Claude Skills by ADu2021

Enable efficient parallel decoding while preserving causal inference properties through progressive distillation with cyclic noise scheduling. Reduce training complexity via block-wise sparse attention, implement rejection recycling and multi-block decoding. Achieve 3.8-4.0× wall-clock speedup on coding benchmarks.

Enable scalable MoE inference by disaggregating attention and expert layers onto independent GPU sub-clusters. Use adaptive two-phase communication, activation load-balanced scheduling, and activation-aware expert management. Achieve 3.9× higher per-GPU throughput than state-of-the-art systems.

Implement techniques from Jet-RL: Enabling On-Policy FP8 Reinforcement Learning with Unified Training and Rollout Precision Flow. Reinforcement learning (RL) is essential for enhancing the complex reasoning capabilities of large language models (LLMs)

Accelerate diffusion transformers by processing only sparse anchor tokens in early stages, then expanding to full spatial resolution using learnable extrapolation. Combines SAG-ODE velocity field lifting with importance-guided token activation for lossless speedup.

Demonstrate that simple single-stage RL with fixed hyperparameters matches complex multi-stage approaches for training small LLMs on mathematical reasoning. Use basic setup: GRPO algorithm, rule-based verification, 16K token context, standard training data without difficulty filtering. Train two 1.5B models to competitive performance using 2× less compute than sophisticated approaches.

Generate optimized GPU kernels by treating LLMs as planning engines that co-evolve with a world model. Decouples high-level algorithmic planning from low-level implementation, enabling structured search through optimization strategies. LLM world model estimates priority scores for pending optimizations while iteratively updating understanding based on execution results. Achieves 2.10× improvement over evolutionary baselines with 14.3× gains on complex MoE kernels.

Pixel-based reinforcement learning agents often fail under purely visual distribution shift even when latent dynamics and rewards are unchanged, but existing benchmarks entangle multiple sources of shift and hinder systematic analysis. We introduce KAGE-Env, a JAX-native 2D platformer that factorizes the observation process into independently controllable visual axes while keeping the underlying control problem fixed. By construction, varying a visual axis affects performance only through the...

Automate compute kernel optimization via agentic AI and retrieval-augmented prompting. Unified context-aware transformation function, hardware-specific constraint KB, self-improving state machine exploring kernel variants—achieving 1.25-17× speedups on production recommendation workloads in hours vs weeks of manual effort.

Reveals that memory bandwidth—not computation—dominates test-time costs; proposes eFLOPs metric incorporating both computation and memory, showing 14B+ minimum threshold for reasoning value.

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

Unify video generation, editing, and reasoning through Multimodal Visual Language (MVL) paradigm processing text instructions, reference images, and video contexts in shared embedding space. Employ prompt enhancer, omni-generator with diffusion transformer, and multimodal super-resolution. Support diverse user inputs with in-context generation and reasoning-based editing.

Improve RL training for LLMs by dynamically allocating exploration budget (rollout count) to tasks based on their difficulty and current learning status. Solves the knapsack problem of maximizing gradient signal within fixed compute budget, increasing non-zero policy gradients by 20-40% and achieving 2-4 point performance gains.

Enable real-time portrait animation via causal AR video generation with temporal coherence. Sliding window with global reference frame caching, temporal knot module overlapping adjacent chunks, global context running ahead—preventing error accumulation while maintaining streaming efficiency and bidirectional-quality consistency.

Build knowledge agents that generalize across grounded reasoning tasks by combining agentic data synthesis via vector search with off-policy reinforcement learning. Use multi-task training on synthetic question-answer pairs to achieve robust cross-domain performance.

Enables 8x KV cache compression with minimal training overhead to improve reasoning accuracy by allowing more token generation within computational budgets.

Guide frozen language models toward multi-step reasoning by modifying cached key-value representations after the prefilling stage. Extract steering vectors from contrastive prompt pairs and apply them to KV cache with scalar coefficients. Improves reasoning on GSM8K, ARC, CommonsenseQA while adding only 10ms overhead per token.

Extract high-quality embeddings from frozen decoder-only LLMs by re-routing internal key-value states without training—outperforming training-free baselines by 10% on MTEB while maintaining robustness across sequences up to 4,096 tokens.

Build an 8B multimodal model specializing in short-form video understanding. Combines four-stage pre-training with instruction-tuning and reinforcement learning to enable advanced reasoning about dynamic video content while maintaining general vision-language capabilities.

Enable LLM agents to actively explore and adapt policies through meta-RL instead of converging to fixed behaviors. Uses cross-episode training with trajectory-level discounting and in-context policy adaptation via textual reflections—achieving 11-19% improvements in exploration-exploitation balance across interactive environments.

Align vision-language-action models with natural language instructions using Bayesian decomposition with latent action queries, improving robotic manipulation generalization. Use when building agents that follow language commands while observing visual scenes and producing motor control.

Advanced reasoning approach for optimizing inference efficiency through meta-cognitive planning, enabling agents to make better decisions with reduced computational overhead.

Compute reasoning rewards from the model's own next-token probability distribution at solution end. Integrates verification-based feedback into single model without separate evaluator, enabling efficient RL training with minimal overhead.

Convert webpage designs to code via Layout-as-Thought reasoning, detecting layout structure and generating HTML/CSS for spatial blocks.

Improve reasoning quality by decoupling reasoning from verbalization. Planner generates deterministic latent trajectories while Decoder grounds them to text. Enables dynamic termination, better Pass@k scaling, and interpretable intermediate states.

Enable LLM agents to collaborate entirely in latent space by sharing layer-wise KV cache representations instead of text, achieving 4× speedup and 71-84% fewer output tokens while maintaining 14.6% higher accuracy through lossless information preservation in continuous embeddings.

Replace VAE in latent diffusion with frozen DINO self-supervised features plus lightweight residual processing, enabling faster training, few-step sampling, and clearer semantic structure.

Reduce hallucinations in multimodal reasoning by detecting high-entropy (uncertain) states and switching to continuous latent embeddings instead of discrete tokens. Use prior-guided visual anchoring during uncertain phases to maintain grounding.

Learn world models by decomposing scenes into latent particles with per-particle dynamics. Each particle represents an object with position, scale, and appearance. Learn distributed latent actions governing per-particle transitions, enabling multimodal video generation from identical initial conditions.

Enables multimodal reasoning by interleaving visual sketches with text. MLLMs generate latent visual representations during reasoning, with sketch decoder converting them to human-interpretable images. Improves reasoning performance while maintaining interpretability through visual thinking aids.

Enable models to reason in continuous latent space via context-prediction fusion, combining hidden state context with vocabulary embeddings to maintain scaling across model sizes.

Unify image understanding and generation with layerwise timestep experts and residual attention reuse, achieving 6x faster inference than comparable unified models while maintaining competitive performance.

Semi-structured latent representation combining efficiency of VecSet with spatial structure guidance via voxel queries and rotary positional embeddings, enabling strong test-time scaling (6K to 30K tokens) and improved model scaling without sparse components.

Improves vision-language model distillation by aligning latent visual reasoning trajectories between teacher and student, enabling 3B parameter models to outperform larger open-source and proprietary systems with +16.9% gains on reasoning tasks.

Improve factual accuracy in LLM generation through decoding-time layer-wise attention suppression. Selectively suppress attention to specific token types at their most influential transformer depths without training or model modifications. Use when you need to reduce hallucinations and improve factual correctness at inference time.

Learning-based approach to reduce key-value cache memory during inference through static channel-wise sparsity masks. Achieves 70% K cache and 16-18% V cache memory reduction with 1.3x attention speedup.

Adapt language model weight matrix scales automatically during training by adding learnable scalar and vector multipliers to each layer. Breaks the noise-weight-decay equilibrium that constrains scales based on hyperparameters rather than data, enabling data-driven scaling adaptation without inference cost or extensive tuning overhead.

Compress LLM weights to 4-bit precision using learned per-row lookup tables that minimize output activation error, achieving better accuracy than fixed formats (int4, fp4, nf4) while maintaining inference speed.

Enable VLMs to perform dynamic spatial reasoning (DSR) by extracting 4D priors from videos and using Geometry Selection Modules (GSM) for selective injection. Provides DSR-Train dataset (50K QA pairs) and benchmark with six reasoning types, balancing geometric specialization with general video understanding—improving VLM 4D reasoning without degradation on general tasks.

Train student models to exceed teacher performance by extrapolating rewards during distillation using a scaling factor λ1. Enables unified students to surpass all individual domain teachers through controlled reward amplification.

Apply reinforcement learning at test time to enable language models to continue adapting on specific problems, achieving state-of-the-art results across mathematics, GPU optimization, algorithms, and biology. Use when you need models to discover domain-specific problem-solving strategies without retraining.

Build autonomous agents that accumulate structured knowledge from task execution into hierarchical memory (strategic, procedural, tool) without human annotation, enabling knowledge transfer to unseen tasks.

Enable models to autonomously improve on target tasks during inference via test-time curricula (TTC-RL). Automatically select task-relevant training examples and apply RL to continue learning. Achieve 1.8x improvement on AIME25 math benchmarks and 2.1x on CodeElo competitive coding by learning task-specific skills at test time without human curation.

Multi-dimensional reward function combining factual precision, response detail, and answer relevance for online RL. Reduces hallucinations 23.1% while maintaining helpfulness and detail.

Learn to predict and generate high-impact research ideas by training models on community feedback signals. Apply reinforcement learning to align research generation with citation-based indicators of scientific impact.

Enable agents to adapt to individual user preferences across extended interactions via memory-equipped architectures. Use learning signals to improve memory updates and enable long-term collaborative relationships with human users.

Decompose visual priors into perception and reasoning components, each optimized by distinct data types: reasoning from code/math corpora, perception from diverse modality-rich sources. Use to construct efficient VLM pretraining pipelines balancing multimodal perception with reasoning capability.

Release 150K-hour multilingual speech corpus with word-level timestamps plus non-autoregressive flow-matching TTS and autoregressive speech editing models. Enable diverse speech generation and editing tasks across 10 languages.

Overcome the shallow exploration trap by explicitly rewarding longer reasoning sequences when models fail to solve problems. Use length-incentivized exploration to enable deeper chain-of-thought reasoning, achieving better test-time scaling and improved generalization across in-domain and out-of-domain tasks.

Achieve complex reasoning with minimal parameters using latent recursion in 2-layer networks. A 7M-parameter Tiny Recursive Model (TRM) solves Sudoku (87% accuracy), mazes (85%), and ARC-AGI with 0.01% the parameters of large LLMs via iterative latent refinement through 6+ recursive steps without fixed-point convergence requirements.

Improves inference efficiency through training-free sparse attention using global token selection patterns aggregated from local attention heads for unified cross-head ranking.