Claude Skills by ADu2021

Bridge the modality gap between language reasoning and low-level robot actions through two-stage alignment training. Use RL with action-accuracy rewards to ground linguistic reasoning in precise motor control, achieving 17.5%-106.6% improvements over SFT baselines.

Train offline safe RL agents using Hamilton-Jacobi reachability principles to learn feasibility-gated policies. Combine reward and safety critics with flow-matching teacher policies, distill to one-step actors, and calibrate safety thresholds via conformal prediction—achieving near-zero constraint violations with 2.5× inference speedup.

Route queries across frontier models using two-stage system: graph-based task discovery identifies ~332 latent task types via semantic similarity + preference patterns; MoE with task-specific adapters estimates quality for candidate models. Achieves <50% inference cost of strongest single model while exceeding its performance; applies when managing pools of frontier models with narrow capability gaps.

Optimize adapter parameter efficiency at scale by decomposing row-wise norm computation into base/cross/BA components (15× memory reduction) and fusing kernel operations. Achieves 1.5–2.0× inference speedup with 77 GB peak VRAM reduction across 8–32B vision-language models; applies when training adapter-based models with strict memory budgets across hundreds of modules.

Achieve post-hoc debiasing of frozen vision-language models by operating in sparse autoencoder latent space. Identify and modulate bias-relevant neurons with neuron-level precision while preserving task-relevant features, requiring no task-specific fine-tuning.

Enable agents to navigate toward sound-emitting objects in continuous 3D spaces with dynamic audio (intermittent sounds, silent periods). Integrate memory-augmented goal descriptors with binaural audio processing and self-motion cues to maintain goal representations even after auditory signals cease.

Add trajectory-level temporal authorization to stateless execution gates by maintaining a semantic centroid of agent behavioral profiles with exponential moving average risk accumulation. Improves F1 from 0.9756 to 1.0 and reduces false positives from 5% to 0% on slow-burn security violations. Use when deploying multi-turn agents and need to detect gradual privilege escalation and data exfiltration.

Replace standard VICReg invariance loss with graph-weighted learnable adjacency matrix to enable self-supervised no-reference image quality assessment without human labels. Improves SRCC by +0.017 (2% relative) on six-dataset average and shows stronger cross-dataset transfer. Use when training quality assessment models without paired quality labels.

Replace dense voxel tokenization with sparse 3D VQ-VAE to reduce token counts by 70% in multimodal 3D understanding. Enables efficient articulated asset decomposition for physics-based simulation. Works best for 3D generation tasks constrained by token budget. Trigger: When working with 3D mesh models and hitting token limits; want to decompose meshes into articulated parts with reduced memory footprint.

Replace fixed full-capacity slot initialization with progressive curriculum-based slot expansion to reduce over-fragmentation in video object-centric learning. Improves FG-ARI by +6.8 on YouTube-VIS and +8.3 on MOVi-C. Use when training slot attention models on videos with variable object counts and sizes.

Detect speech deepfakes by decomposing features into speaker, artifact, and context subspaces—nulling speaker information via PCA projection to isolate artifact patterns. Train minimal logistic regression classifier on orthogonally-projected representations, achieving 56% error reduction while maintaining cross-speaker and cross-TTS generalization.

Analyzes how Reinforcement Learning from Verification Rewards (RLVR) improves reasoning by examining token-level probability distributions. Finds that >83% of token positions exhibit near-zero divergence—RL operates through sparse, targeted refinements. Cross-sampling experiments show 1.5-7.8% RL-selected tokens recover full gains, while reverting 5-10% of RL tokens collapses performance. Reveals that RL primarily reallocates probability within existing candidates (80% overlap in top-k tokens...

Inject spatial understanding into VLMs via language-guided multi-turn Chain-of-Thought reasoning over hierarchical spatial knowledge (pixel→object→scene). Implement dual-channel attention mechanism preserving pre-trained vision features while progressively incorporating dense 3D spatial information. Frozen original parameters prevent catastrophic forgetting; only new channel and mixture weights update. Maintains visual fidelity while enabling precise spatial reasoning (depth, relative positio...

Accelerate agentic multimodal LLMs via speculative execution without sacrificing accuracy. Use a lightweight tool-free MLLM to predict the main model's decisions and pre-compute tool calls before the main model confirms them. Cognitive gating enables the model to self-assess confidence. Achieves 1.1-3.35x speedup with accuracy improvements up to +6.7%. Use when reducing latency in multi-step agentic workflows, have compute budget for a second model, or need to parallelize tool execution with ...

Replace multi-stream modality-specific pathways with a unified Transformer backbone processing text, video, and audio tokens in shared sequence via self-attention. Achieves superior visual quality (4.80 vs 4.76), 75% better speech clarity (14.6% WER vs 19.23%), and 80% human preference wins—particularly strong for human-centric scenarios with expressive facial performance and audio-video sync.

Design modular AI agent systems that work across multiple interaction protocols (A2A, AG-UI, A2UI, UCP, AP2) without protocol-specific rewiring. STEM Agent uses biological pluripotency as a metaphor: a generic core differentiates into specialized handlers, tool bindings, and memory subsystems. Validated with 413 tests. Use when building agents that must support diverse interfaces, require adaptive learning from user interactions, or need scalable memory management.

Replace single-pathway JEPA with a dual-temporal architecture combining dense frame sampling (fine-grained dynamics) and uniformly-sampled VLM guidance (semantic coherence) to improve egocentric trajectory prediction by 14-27% on ADE/accuracy metrics. Effective when predicting hand-object interactions where both low-level dynamics and high-level semantic context matter, and long-horizon predictions benefit from hierarchical visual representations.

Automate conversion of 630M+ heterogeneous GitHub repositories into standardized Model Context Protocol (MCP) services via hierarchical multi-agent system. Achieves 68.4% success rate after three repair cycles, 210s per repository versus 1589s manual; increases agent performance by 10.6–13.4% when integrated into existing systems. Use when scaling tool availability beyond manually curated sets.

Improve dense trajectory generation by replacing absolute coordinate encoding with grid-anchor offset encoding (reduces location variance 90%→10%), adding spatiotemporal consistency regularizers to VAE (30-40× improvement in 81-frame prediction), and using boundary-anchored fine-tuning for flow models. Effective for predicting extended trajectories (81 frames vs. prior 24-frame limits) in autonomous driving and video prediction where motion coherence across time steps is critical.

Swap fixed entailment loss with uncertainty-guided piecewise-continuous formulation to improve part-to-whole compositional alignment in hyperbolic vision-language models by +3.0-3.1% on zero-shot classification. Use when training hyperbolic VLMs on part-object relationships and need better multi-object compositional understanding.

Replace classifier-free guidance and KL-based regularization in flow matching with velocity-based MSE regularization and GRPO for joint text-image optimization. Achieves 0.8381 TextAlign and 0.90 GenEval without CFG overhead. Works best for multimodal generation where text and image must stay coherent. Trigger: When optimizing vision-language models with flow matching and need better joint text-image policy.

Discover that generative models and vision encoders share an underlying approximately Gaussian latent space (UNE). Enable controllable image editing by operating on linear directions in diffusion noise space, revealing that semantic information is present without explicit supervision.

Enable VLMs to find relevant clips in long videos through sparse observation and graph-based propagation. Iteratively hypothesize promising segments, extract multimodal evidence (captions, OCR, speech), and propagate relevance scores via visual-temporal affinity graph. Jointly model extrinsic relevance (query-to-segment matching) and intrinsic relevance (video internal structure) to infer unobserved segments. Achieves competitive accuracy with significantly fewer tokens than dense sampling.

Optimize vision-language model inference by sparsifying the interactions between vision and language tokens instead of compressing images. Uses a dynamic policy to allocate visual computation per sample based on complexity, enabling a universal network across different compute budgets. Maintains high-resolution reasoning when needed. Use when deploying VLMs under varying compute constraints, need per-sample efficiency adaptation, or want to preserve fine visual details while reducing compute.

Replace monolithic VLA single-pathway decision-making with a decomposed System-2/System-1 architecture where a pretrained VLM planner identifies targets as visual prompts (crosshairs, bounding boxes) and a VLA controller executes on grounded observations, improving success rates by 5-8% on manipulation tasks. Use when spatial precision is critical, multi-step reasoning is needed, and you have access to pretrained segmentation and vision-language models.

Replace vision-only Video-Action Models with a video-tactile fusion architecture using tactile regularization loss to prevent visual dominance, improving contact-rich manipulation success from baseline to 90% on complex tasks (80% improvement on high-precision pick-and-place). Effective when robots interact with objects requiring fine-grained force awareness, partial visual observability, or contact state transitions that vision alone cannot capture.

Build action-conditioned world models with explicit state tracking using WildWorld's 108M+ frames from Monster Hunter: Wilds. Includes data acquisition protocol with skeleton and world state annotations, quality filtering pipeline removing temporal discontinuities and cutscenes, and WildBench evaluation metrics (video quality, camera control, action following, state alignment) for assessing long-horizon consistency and state-aware predictions.

Replace zero-order hold assumptions with perception-constrained approximation via four-module caching system: motion-aware skip thresholds, saliency-weighted drift, least-squares blending, and adaptive scheduling. Achieves 2.1–2.3× speedup at 2B scale with 99.4–99.6% quality retention on Cosmos-Predict video generation; apply when optimizing diffusion world model inference under real-time constraints.