Claude Skills by ADu2021

Build LLM-driven data science agents grounded in empirical knowledge through expert knowledge base, tree search algorithms, and complexity-adaptive code generation, surpassing SOTA by 8% on MLE-Bench.

Build autonomous research agents using pre-computed knowledge graphs instead of online reasoning. Extract methodological patterns from literature, organize them into structured knowledge, and enable agents to align user research intents with established paradigms for efficient, grounded research planning and execution.

Autoregressive U-Net operating directly on raw bytes with hierarchical multi-scale pooling for adaptive token embedding, eliminating fixed vocabularies.

Generate synthetic web environments at scale by specifying websites as Finite State Machines with explicit state transitions, then programmatically executing GUI actions to collect verified interaction trajectories. Reduces trajectory cost from $0.15–$1.00 to $0.04 per sample while generating 11,000+ verified trajectories with deterministic, executable validation requiring no external judges.

Avey architecture pairs a ranker with autoregressive processor to select relevant tokens, decoupling context window from sequence length for efficient long-range processing.

Accelerate agentic AI training by distributing task execution across clusters, achieving 14.6x speedup in experience collection and enabling practical large-scale agent development

Stabilize off-policy RL for LLMs using adaptive clipping that dynamically rebalances positive/negative gradients and preserves entropy, improving mathematical reasoning performance vs standard PPO.

Fixes batch speculative decoding ragged tensor problem where sequences in batches accept different token counts, desynchronizing state. EQSPEC guarantees output equivalence through proper synchronization. EXSPEC reduces overhead 40% via cross-batch scheduling. Enables efficient parallel decoding with 95% equivalence.

Allocate LLM reasoning budget optimally via value tree search: use residual value prediction to estimate step utility, then dynamically shift exploration-exploitation balance as budget depletes. Outperform high-budget baselines at 1/4 cost.

Compress long prompts to 1/26th of original size while maintaining retrieval accuracy using hierarchical page-level pooling, without model fine-tuning.

Reinforcement learning (RL) is central to post-training, particularly for agentic models that require specialized reasoning behaviors. In this setting, model merging offers a practical mechanism for integrating multiple RL-trained agents from different tasks into a single generalist model. However, existing merging methods are designed for supervised fine-tuning (SFT), and they are suboptimal to preserve task-specific capabilities on RL-trained agentic models. The root is a task-vector mismat...

We introduce Being-H0.5, a foundational Vision-Language-Action (VLA) model designed for robust cross-embodiment generalization across diverse robotic platforms. While existing VLAs often struggle with morphological heterogeneity and data scarcity, we propose a human-centric learning paradigm that treats human interaction traces as a universal 'mother tongue' for physical interaction. To support this, we present UniHand-2.0, the largest embodied pre-training recipe to date, comprising over 35,...

Systematic evaluation toolkit for assessing large language models across multiple dimensions, enabling comprehensive benchmarking of agent capabilities and comparative analysis of model performance.

Align diffusion models to hierarchical fine-grained criteria rather than binary preferences. Decompose expert knowledge into attribute hierarchies and apply Complex Preference Optimization to simultaneously maximize positive attributes while minimizing negative ones.

Improve long-context performance by incorporating imaginary components discarded in standard RoPE implementations. Use phase information from complex-valued attention for richer positional encoding—especially valuable as context length increases beyond normal ranges.

Generate high-quality synthetic training data that enables 7.7x faster training than web data, with smaller models achieving better performance through strategic content rephasing and data optimization.

Connects multimodal language models with diffusion models using patch-level CLIP embeddings as shared latent variables, enabling controllable image generation with minimal training overhead.

Build fully ternary quantized vision-language-action models for robotic manipulation, achieving 11x memory reduction and 4.4x speedup while maintaining task performance on edge devices.

Improve computer-use agent performance by running multiple rollouts and selecting the best trajectory using narrative-level reasoning. The Behavior Judge (BJudge) converts raw execution traces into behavior narratives, enabling intelligent trajectory selection that scales agent effectiveness beyond single-rollout limitations.

Deploy efficient MoE models on resource-constrained edge devices by learning chunk-level activation sparsity that achieves 3.67× speedup. Use when you need to compress LLMs for on-device inference while maintaining reasoning quality and supporting speculative decoding acceleration.

Improve credit assignment in multi-objective RL by decomposing advantages into segment-specific values. Use Outcome-Conditioned Baselines to reduce cross-objective interference without expensive rollouts, enabling better training signals for multi-step completions with different reward functions per segment.

Optimize language model policies layer-by-layer rather than monolithically to understand internal reasoning structure. Decompose models into per-layer and per-module policies via residual streams, analyze entropy patterns revealing exploration→convergence phases, and optimize layers sequentially—improving reasoning on math tasks by up to 4.69 points.

Overcome reasoning model training plateaus by increasing rollouts per prompt (N=512) rather than training steps, addressing unsampled coupling that destabilizes learning. Theoretical analysis shows broad exploration eliminates plateau bottleneck.

Build web agents using human-inspired browser actions (scrolling, clicking, typing) operated directly on raw HTML via Playwright. Combine supervised fine-tuning and rejection fine-tuning with explicit memory for strong generalization on web tasks.

Automates artistic typography customization through self-distilled learning and localized style injection. Generates stylized text images by encoding reference style and injecting it into diffusion denoising. Use for digital design workflows, text-based visual content creation, or applications needing artistic typography control without manual annotation.

Replace discrete token prediction with continuous vector prediction by training a high-fidelity autoencoder to compress K tokens into single latent vectors, enabling K-fold sequence length reduction while maintaining likelihood-free generation through energy-based scoring rules.

Adapt large reasoning models for optimization tasks using expert-guided hint correction. Generate high-quality training data with minimal expert intervention (<2.6% token modification). Trigger: fine-tune reasoning models on domain-specific tasks without large supervised datasets.

Protects computer use agents from prompt injection by using single-shot execution planning that generates complete control flow graphs before UI observation, preventing instruction hijacking while maintaining 57% performance on frontier models.

Implement techniques from Can LLMs Clean Up Your Mess? A Survey of Application-Ready Data Preparation with LLMs. Data preparation aims to denoise raw datasets, uncover cross-dataset relationships, and extract valuable insights from them, which is essential for a wide range of data-centric applications

Generate images with unified control over identity, spatial position, pose, and layout: encode diverse control modalities (spatial canvas, pose canvas, box canvas) into single RGB image, train diffusion model jointly across all control types, and enable flexible multi-modal composition at inference without task-specific fine-tuning.

CapImagine teaches models to explicitly imagine through text rather than latent reasoning, significantly improving visual reasoning performance.

Enhance LLM reasoning by combining contrastive learning on reasoning representations with reinforced fine-tuning, leveraging both annotated chains and unsupervised signals.

Train reusable pre-computed KV cache representations of large text corpora for efficient retrieval, achieving 38.6x memory reduction and 26.4x throughput improvement.

Replace token-insertion for fusing vision and language with efficient cross-attention that maintains separate text self-attention. Enables text tokens to attend images within local windows, preserves gist tokens from prior images, and maintains near-constant memory costs for streaming video—more practical than direct token insertion for resource-constrained applications.

Defend against indirect prompt injection attacks by detecting dominance shifts using leave-one-out attribution, enabling selective sanitization without sacrificing latency or utility.

Implement CASTLE, a causal attention mechanism that dynamically updates key representations as context expands. Reduces validation loss by 0.006-0.037 across model scales while maintaining O(L²d) training complexity and O(td) decoding speed. Deploy for improved language model perplexity without inference overhead.

Extract sparse causal concept graphs from LLM activations using SAE and DAGMA, then validate through ablation to identify causally influential features. Bridges mechanistic interpretability with causal inference for understanding reasoning flow.

Control policy entropy dynamics in RL by reweighting gradients from clipped tokens. CE-GPPO preserves out-of-clip gradients with beta parameters to stabilize exploration-exploitation balance, preventing entropy collapse while maintaining training stability in LLM fine-tuning.

Implement techniques from CGPT: Cluster-Guided Partial Tables with LLM-Generated Supervision for Table Retrieval. General-purpose embedding models have demonstrated strong performance in text retrieval but remain suboptimal for table retrieval, where highly structured content leads to semantic compression and query-table mismatch

Train single models to simulate multi-agent collaboration through distillation from complex multi-agent systems and agentic RL, creating efficient Agent Foundation Models for tool use and web navigation.

Enable language models to dynamically switch between four cognitive modes (spatial, convergent, divergent, algorithmic) during problem-solving. Meta-agent observes state and selects optimal mode per step, improving reasoning across math, coding, and spatial tasks without requiring model training.

Analyze when CoT reasoning succeeds or fails using DataAlchemy synthetic environment and distribution discrepancy measurement.

Improve diffusion model sampling by planning content-adaptive denoising trajectories. Extract Diffusion DNA signatures quantifying per-stage difficulty, then apply graph planning to allocate computation to challenging generative phases.

Train search agents using citation-aware rubric rewards that decompose complex questions into verifiable single-hop facts. Agents learn to chain evidence through explicit source citations, preventing hallucinations and shortcut exploitation. Citation-aware Group Relative Policy Optimization (C-GRPO) combines rubric and outcome rewards, enabling agents to solve multi-hop reasoning tasks with high factual grounding and transparency.

Implement techniques from ChartVerse: Scaling Chart Reasoning via Reliable Programmatic Synthesis from Scratch. Chart reasoning is a critical capability for Vision Language Models (VLMs)

Harmonize supervised fine-tuning and reinforcement learning through dynamic weighting, balancing expert imitation and on-policy exploration to prevent response pattern disruption.

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

Unify retrieval and generation in RAG systems by compressing documents into shared continuous embeddings that serve both retrieval and generation: implement joint training with differentiable selection, achieving up to 16× context compression while improving generation quality.

To teach robots complex manipulation tasks, it is now a common practice to fine-tune a pre-trained vision-language-action model (VLA) on task-specific data. However, since this recipe updates existing representations, it is unsuitable for long-term operation in the real world, where robots must continually adapt to new tasks and environments while retaining the knowledge they have already acquired. Existing continual learning methods for robotics commonly require storing previous data (exempl...

Accelerate LLM inference by dynamically skipping transformer layers based on in-context signals, without training auxiliary draft models or changing model weights.