Claude Skills by ADu2021

Uses LLMs with RL to automatically optimize HGEMM CUDA kernels across 1,000 configurations, systematically outperforming NVIDIA's cuBLAS and cuBLASLt through continued pretraining, general RL, and specialized HGEMM RL stages.

Improve code and test generation through co-evolution where LLMs generate both solutions and tests, optimizing each based on mutual evaluation and discriminative testing performance.

While large language models (LLMs) have shown to perform well on monolingual mathematical and commonsense reasoning, they remain unreliable for multilingual medical reasoning applications, hindering their deployment in multilingual healthcare settings. We address this by first introducing CUREMED-BENCH, a high-quality multilingual medical reasoning dataset with open-ended reasoning queries with a single verifiable answer, spanning thirteen languages, including underrepresented languages such ...

Improves reasoning efficiency through curriculum learning that progressively constrains token budgets, enabling models to first discover solution strategies then distill them into concise traces.

Framework for training cybersecurity agents without access to live environments. Uses CTF writeups and persona-driven LLM simulation to synthesize training trajectories, achieving performance matching proprietary systems like Claude-3.5-Sonnet.

Mitigate lazy reasoning in Large Reasoning Models via self-distillation teaching task decomposition, followed by Diversity-Aware GRPO with entropy-based advantage functions, enabling effective decomposition without external teachers while balancing structured reasoning with diversity.

Enable autonomous agent self-improvement through evolutionary mutation of agent codebases, using LLM-generated variants and empirical validation to discover beneficial modifications like enhanced tools and context management.

Accelerate the Shampoo optimizer 4.8x using batched block-wise preconditioning and numerical approximations, enabling more frequent preconditioner updates without computational bottleneck.

Boost language model performance by strategically ordering training data without changing content or model size. Uses learnability-quality scoring and folding schedules to improve convergence and knowledge retention, achieving consistent gains across all model scales.

Train efficient robot manipulation policies by strategically applying task diversity and debiasing expert demonstrations to remove execution speed variations that degrade learning.

Automatically synthesize and optimize training data using GRPO to generate data recipes (specifications for dataset creation). Use a Data Verifier to efficiently evaluate sample quality without full model training. Achieve performance comparable to human expert curation across diverse domains.

Accelerate video generation by 14.8x through deep compression autoencoder (32x-64x spatial, 4x temporal compression) combined with lightweight adapter-based model adaptation. Use when deploying video diffusion models under compute or latency constraints.

Accelerate video generation through dual-expert consistency distillation, using separate denoisers for semantic layout/motion and detail refinement to resolve conflicting optimization gradients.

DCPO eliminates zero-gradient dead zones in policy optimization by adaptively adjusting token-level clipping bounds based on prior probabilities and smoothing advantage standardization across cumulative training steps, achieving 28% improvement in effective response utilization and 10x reduction in token clipping ratio on mathematical reasoning benchmarks.

Accelerate diffusion transformer inference by dynamically adjusting patch granularity during generation based on detail complexity at each timestep. Early denoising steps (establishing low-frequency structure) use coarse patches; later steps (adding high-frequency detail) use fine patches. Achieves 3.52× speedup on FLUX-1.Dev and 3.2× on video models while maintaining quality through variance-based adaptive scheduling.

Bridge the gap between informal mathematical reasoning (80% accuracy) and formal proof synthesis (8% success) by decoupling them: a general-purpose reasoner generates strategic lemmas, then a specialized prover verifies them formally. First open-source solver of post-2000 IMO problems.

Enables autonomous reasoning agents to discover and invoke tools efficiently through end-to-end training. Uses autonomous memory folding to compress interaction history and ToolPO to learn general-purpose tool use, applicable across diverse benchmarks from QA to web automation.

Maintains half of sliding window as attention sinks with dynamic temporal RoPE alignment plus importance-aware KV cache pruning, enabling 12× extrapolation beyond training length (60+ seconds from 5-second training) without fine-tuning.

Enhances model safety by filtering dual-use topics from pretraining data, creating tamper-resistant models robust to adversarial fine-tuning without degrading unrelated capabilities.

Reduce redundant tokens in parallel reasoning by 80% while maintaining accuracy via dynamic pruning of equivalent reasoning paths. Trigger: improve efficiency of consensus-based reasoning (multiple CoT generation).

DeepResearchEval: An Automated Framework for Deep Research Task Construction and Agentic Evaluation. From arXiv:2601.09688

Monitor search agent reasoning quality via hierarchical uncertainty detection. Fast consistency checks identify anomalies; slow experience-driven feedback provides corrections. Minimal overhead while catching misalignment.

Overcome exploration bottlenecks in reasoning RL by integrating Monte Carlo Tree Search during training (not just inference). Global frontier selection and entropy-guided sampling reduce GPU hours by 5.7x while improving performance.

Build research agents that systematically search for comprehensive answers to complex questions by maintaining search state, iterating on queries, and validating answer completeness. Implement strategies for identifying knowledge gaps and conducting follow-up searches to ensure thorough coverage of topics.

Transform research specifications into production-grade codebases through strategic information management and autonomous agent orchestration. DeepCode surpasses PhD experts and commercial tools—critical when you need scientific code reproducibility at scale.

Filter low-quality reasoning traces using model-internal confidence signals at test time, eliminating weak paths during generation to achieve 99.9% accuracy while reducing token generation by up to 84.7%.

Train multimodal agents to dynamically invoke tools (code execution, web search) within reasoning loops through a two-stage pipeline combining cold-start supervised learning with reinforcement learning—enabling task-adaptive tool invocation for perception, reasoning, and retrieval tasks.

Synergistic verifier-generator training loop enabling LLMs to identify logical issues in mathematical proofs without reference solutions, improving reasoning rigor through meta-verification. Apply when you need to scale mathematical reasoning without hand-labeled proof annotations.

Combines DeepSeek Sparse Attention (DSA) achieving O(Lk) complexity, scalable RL framework allocating 10%+ compute to post-training, and large-scale agentic task synthesis with 1,800+ environments. Matches proprietary models in reasoning; DeepSeek-V3.2-Speciale achieves olympiad gold.

Enable efficient speculative decoding by training discrete diffusion language models for parallel draft generation. Use AR-style distillation and scribe refinement to train dLLMs. Eliminate left-to-right error accumulation through independent parallel proposals. Achieve 5.54× speedup on HumanEval vs. 2.41× for AR-based methods.

Improve diffusion model alignment by assigning step-wise rewards during denoising instead of terminal rewards. Fixes sparse reward signal mismatch in multi-step generation processes through ODE-based reward estimation.

Extend monocular depth models to handle adverse conditions (weather, darkness, sensor noise) using only 540K training samples. Applies perturbation-based consistency learning and spatial constraints to maintain robust depth prediction across challenging real-world scenarios.

Replace LayerNorm with Derf(x) = erf(αx + s) for improved generalization in transformers. Derf outperforms LayerNorm across vision, speech, and DNA modeling—ideal when normalization-free training provides benefits without architectural complexity.

Accelerate LLM inference 6x by using block diffusion for parallel token drafting with tight coupling to the target model's hidden representations, achieving higher speedups than existing speculative methods without quality loss.

Unlocks parallelism in recurrent memory transformers through diagonal batching of the layers-segments grid, achieving 3.3x speedup on 131K-token sequences without model retraining.

Train diffusion language models to generate optimized CUDA kernels using bi-phase reinforcement learning. First phase masks and regenerates core kernel logic with provided scaffolding to prevent PyTorch shortcuts. Second phase enables end-to-end generation. Leverage diffusion's global context awareness for non-sequential code generation.

Recast chain-of-thought reasoning as iterative denoising using diffusion principles to overcome exposure bias in autoregressive reasoning. DiffCoT enables retrospective refinement of intermediate steps while maintaining temporal consistency through causal noise scheduling.

Improve state-space model performance on language modeling and retrieval by applying differential mechanisms to reduce noise in intermediate representations and enhance long-context reasoning.

Apply diffusion models as native generative agents for vision-centric reasoning tasks (sequential planning, constraint satisfaction, spatial configuration) instead of text-based LLM chains. Achieves 3x+ improvements over GPT-5 and Gemini-3 on visual reasoning. Use when image-to-image generation better captures the reasoning constraints than text-based problem decomposition.

Train masked diffusion models for code generation using coupled-GRPO to optimize non-autoregressive generation. Achieves 4.4% improvement on code benchmarks while reducing autoregressive bias.

Extract fine-grained visual features from pretrained text-to-image diffusion models to improve multimodal LLMs beyond CLIP, achieving up to 6% gains through strategic fusion of unconditional and conditional diffusion representations.

Convert pre-trained autoregressive vision-language models into diffusion VLMs without architectural modifications. Use block diffusion strategy enabling arbitrary-length generation and KV-cache reuse. Hybrid attention enforces bidirectional within blocks, causal between blocks. Requires less than 5% of data compared to prior diffusion VLM methods.

Research contribution advancing agent and reasoning capabilities through novel approaches to model development, training, and evaluation.

Apply reasoning-focused RL to vision foundation models using GRPO to develop deep visual understanding and abstract reasoning beyond visual recognition.

Improve diffusion model capacity efficiency by directly predicting clean data instead of noise—leverage the manifold assumption that natural data occupies low-dimensional space while noise spans full dimensionality.

Optimize diffusion models for preference alignment by learning directly from group-level preferences without stochastic policies, enabling efficient and stable training.

Enable effective RL for diffusion language models via DiPO (unbiased GRPO for dLLMs) and framework optimizations. FlexAttention accelerates blockwise training, LMDeploy optimizes inference, achieving training-inference consistency—improving dLLM math performance to rival larger autoregressive models.

Enables diffusion LLMs to achieve 2.5× faster inference than autoregressive models through block-wise generation with parallel inter-block decoding.

Reformulate discrete diffusion using signal-to-noise ratio for hybrid masked-uniform noise scheduling. Derive compute-optimal scaling laws through careful hyperparameter tuning, showing uniform diffusion scales more favorably in token-constrained settings than autoregressive models.

Understand fundamental limitations of applying diffusion to discrete text: position-agnostic corruption ignores linguistic structure, and token-wise training misses multi-token dependencies. Design text diffusion systems satisfying five essential properties: position-aware corruption, dependency-aware training, parallel consistency, linguistic structure respecting, and robust handling of token boundaries.