Claude Skills by hiyenwong

Use when modeling emotion-driven decisions or affective networks. Emotions as colored noise framework.

Use when combining multiple vision-language foundation models in few-shot settings. Closed-form Orthogonal Procrustes alignment of heterogeneous representations plus unified graph refinement of support features and prototypes.

Finite-N FRR for noisy coupled phase oscillators.

FlashLoop methodology — training-free inference acceleration for Looped Transformers via lazy/sparse updates and KV-residual quantization. Use when: (1) accelerating looped/recurrent Transformer inference that re-runs shared blocks, (2) KV-cache memory grows linearly with loop depth, (3) reducing cross-loop computational redundancy without retraining. Keywords: looped transformer, lazy update, KV-cache compression, sparse attention, inference speedup.

Use when removing specific motions from RL-trained humanoid policies. First motion-level unlearning method for physical humanoid control — degrades target motions while preserving the rest, resolving two unlearning-failure training mechanisms.

Use when designing cyclic contextuality or binding tests.

Decomposes neural population state changes into on-manifold movement, multiplicative gain modulation, and genuine off-manifold novelty via a tangent/normal bundle geometry with a radial gain axis, plus a 7-gate identifiability cascade (chart existence, intrinsic dimension, tangent frame, anchor bias, gain-axis existence, curvature booking, ratio conditioning). Use when analyzing neural state transitions, memory segmentation decorrelation, neuromodulator-driven gain vs. novelty attribution, ne...

Governed self-evolution: LLM proposes, frozen anytime-valid e-process referee judges. Controls FDR in agentic loops.

GridSFM methodology — physics-inspired GNN foundation model for AC Optimal Power Flow with disconnected-feasible-set repair via log-penalized slack lifting, plus Newton-based physics-informed fine-tuning for unseen grids. Use when: (1) neural solvers fail because the feasible set is disconnected (non-approximable by continuous nets), (2) building pretrain-then-finetune neural PDE/optimization solvers across topologies, (3) few-shot adaptation (100 instances) to new system instances. Keywords:...

Quenched disorder structure (not magnitude) controls SR and LSM performance in FHN reservoirs. Use for heterogeneous reservoir design.

Use for low-rank (LoRA) control of NCA morphogenesis. Maps D'Arcy Thompson transformations to reusable regulatory weight-space hyper-directions.

Use for gradient-free online task adaptation via NCA cellwise fast memory.

Use for multimode NEMS reservoir computing via drive-amplitude multiplexing.

Two-level fast-slow neural-astrocyte architecture implementing a hybrid automaton for latent-context inference. Slow astrocytic RNN accumulates reward-prediction-error evidence; context switches emerge from a cascade of two bifurcations (attractor annihilation + level-set boundary crossing). Stickiness is mediated by environmental entropy flattening the unrewarded vector field. Astrocytic tiling/lattice topology suffices up to 90% sparsity.

Noise-assisted internal simulation methodology. Deterministic attractor replay mis-represents rare events (greedy-recurrent over-representation OR low-prior-veto under-representation); moderate Ornstein-Uhlenbeck noise on unit support restores marginal AND conditional occurrence fidelity, with an inverted-U (stochastic resonance) signature and parameter-tolerance broadening.

NP-aided shadow tomography no-gos for Microcrypt PRS/PRU.

OED zero-storage neural synthesis via Mandelbrot boundary.

Use when probing why models fail equivalent problems or analyzing internal representations of reasoning. Permutation SNR metric distinguishing answer invariance from representation invariance; higher SNR correlates with accuracy on reordered problems.

Soma voltage from dendritic-only recordings via HH PINN.

Plasticity as prospective property: inverse configurational constraint P=1/C.

Two zero-private-capacity channels jointly send private bits (arXiv 2609.10520).

Use when auditing LLM chat privacy. PrivDrift method.

Use when learning MPS or TTN states from arbitrary noisy states.

Use for ATDM BCI: prototype-based EEG evidence encoding.

Task-resolved Fisher spectroscopy for quantum reservoir computing: separates encoding vs measurement vs compression loss.

Use when integrating LLMs into safety-critical commissioning or requirement acceptance. Separates candidate generation from release authority via deterministic entailment gates.

Use for private inference on manifold-valued optimization.

Use when running tabular foundation models on wide tables under GPU memory limits. Training-free inference framework converting quadratic feature-interaction cost to linear-in-width with bounded working set, no ensembling or new parameters.

Task-aligned receptive fields are a computational prior beyond sparsity; value shrinks as single-neuron expressivity grows. Use for wiring priors in SNNs.

Use when bounding PGM error in quantum state discrimination.

Use for graph-spectral transfer analysis of Kuramoto sync.

Use when implementing fault-tolerant logical T gates on stabilizer codes.

Use when validating time-series models. Prices the causality law.

Use when standardizing verification/validation workflows or building requirements-to-evidence traceability schemas. JPL relationship-based V&V schema.

Use when designing state-agnostic tests or universal codes.

**arXiv ID:** 2609.29322v1 **Authors:** Seungwoo Han, Ingon Chanpornpakdi, Motoi Noda, Puwadej Leelasiri, Ibuki Hiruma, Toshihisa Tanaka **URL:** http://arxiv.org/abs/2609.29322v1 **Utility Score:** 1.00

**arXiv ID:** 2609.29372v1 **Authors:** Kwok-Ho Ng, Tingting Song, Bingwen Feng, Zhihua Xia **URL:** http://arxiv.org/abs/2609.29372v1 **Utility Score:** 1.00

**arXiv ID:** 2609.29951v1 **Authors:** Zhiyu Zhang, Yupeng Li **URL:** http://arxiv.org/abs/2609.29951v1 **Utility Score:** 1.00

**arXiv ID:** 2609.30079v1 **Authors:** Anne M. Tumlin, Ben Wooding, Zhenxuan Shao, Diego Manzanas Lopez, Tyler Derr, Taylor T. Johnson **URL:** http://arxiv.org/abs/2609.30079v1 **Utility Score:** 1.00

**arXiv ID:** 2609.30150v1 **Authors:** Giacomo Arcieri, Gregory Duthé, Christophe Muller, Konstantinos G. Papakonstantinou, Daniel Straub, Eleni Chatzi **URL:** http://arxiv.org/abs/2609.30150v1 **Utility Score:** 1.00

CDRTL framework coupling Lorenz oscillator networks with persistent Laplacian fingerprints for preictal EEG node classification. Use for EEG topology-dynamics feature engineering.

Use on raw EEG; ordinal entropy survives blink artifacts.

Dynamic mean-field theory for how four second-order synaptic motifs (chain, reciprocal, convergent, divergent) shape nonlinear random-network dynamics. Covers outlier/bulk eigenvalue reshaping, ferromagnetic/limit-cycle/glassy regimes, quenched-heterogeneity chaos suppression, and Lyapunov geometry. Activation: synaptic motifs, DMFT, path-integral, network motifs, glassy dynamics, limit cycles, chaos dimensionality.

Use when mapping VQA losses to random fields.

**arXiv ID:** 2609.29398v1 **Authors:** Xunkai Li, Xu Wang, Yinlin Zhu, Xiong Yongfu, Yi Liu, Rong-Hua Li, Guoren Wang **URL:** http://arxiv.org/abs/2609.29398v1 **Utility Score:** 1.00

Use when scheduling quantum network entanglement requests.

MPS + shared physical legs encode long-range correlation.

Use when optimizing QAOA on hierarchical problems. Learned projectors.

Use when budgeting MPO rank in quantum simulation.

Use when quantum privacy clashes with entanglement needs.