. Sampler Agent proposes candidates, Parameter Optimizer refines equations, Scientist Agent uses LLM for combined qualitative (domain knowledge) and quantitative (fit metrics) evaluation to iteratively guide search. Accepted at ICML 2026. Use when: ODE discovery, symbolic regression, scientific ML, equation discov...

LLM-based Design Structure Matrix (DSM) modularization methodology. Use when partitioning complex systems into cohesive modules, optimizing system architecture, or applying LLMs to combinatorial engineering problems. Activation triggers: DSM, design structure matrix, system modularization, architecture decomposition, LLM combinatorial optimization, semantic alignment hypothesis, engineering design optimization.

Dynamic Gradient Gating (DGG) methodology for sample-efficient RLVR - monitoring lm_head gradient norm to detect harmful policy shift and intercept gradients before corruption

EARLY (Evolutionary Algorithm for Reservoir Learning and Yielding) - evolutionary framework for discovering multi-reservoir ESN architectures. Graph-based genomes encode modular ESN topologies, evolves both structure and hyperparameters. Outperforms random search on CogScale temporal tasks, adapts to cross-situational learning. Activation: evolutionary reservoir, ESN topology search, multi-reservoir, temporal learning, modular brain-inspired.

Economy of Minds - Multi-agent intelligence emerging from economic interactions (auctions, payments, wealth accumulation). Decentralized self-orchestration without explicit communication protocols. Inspired by Hayek's economic theory. Activation: multi-agent economy, decentralized coordination, economic selection, agent auctions, wealth accumulation, Hayek theory, emergent intelligence, self-organization. Tags: multi-agent, economics, decentralized, self-organization, coordination, auctions, ...

Theoretical analysis of effective target shift in online learning and methods to correct for it. Explains why online learning struggles under distributional shift and how to characterize the relationship between online and offline learning. Activation triggers: online learning, target shift, distributional shift, online vs offline learning, sequential learning theory

SR²AM (Self-Regulated Simulative Reasoning Agentic LLM) methodology from arXiv:2605.22138 (May 2026). Three-system framework decomposing agent reasoning: System II (simulative planning via world model), System III (self-regulation deciding when/how to plan), and System I (reactive execution). Use when working on: agent reasoning architectures, adaptive computation for LLM agents, self-regulated planning, token-efficient reasoning, or world-model-based planning.

Extended Sompolinsky-Crisanti-Sommers (SCS) chaos theory for Excitatory-Inhibitory recurrent networks with target-specific inhibition. Derives mean-field theory for E/I networks, identifies three dynamical regimes (quiescence, asynchronous chaos, coherent oscillations), and shows coherent oscillations suppress chaos. Activation: chaos-synchrony, SCS theory, E/I balance, target-specific inhibition, dynamical mean-field theory, neural phase diagram, recurrent network dynamics, excitation-inhibi...

Expert guidance for designing modular Mixture-of-Experts (MoE) architectures using emergent document-level expert grouping. Based on EMO paper (arXiv:2605.06663). Use when designing sparse LLM architectures, MoE modularity, expert specialization, memory-efficient LLM deployment, or composable model architectures.

E-SPL: Evolutionary System Prompt Learning

Evolutionary Generative Merging (EvoGM) framework for training-free LLM composition via learnable generative modeling and dual-generator architecture with cycle-consistent learning

Excitation-driven data generation and distributed control optimization for building thermal systems and district heating networks. Combines BuilDyn framework (arXiv:2605.29849) and distributed NMPC with ADMM (arXiv:2605.29841).

Factorized Low-Rank RNN (FacRNN) framework for uncovering independent neural latent dynamics and connectivity. Group-wise independence among latent dimensions with variational autoencoder formulation and partial correlation penalty. Disentangles interpretable latent trajectories in low-dimensional space for neural population activity analysis. Use for: neural latent dynamics discovery, low-rank connectivity interpretation, disentangled representation learning, neural population modeling, inde...

FCN-LLM: Empowering LLMs for Brain Functional Connectivity Network Understanding via Graph-level Multi-task Instruction Tuning. Covers the multi-scale FCN encoder, semantic projection into LLM, 19-attribute multi-paradigm instruction tuning, two-stage learning strategy, and zero-shot generalization. Based on arXiv 2603.01135.

Backpropagation-free continual learning using feedback-aligned Hebbian plasticity. Replaces backpropagation with local Hebbian updates guided by random feedback connections, enabling biologically plausible continual learning without catastrophic forgetting. Use for bio-inspired learning, continual/incremental learning, and backprop-free neural network training. Activation: backprop-free learning, feedback alignment, Hebbian continual learning, local learning rules, biologically plausible trai...

FLAME: Adaptive Mixture-of-Experts for continual multimodal multi-task learning. Handles both co-available multi-task pretraining and sequential continual adaptation. Activation triggers: FLAME MoE, continual multimodal learning, adaptive mixture of experts, multi-task continual learning, sequential task adaptation

Free Energy Principle-based Mixture-of-Experts routing methodology. Uses LIF membrane potentials (beta) for temporal memory, precision-weighted gating (Pi) for reliability assessment, and anticipatory routing to solve domain transition failures in sparse MoE. Trigger words: free energy MoE, MoE routing failure, domain transition, LIF gating, precision-weighted routing, anticipatory routing, mixture of experts failure, sparse MoE, expert affinity.

Free Energy Principle-based MoE routing using LIF membrane dynamics. Solves domain transition failures in sparse MoE with three mechanisms: temporal memory (beta), precision-weighted gating (Pi), and anticipatory routing. 124x improvement at transitions. Activation: free energy principle, MoE routing, domain transition, predictive routing, LIF gating, Friston, mixture of experts.

**Problem**: Classical energy-based models require symmetric weight matrices, excluding biologically realistic E-I networks with asymmetric connectivity.

Geno-Synthetic Algorithm: type-factored coevolutionary optimization for heterogeneous genotypes and assembled phenotypes. Use when: coevolutionary algorithms, heterogeneous genotype optimization, assembled phenotype synthesis, evolutionary computation, neural architecture search, modular evolutionary design. Activation: geno-synthetic, coevolutionary optimization, heterogeneous genotype, assembled phenotype, type-factored evolution, modular evolutionary algorithm.

GFFMERGE methodology for efficient closed-form model merging in Graph Neural Networks. Exploits linear structure of message-passing layers to enable near-quantum accuracy atomistic simulations without retraining foundation models. Applicable to drug discovery, materials science, and general GNN transfer. Activation: GNN model merging, graph neural network transfer, neural force field merging, molecular simulation, atomistic GNN, quantum accuracy GNN, convex embedding alignment

图神经网络与 Transformer 融合的多模态数据融合方法论。 整合非欧几里得脑影像数据与欧几里得表格数据，支持时间感知的纵向预测。 触发词：多模态融合、脑网络、GNN、Transformer、时序预测、纵向分析、 multimodal fusion, brain connectivity, GNN-TF, temporal fusion。

Graph-based workflow management paradigm for efficient LLM agent serving using unified directed graphs (wGraph) for dynamic workflow instantiation with KV-cache optimization. Use for LLM agent serving, workflow management, agent serving optimization, agent orchestration, KV-cache management.

GTaS Generative Spike Train Model

GyralNet Subnetwork Partitioning

Hardware-motivated noise modeling methodology for fault-tolerant quantum computing benchmarking. Uses structured noise families (Pauli bias, measurement bias, spatial non-uniformity) instead of uniform depolarizing model to faithfully reflect real device characteristics. Enables joint code-hardware co-design. Use when evaluating QEC protocols, designing fault-tolerant quantum architectures, or benchmarking logical primitives under realistic noise conditions.

Benchmarking 7 local Hebbian learning rules for associative memory storage and prototype extraction. Bayesian-Hebbian rules achieve highest capacity. Activation: hebbian learning benchmark, associative memory capacity, prototype extraction, Bayesian-Hebbian learning, covariance learning.

分层 MoE 架构技能 - 用于对象检测的分层实例条件化混合专家模型 (HI-MoE)。通过两级路由机制实现稀疏计算与实例中心结构的匹配。基于论文 HI-MoE: Hierarchical Instance-Conditioned Mixture-of-Experts (arXiv 2604.04908)。激活关键词: MoE, mixture of experts, object detection MoE, instance routing, 分层路由, 实例条件化。

Web-scale LLM-based multi-agent system architecture for the Agentic Web. Focuses on five-layer coordination architecture, heterogeneous agent interaction, and open-world scaling challenges. Use when: (1) Designing large-scale multi-agent systems, (2) Implementing web-scale agent coordination, (3) Building agentic web architectures, (4) Studying LLM-based multi-agent systems, (5) Understanding agent ecosystem evolution toward AGI.

Modern Hopfield Networks for continual learning in diffusion models via energy-based intrinsic forgetting and replay selection

Hippocampal-Entorhinal (HPC-MEC) inspired hierarchical world model for structure abstraction and generalization from video sequences. Based on arXiv:2605.15733 (May 2026). Use when: designing brain-inspired world models, HPC-MEC cognitive architecture, structure abstraction from video, latent transition learning, hippocampal-entorhinal coupling models, continuous attractor neural networks for AI, path integration in abstract spaces, self-supervised world model learning, zero-shot structural t...

Integrative neurocybernetic modeling framework for large-scale neuroscience. Treats brain as controller pursuing latent objectives in closed-loop coupling with body and environment. Keywords: neurocybernetics, closed-loop modeling, large-scale neuroscience, brain-body-environment coupling, nonlinear state-space models

Bayesian modeling methodology for ion channel gating with missed event correction. Integrates temporal resolution limitations of patch clamp recordings into hierarchical Markov chain likelihood functions, enabling unbiased kinetic parameter inference and model selection for IP3R calcium channels. Reveals multimodal gating behavior with Park/Drive mode switching regulated by Ca2+ concentration. Use when: modeling ion channel kinetics, analyzing single-channel patch clamp data, correcting for m...

DMOSOPT — scalable optimization framework using jointly learned surrogate models for constrained multi-objective optimization of neural dynamical systems. Learns smooth approximations of objective landscapes and feasibility boundaries to guide search with unified gradients.

Kinetic energy in random recurrent neural networks - links chaotic dynamics and unstable fixed points through dynamical mean-field theory. Cubic scaling at critical point, shell-like chaotic manifold geometry. Activation: kinetic energy, random RNN, chaos, dynamical mean-field theory, chaotic dynamics, fixed points.

Neuro-inspired phase encoding using Kuramoto oscillators (KoPE) applied to Vision Transformers, combining oscillatory dynamics with attention mechanisms for improved learning efficiency. Based on ICLR 2025 paper by Xiao et al. (Microsoft Research).

基于格论的RNN剪枝方法论。将RNN建模为偏序集，构建依赖格，识别不可约元进行选择性剪枝。相比传统幅度剪枝更好地保留功能连接性。触发词：RNN剪枝、格剪枝、偏序集、依赖格、不可约元、网络压缩、lattice pruning、poset、meet irreducible。

Distributionally robust free energy principle for reliable robotic control. Jointly learns environment dynamics and rewards while ensuring robustness to epistemic uncertainties. Validated on Franka Research 3 arm manipulation tasks.

Developmental scaffoldings methodology for guiding self-organisation through learned pre-patterns. Joint NCA+SIREN model that offloads information to initial conditions, enabling robustness, encoding capacity, and symmetry breaking improvements. Activation: developmental scaffoldings, self-organisation, neural cellular automata, NCA, pre-patterns, morphogenetic, developmental biology, SIREN, information offloading.

Learning Dynamic Stability Landscapes in Synchronization Networks methodology - graph-to-image prediction paradigm for predicting stability landscapes from network topology. Pioneers image-like per-node stability landscapes beyond scalar indices. Applicable to neuroscience, power grids, biological synchronization. Activation: stability landscape, synchronization stability, graph-to-image prediction, dynamic stability, oscillator networks, power grid stability.

Long Context Pre-Training with Lighthouse Attention — NousResearch 提出的对称选择式分层注意力算法。面向超长上下文预训练，通过对称 Q/K/V 金字塔池化、无参数打分 top-K 选择、选择与注意力解耦、两阶段训练，实现 O(N·d) 复杂度。触发词：lighthouse attention, long context pretraining, hierarchical attention, symmetric pooling, sparse attention, FlashAttention wrapper

Design LLM agent systems using the externalization framework from cognitive artifacts theory (Norman). Covers memory externalization (state across time), skills externalization (procedural expertise), protocol externalization (interaction structure), and harness engineering (unification layer). Use when architecting multi-tool LLM agents, building agent frameworks, designing memory/skills/protocol systems, or unifying agent components. Keywords: agent externalization, cognitive artifacts, mem...

Decision-Centric framework for LLM systems that separates decision signals from action policies. Apply this when designing LLM control flow, routing, adaptive inference, or building diagnosable agent systems.

Methodology for identifying and analyzing functional emotion representations in LLM internals. Covers finding emotion-related neural activity patterns, testing their causal influence via activation steering, and understanding how abstract emotion concepts shape model behavior. Use when: (1) analyzing LLM emotional behavior, (2) studying representation causality, (3) investigating model decision-making driven by internal states, (4) safety research on models taking undesirable actions under em...

Neuroscience-inspired geometric account of in-context learning (ICL) in LLMs — how representational geometry reshapes to support online untangling and classification without parameter updates.