Claude Skills by hiyenwong

Persistent-negative adversarial distillation for black-box on-policy distillation - anchor the discriminator with historical teacher-student comparisons to stop the moving-target reward problem. Use when distilling from API-only teachers (no token probabilities), training discriminator-based rewards, or stabilizing adversarial RLHF pipelines.

Use when adding short-term synaptic plasticity to CNNs.

Use when compiling 2-local Hamiltonians for chiplet QPUs.

Symmetry sets untrainable error floors in physical nets.

Reward vs gradient channels shape emergent rule codes differently.

Spectral-window transfer analysis for cross-domain foundation models - pretraining transfer quality is governed by the overlap between the pretrained model's spectral window and the target domain's spectra (infrared-pretrained model falls below untrained control on optical/UV tasks). Use for deciding pretrained-vs-in-domain pretraining for spectroscopy, sensor, or any wavelength/frequency-structured scientific data.

HySTAR - fix the value-decomposition basis (overlapping sparse hypergraph scaffold) while learning adaptive spatiotemporal representations, eliminating structural target drift in cooperative multi-agent RL. Use for credit assignment under partial observability, stable high-order coalition value decomposition, or when dynamic grouping keeps reshuffling your learning targets.

Data-driven machine learning approach for medium-term degradation forecasting in Anion Exchange Membrane water electrolyzers using LSTMs and CNNs

Use when splitting a large QUBO into sub-problems for hybrid solvers.

Benchmark GKP-concatenated qLDPC codes (BB vs tricycle) under circuit-level noise.

Use when improving sample efficiency of biological sequence prediction (peptide-HLA/neoantigen binding) with hybrid quantum-classical neural networks in low-data regimes.

Use when OGP separates classical heuristics from QAOA depth barriers.

Use when evaluating quantum CNN generalization in small-data regimes — encoding bottleneck analysis (amplitude vs angle), parameter-matched comparisons, mid-circuit measurement QCNN design.

Use when replacing softmax attention with quantum-derived doubly stochastic matrices (QDSM) for data-limited molecular/gene-expression prediction from histopathology.

Reductions between quantum auctions and secure summation.

Quantum speedup for Monte Carlo tree search. Use when combining quantum oracles with game tree search.

Unique Cramér-Rao-efficient quantum U-statistic estimators.

RL agent emits QNN gradients to bypass barren plateau variance decay.

Use for real-time superconducting qubit readout with spiking neural networks on FPGA. Streaming SNN discriminators.

GMRF-space + AR(1)-time Bayesian receptive fields via INLA. Use for spike-count RF estimation.

We present an overview of efforts to bridge the gap between overhead and practicality for privacy-preserving learning systems using MPC, ZKPs, and FHE. Through hardware/software/algorithm co-design, we show progress towards enabling LLM-scale applications in privacy-preserving settings including DNN IP ownership, ethical LLM usage enforcement, and transformer inference.

Research paper: Progressive Skill Discovery as Access Control for Tool-Using LLM Agents: Structural Governance through Role-Scoped Capability Delivery.

Research paper: Codetta: High-Capacity, Keyless, and Undetectable Multi-Agent Collusion.

Research paper: Subjects, Not Authors: The Authorship Hazard in Agentic Dataspaces.

Quaternion chart makes SU(2) updates degree-2 → depth-1 homomorphic FL.

We propose a novel hybrid ANN-SNN framework that integrates layer-wise encode-decode SNN blocks within conventional ANN pipelines. Central to our method is the use of surrogate gradients for a bit-plane-based spike encoding function, enabling end-to-end differentiable training across ANN and SNN layers. This design achieves competitive accuracy with energy-efficient computing.

Use when analysing SNN expressivity or choosing SNN weight init. Causal pieces theory.

Use when building deep SNNs trained with local STDP and temporal fusion.

Use when evolving native SNN architectures with LLM search

URCHIN: A Horizontal Spiking Language Model for Data-Constrained Pretraining. Single-layer Dale's-law E/I LIF connectome with multi-transmission loop, dual parallel-SSM/serial-RSNN deployment from one weight set. Use when building biologically-plausible spiking LMs, BabyLM-scale training, or event-driven edge language models.

We propose BRIDGE, an RL-optimized depth-to-image generation framework that synthesizes over 20M realistic and geometrically accurate RGB images from diverse source depth maps. Then we train depth estimation models with hybrid supervision integrating teacher pseudo-labels with ground truth depth. BRIDGE achieves breakthroughs in scale and domain diversity, consistently outperforming state-of-the-art approaches.

LLMs develop rich visual priors despite text-only training. We reveal that visual priors are composed of separable perception and reasoning priors with unique scaling trends and origins. Visual reasoning ability is predominantly developed by pre-training on reasoning-centric data (code, math, academia). We propose a data-centric recipe for pre-training vision-aware LLMs verified in 1T token scale pre-training across 100+ controlled experiments consuming 500,000 GPU-hours.

Cortical flatmaps + frozen SigLIP2 for fMRI prediction.

HierINRSeg - hierarchical multi-layer INR aggregation for parameter-efficient semantic segmentation; INR advantages concentrate in low-parameter and limited-augmentation regimes, so aggregate complementary semantic structure distributed across INR layers instead of scaling parameters. Use for segmentation under memory budgets, cross-site domain shift, or analyzing when INRs beat CNNs.