Claude Skills by ADu2021

Enables long-term consistent video generation through three-tier memory architecture combining working memory, geometry-grounded point clouds, and episodic keyframes.

Reduce inference latency and token cost in video question-answering by selectively invoking chain-of-thought reasoning. VideoAuto-R1 uses confidence-based early-exit to determine when explicit reasoning is necessary, achieving 3.3× response length reduction while maintaining state-of-the-art accuracy.

Generate minute-scale coherent videos using state-space models as evolving memory for scene dynamics. VideoSSM achieves linear computational complexity while reducing motion drift—ideal when temporal consistency matters across long video sequences.

Transform video generation models into robot manipulators by jointly predicting actions and future visual outcomes. VideoVLA demonstrates that imagining futures improves action reliability—ideal when you need generalizable robot learning from video imagination.

Implement techniques from VisGym: Diverse, Customizable, Scalable Environments for Multimodal Agents. The suite spans symbolic puzzles, real-image understanding, navigation, and manipulation, and provides flexible controls over difficulty, input representation, planning horizon, and feedback

Predict robot actions through a perception-prediction-action loop: forecast dynamic regions, depth, and semantic features from visual observations, then generate action sequences via diffusion transformers.

Transfer reasoning capabilities from text LLMs to visual domains using reinforcement learning, achieving human-level visual reasoning on complex benchmarks. 38B parameters match closed-source VLMs by optimizing cross-modal connector alignment and entropy-based reasoning signals.

Build efficient vision-language models by distilling knowledge from frozen diffusion decoders and vision encoders. Achieve GPT-4o-level captioning with <$1000 training cost by leveraging pre-trained components. Use when you need high-quality vision-language understanding without expensive end-to-end training.

Apply test-time register token injection to pre-trained Vision Transformers without retraining, eliminating high-norm outlier artifacts and improving attention map quality.

Enable Vision-Language Models to improve without human annotations through competitive multi-agent self-play games (Who-Is-the-Spy format) combined with Iterative Self-Play Policy Optimization. Use when annotation costs limit multimodal dataset scaling or domain diversity.

Enable VLMs to self-improve from unlabeled images via dual-role framework—questioner generates challenging visual questions while reasoner answers them, trained jointly with GRPO using difficulty and diversity rewards.

Implement techniques from VISTA-PATH: An interactive foundation model for pathology image segmentation and quantitative analysis in computational pathology. Accurate semantic segmentation for histopathology image is crucial for quantitative tissue analysis and downstream clinical modeling

Improve LMM performance on high-resolution images by training models to ground reasoning on image regions through RL, learning spatial localization without requiring expensive grounding annotations.

Enable multimodal models to dynamically revisit and re-ground reasoning steps in images using point-and-copy mechanisms for better long-horizon reasoning.

Enhance VLA models with spatiotemporal awareness by embedding both 3D spatial coordinates and temporal sequences: predict actions that include temporal parameters (duration) alongside spatial movements, achieving 97.4% robotic manipulation success by grounding reasoning in coherent 4D representations.

Enhance embodied robot reasoning by integrating explicit chain-of-thought supervision with reinforcement learning from verifiable rewards (GRPO+RL). Use when improving robot decision-making for tasks requiring spatial reasoning and constraint satisfaction.

Fine-tune Vision-Language-Action models using learned world models as simulators, eliminating costly real-world or physics-simulation RL. Train robust robot policies in 400 steps via GRPO with model-generated verified rewards.

Rolls forward robot state using previously generated actions to condition predictions on estimated future states, paired with temporal-offset training augmentation. Achieve 2× speedup and 17.4× latency reduction in vision-language-action models without architectural changes.

VLingNav: Embodied Navigation with Adaptive Reasoning and Visual-Assisted Linguistic Memory. From arXiv:2601.08665

Reduces video diffusion training compute 2.92× through spatio-temporal sparse attention with layer-wise cyclic block partitioning. Maintains generation quality while enabling long-sequence video training. Use when training video models on memory-constrained hardware or processing longer video sequences.

Accelerate speculative decoding by pruning drafter vocabulary to high-frequency tokens. Achieves 16% speedup in memory-bound settings by eliminating unused vocabulary entries without retraining.

Generate spatially-coherent 3D point-cloud videos from single images using depth-fused diffusion with efficient world caching for infinite scene exploration.

Unify multimodal understanding, generation, and reconstruction using high-dimensional codebooks for semantic information. VQRAE achieves 100% codebook utilization at 1536 dimensions—ideal when you need a single tokenizer for vision-language tasks.

Enable precise motion control in video generation using dense point trajectories as latent space features. Wan-Move integrates with existing models without architecture changes—ideal when you need fine-grained scene control without auxiliary motion encoders.

Convert internet tutorial videos into executable UI action trajectories using inverse dynamics models trained on 600K synthetic screen transitions. Generate 53K high-quality demonstrations spanning 69 applications, achieving state-of-the-art 7B agent performance on WindowsAgentArena via both in-context learning and supervised fine-tuning.

Break through supervised fine-tuning saturation by mixing logits from weaker model checkpoints into strong model training targets. Amplifies hard negatives that strong models have already suppressed, enabling continued learning after standard training plateaus.

Enable safe and efficient AI agent interactions with websites via declarative HTML extensions—define agent-executable tools and context using <tool> and <context> tags instead of relying on brittle UI inference.

Create principle-guided reward models for web automation agents that evaluate reasoning process quality rather than just outcomes. Implement domain-specific principles (HTML understanding, interaction patterns, state tracking) to guide agent behavior in web navigation and task completion.

Optimize open-source LLM web agent training through systematic analysis of supervised fine-tuning vs. reinforcement learning trade-offs. Achieve 45% lower compute cost by branching into RL at strategic SFT checkpoints.

Enable web agents to improve over time by curating episodic memory from navigation trajectories and dynamically injecting task-specific guidance at runtime—no retraining required, persistent improvement across sessions.

Bridge deterministic web code and generative LLMs via hybrid world models. State and physics defined by TypeScript code, LLMs generate narrative/aesthetics on top. Uses typed interfaces, deterministic hashing, graceful degradation—enabling scalable interactive environments from travel atlases to fictional worlds without databases.

Systematically train 8B web agents to solve complex information-seeking tasks through model-based exploration and long-to-short query evolution, achieving state-of-the-art performance via SFT+RL training pipeline with progressive context expansion to 128K tokens and 100 tool turns.

Training platform with thousands of realistic web navigation tasks for developing visual web agents, supporting scalable training and evaluation of agents that can generalize to arbitrary websites.

Enable autonomous web agents to navigate partially observable environments through action-aware tree search. Handle destructive actions via snapshot validation, implement speculative backtracking with parallel tabs, and dynamically prioritize actions based on safety and reversibility. Achieves 54.6% success on WebArena.

Develops multimodal research agents that combine visual and textual reasoning for complex information-seeking tasks, using synthetic training trajectories and reinforcement learning.

Train autoregressive world models on 1M+ real web interactions for accurate browser state prediction. Enables agent training with 100× more data than prior approaches, achieving GPT-4o comparable performance with format flexibility and cross-domain generalization.

Improve downstream task performance by increasing weight decay during pretraining (0.3-1.0 vs. default 0.1), enabling better model plasticity and representation structure.

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

Dynamically fork sub-agents at any step based on task requirements instead of using fixed agent counts. Linearize hierarchical trajectories into unified sequences for GRPO training. Demonstrates 6.36x more sub-agents than baselines via learned orchestration.

Train Vision-Language-Action models for robotic control through world model simulation without real-world interaction—using pixel-based world models aligned with VLA features to enable self-correction and robust policy optimization.

Accelerates iterative world model inference by classifying tokens by temporal curvature (predictability) and applying differentiated caching: stable tokens reused, linear tokens extrapolated, chaotic tokens updated. Achieves 3.7x speedup with 98% rollout quality.

Improve long-horizon world model fidelity using RL with clip-level rollouts and complementary reward functions for action accuracy and visual quality. Breaks computational constraints by evaluating candidate clips incrementally rather than full sequences, enabling efficient multi-objective optimization.

Quantify confidence in video generation by estimating latent-space uncertainty and mapping to interpretable heatmaps. Detect untrustworthy regions in generated frames for robot planning and policy evaluation—critical when video hallucinations could cause real-world failures.

Build unified world models using flow-based video generation architecture that handles Text2World, Image2World, and Video2World in a single model, trained on 200M video clips with RL post-training for improved instruction-following and video quality.

Derives closed-form optimal linear blockwise transforms combining fixed Hadamard matrices with data-dependent components from second-order statistics, providing provably near-optimal quantization for both weights and activations with GPU efficiency.

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

Use soft-prompted transformer architecture for multi-robot learning. Add learnable embodiment-specific prompt embeddings to handle different robot types while maintaining single shared backbone. Scale to 0.9B parameters across 6 simulators and 3 real robots.

Reduce KV cache memory by 7.7-10x through quantization and rematerialization of input activations instead of caching Keys and Values, trading recomputation for memory efficiency.

MoE-enhanced YOLOv9-Tiny using lightweight routers to adaptively weight specialized expert outputs at each feature resolution, improving detection quality while maintaining end-to-end differentiability. Deploy for efficient multi-scale object detection with dynamic specialization.

Build scalable LLM agent frameworks with automated generation of tools, prompts, and configurations, plus continuous optimization through experience accumulation and reinforcement learning—achieving 71.47% on WebWalkerQA and 72.8% on GAIA.