Claude Skills by ADu2021

Implement low-rank attention routing using encode-decode factorization to achieve linear-time complexity on long sequences while maintaining compatibility with optimized attention kernels.

Accelerates long-context LLM prefilling by identifying sparse attention patterns without expensive scoring, using block-level approximations and dynamic thresholding. Achieves 27.78x speedup at 256K tokens while maintaining accuracy.

Fuse categorical sampling directly into LM-head matrix multiplication to eliminate logits materialization. Use Gumbel noise during computation and hierarchical reduction to achieve 19% token-level latency reduction.

Reduce agent execution steps by 35% and latency by parallelizing sequential tool calls through task dependency graphs (DAGs). Use when deploying information-retrieval agents where tool execution ordering is flexible.

Enable LLM agents to improve continuously during deployment by constructing structured experience libraries through self-reflection on successes and failures—achieving 23% improvement on reasoning without gradient-based parameter updates or external training.

Understand how token generation flexibility in diffusion LMs paradoxically constrains reasoning, as models exploit ordering flexibility to avoid uncertain tokens, and apply simplified approaches that preserve parallel decoding benefits. Use when optimizing diffusion-based language models for reasoning tasks.

Train language models where each expert learns independently on closed datasets, enabling flexible inference with selective data inclusion or exclusion. 41% performance improvement while allowing users to opt out of specific data sources without retraining.

Uses flow maps as look-ahead operators to enable principled reward-guided diffusion by predicting trajectory endpoints at any denoising step. Deploy when applying rewards or preferences to diffusion trajectories with meaningful gradients throughout generation.

Accelerate video generation by allocating smaller models to intermediate diffusion timesteps and larger models to capacity-critical early and late stages. Achieves 1.65x speedup and 57% FLOP reduction while maintaining visual quality. Use when video generation latency or computational cost is critical and you have multiple model sizes available.

Improve LLM serving under mixed workloads by decoupling execution granularity from scheduling frequency. Operator-level preemption allows fine-grained interruption at natural boundaries (attention, feed-forward layers) without efficiency loss. Event-driven scheduling triggers decisions only on request arrival/completion. Eliminates head-of-line blocking where long requests starve short time-sensitive ones. Achieves 5.6× higher goodput vs. baselines in production workloads.

Train LLMs with distribution-matching rewards instead of reward maximization to achieve 10% improvement on math reasoning while improving solution diversity by matching the full reward distribution via flow balancing, addressing mode collapse in long chain-of-thought tasks.

Reduce web agent context size by 51% while maintaining task performance using lightweight LLM retrieval to extract relevant accessibility tree lines. Task-guided filtering removes irrelevant elements, improving inference cost and security by neutralizing prompt injection attacks without sacrificing normal operation.

Use a lightweight LLM to filter accessibility tree observations by task relevance, reducing agent context size by 50-80% while maintaining equivalent task performance.

Training-free prompting strategy that pre-organizes query information into compact structured format, reducing generation tokens by 2-3× while maintaining reasoning accuracy. Apply when reasoning performance is bottlenecked by verbose input formatting.

Replace expensive test-based execution loops with learned prediction models that forecast agent action outcomes before commitment. Framework uses internalized execution priors and structured analysis reports to achieve 6x faster convergence and 6% higher performance compared to execute-first baselines. Applicable to scientific discovery, optimization, and hypothesis testing where verification is computationally or financially expensive.

Predict when LLM outputs are trustworthy for formal reasoning by analyzing domain-specific uncertainty signals.

Training-free framework compressing KV caches using Fourier basis functions, exploiting heterogeneous transformer head roles for memory-efficient long-context LLMs.

Diagnose and solve LLM reproducibility failures caused by floating-point precision across hardware configurations using LayerCast optimization for deterministic inference with minimal memory overhead.

Extend video diffusion models to generate 4-8× longer sequences without retraining. Uses frequency-aware attention to blend local detail preservation with global consistency, identifying and fixing high-frequency distortion in extended videos.

Generate smooth morphing sequences between images without fine-tuning or alignment. Uses guidance-aware spherical interpolation and step-oriented attention blending to handle diverse semantic and layout scenarios, completing morphs 50× faster than fine-tuning methods.

Achieve emergent brevity in reasoning by retaining and up-weighting easy problems during RL training, implicitly regularizing solution length without explicit penalties while maintaining accuracy on hard problems.

Scale research agent capability using persistent filesystem as external memory. Dual-agent architecture with context builder accumulating knowledge and report writer composing outputs enables computation scaling beyond context windows.

Build unified multimodal embeddings with a single transformer encoder processing image and text tokens together, improving performance on structure-aware tasks through early fusion.

Combine multiple specialized language models at token-level granularity without joint training or architectural compatibility. FusionRoute performs expert selection and generates complementary logits to overcome routing limitations, achieving superior cross-domain performance at inference time.

Guide LLM exploration through the model's own gradient geometry rather than external signals. Extract sequence-level gradient features measuring how tokens would reshape output distributions. Reward responses introducing novel gradient directions while deemphasizing redundant ones. Achieve orthogonal gradient directions and improved accuracy.

Build a conversational AI that combines image geolocalization with contextual geographical knowledge. GAEA enables users to query precise GPS locations from images while receiving conversational responses about places, their attributes, and regional context—outperforming GPT-4o by 7.2% on geography-aware visual QA tasks.

Improve RL training efficiency by 2.5× using angle concentration between token hidden states as a cost-effective data scheduling signal, selecting high-gradient samples dynamically.

Implement techniques from GameTalk: Training LLMs for Strategic Conversation. Strategic decision-making in multi-agent settings is a key challenge for large language models (LLMs), particularly when coordination and negotiation must unfold over extended conversations

Optimize language models against multiple reward signals simultaneously by decoupling reward normalization. GDPO prevents reward combination collapse that undermines training signal quality when aligning models to multiple human preferences like accuracy, safety, efficiency, and format compliance.

A standardized environment framework for training and evaluating LLM agents, providing 24+ tasks with asynchronous vectorized execution, extensible wrappers, and integration examples for five RL frameworks. Enables reproducible agent research and training at scale.

Generative approach for creating complex dynamic scenes and content, supporting agent capabilities in understanding and reasoning about multi-agent environments.

Build persistent, lossless agent memory using just-in-time compilation: store complete history in a universal page-store while performing dynamic deep research at query time, enabling test-time scalability through iterative information synthesis and reflection.

Segment novel 3D point cloud classes with few support samples by combining dense but noisy pseudo-labels from 3D vision-language models with precise sparse few-shot annotations. GFS-VL adapts to new classes while retaining base class performance, using prototype-guided filtering and adaptive infilling strategies ideal for applications with limited labeled training data.

Unified platform combining instruction-conditioned video diffusion, flow-matching action decoder, and action-conditioned simulator. Enables scalable robot learning without extensive labeled demonstrations.

Improve video diffusion consistency by aligning intermediate diffusion features with 3D geometric representations from pretrained foundation models, enabling spatially coherent and temporally stable video generation through angular and scale alignment losses.

Extend vision-language models with 3D spatial understanding by adding geometric expert stream alongside semantic expert: predict pixel-aligned 3D point maps, surface normals, and camera poses from 2D images, enabling unified reasoning across 2D semantic and 3D geometric domains.

Prevents catastrophic forgetting in continual LLM learning using threshold-based margin loss with fixed general replay samples from pretraining data.

Evolve Python algorithms and programs using LLMs as mutation operators combined with MAP-Elites quality-diversity search, achieving competitive results on geometric optimization and algorithmic problems by iteratively mutating code informed by historical performance and lineage context.

Control video camera motion using gravity-aligned absolute coordinates instead of relative trajectories. GimbalDiffusion enables precise camera control with null-pitch conditioning—ideal when you need interpretable, physics-aware camera motion in text-to-video.

Phase-aware acceleration using two lightweight LoRA adapters (Slow-LoRA for semantic reconstruction, Fast-LoRA for texture refinement) trained on a single image in one GPU hour, achieving 5× speedup via smart per-phase acceleration rather than uniform speedup.

Train vision-language models with curriculum-based reinforcement learning (RLCS) to improve reasoning across diverse multimodal tasks. Dynamically adjust training difficulty to model capability, preventing both trivial and overly-hard examples.

Enable frozen LLMs to predict their own correctness by decoding signals from internal hidden states and attention patterns, achieving reliable self-verification without external judges—adding only 5M parameters while reducing inference cost and improving calibration.

Train language models for formal theorem proving via expert iteration with verifier-guided self-correction and checkpoint merging.

Synthesize unlimited verifiable training tasks from unverifiable text by converting reasoning passages into multiple-choice problems. Creates higher-quality training data for RLHF systems without requiring new human labels.

Separates generative policy optimization through latent encoder (standard RL algorithms) and conditional decoder (frozen then refined), using two-timescale alternating schedule to eliminate gradient instability from direct generative policy optimization.

Balance LLM flexibility with domain rigor in scientific agents through dual-layer architecture. Enforce role-based access control and artifact-centric state management to prevent hallucinations, while preserving free-form reasoning for lower-risk tasks.

Improve adaptive learning rates by clustering gradient statistics within layers and applying cluster-specific scaling.

Compress long context into compact memory tokens via iterative gradient descent. Learn to write information into prefix memory without storing full KV-caches, enabling efficient long-context reasoning and retrieval.

Unifies supervised fine-tuning and reinforcement learning through GRAO framework that combines multiple-output generation with group direct alignment loss for improved preference learning.

Unify positional encoding methods via group action theory, encompassing RoPE and ALiBi as special cases. GRAPE enables exploration of cross-subspace feature coupling—ideal when you need principled positional encoding beyond standard implementations.