Claude Skills by ADu2021

Unified text-image token processing in a compact 6B-parameter transformer enabling sub-second inference on enterprise GPUs through hierarchical distillation and single-stream architecture. Use when generating high-quality images with tight latency budgets or limited GPU memory.

Generate diverse human motions from text descriptions using million-scale datasets and wavelet-enhanced quantization. Achieves state-of-the-art zero-shot generalization on out-of-domain motions, compositional descriptions, and complex choreography through efficient tokenization and scalable transformer-decoder architectures.

Reduce token count by 15-40% at inference through context-adaptive compression, merging frequent token sequences into hypertokens using online Lempel-Ziv-Welch compression without retraining entire models.

Replace monolithic 3D Gaussian Splatting with two-expert architecture separating geometry (pose) estimation from appearance synthesis. Converges 30× faster (5K vs 150K iterations) while matching pose-dependent methods. Works best for multi-view reconstruction when geometry and appearance have conflicting optimization dynamics. Trigger: When doing feed-forward 3DGS and need faster convergence without sacrificing quality.

Replace standard likelihood-based video diffusion training with decoupled physics discriminators and DPO post-training to suppress physically implausible behaviors (object penetration, anti-gravity motion) in robotic manipulation videos. Use when generating physics-realistic video predictions for embodied AI and want to maintain visual quality without physical violations.

Replace standard supervised fine-tuning loss with a dual-objective loss that jointly optimizes over both concrete instances and their abstract representations (entity-masked versions), eliminating need for replay buffers and improving cumulative accuracy by 2-5% on continual learning benchmarks. Use when streaming data contains latent relational structure, catastrophic forgetting is problematic, and you want to maintain structural understanding without memory overhead.

Establish regret bounds for behavior cloning with discretized actions combining statistical error and quantization error terms. Prove smoothness requirements for safe quantizer design, show that learning-based quantizers fail these requirements, and propose model-based augmentation to reduce error dependence from H² to H.

Dynamically allocate LoRA ranks per-layer during fine-tuning instead of using fixed uniform ranks. Learn optimal rank for each layer and subject via variational framework with discretized exponential distribution, reducing memory footprint while maintaining fidelity and text-alignment.

Adapt general LLMs to specialized manufacturing domains via domain-adaptive pretraining on open-access journals and visual instruction tuning. Extract 50M tokens and 24K images from peer-reviewed papers, achieve >90% accuracy on domain knowledge tasks, and enable real-time defect identification from manufacturing images.

Future intelligence explosions will be plural, social, and entangled with humanity through distributed collaborative systems rather than singular superintelligence. Intelligence is inherently social, demanding infrastructure matching agent development; integrate governance, institutional frameworks, and constitutional checks across hierarchies of autonomous agents and human-AI centaurs in shifting configurations.

Automate systematic literature reviews in epidemiology using agentic AI pipelines. Achieves 58x speed-up (7 weeks to 20 hours) by automating article retrieval, screening, data extraction, and report synthesis. Demonstrates that review quality depends on model capabilities rather than scale. Use when conducting evidence-based reviews in specialized domains, need to validate against human expertise, or require cost-effective evidence synthesis at scale.

Build taxonomy-aware audio-text pretraining systems for species recognition from animal vocalizations. Train contrastive models that augment text prompts with hierarchical taxonomic structure (scientific/common names, phylogenetic sequences), evaluate on unseen species via rare-species test sets, and predict ecological traits directly from audio.

Reduce video token overhead by 4-100x through autoregressive patch selection, enabling MLLMs to process 1K-frame 4K video efficiently. Uses next-token prediction to identify multi-scale patches that matter. Achieves 19x speedup on vision transformers. Use when processing long, high-resolution videos with MLLMs, have budget constraints on tokens/compute, or need to handle 4K resolution at scale.

Address hallucinations in LLM QA over black-box knowledge graphs using evidence-driven retrieval. Formalize Optimal Informative Subgraph Retrieval and employ bubble expansion to discover candidate evidence graphs, achieving state-of-the-art multi-hop QA performance.

Coordinate multiple LLM agents via CAID framework: centralized task delegation, asynchronous execution in isolated git worktrees, structured integration through git merges. Branch-and-merge with worktree isolation yields +26.7% absolute improvement on PaperBench; ranked strategies show structured JSON communication and dependency-aware delegation outperform soft isolation.

Craft stealthy vehicle appearances that fool detectors by formulating attacks as conditional image editing. Apply image-level stylization (match surroundings) and scene-level strategies (match semantic concepts), achieve 38% AP50 reduction with 85%+ human perceptual success, and transfer to black-box detectors.

Establishes Active-Vision Foundation Models (AVFM) as a new problem class and proposes CanViT: a retinotopic ViT backbone with Canvas Attention that decouples thinking (glimpse processing) from memory (scene canvas). Dense latent distillation from DINOv3 enables unsupervised pretraining on 1B random glimpses. Achieves 81.2% ImageNet accuracy with frozen probes—proving foundation models can be adapted to active-vision tasks. Trigger: When building systems requiring selective visual attention t...

Overturn the assumption that scaling foundation models improves RNA structure prediction by understanding why they fail out-of-distribution. Includes structure-aware deduplication revealing 33-fold residual redundancy in prior benchmarks, out-of-distribution test regimes (GenA, GenC, GenF), and root cause analysis showing coverage and wiring failures. Foundation models achieving 67.3% on held-out test drop to 18.0% OOD (26.7% retention), while structured decoders retain 92.3%, enabling practi...

A single insight reframes optical flow in degraded video as correspondence matching via diffusion features: restoration models naturally encode degradation patterns while preserving geometry. Add temporal reasoning via cross-frame attention to enable zero-shot correspondence without task-specific training. Trigger: When matching pixels across corrupted frames, apply diffusion features with temporal awareness to estimate optical flow without labeled degraded-video data.

Evaluate data agents on realistic multi-database queries across 54 tasks spanning 12 datasets, 9 domains, and 4 DBMS systems. Reveals frontier models achieve only 38% pass@1 accuracy, with 85% of failures from incorrect planning rather than data selection. Benchmark captures critical challenges: heterogeneous database integration, identifier reconciliation, unstructured text extraction, and domain knowledge. Identifies optimal exploration ratio (~20% of operations) and cost-efficiency differe...

Comprehensive recipe for RL-training tool-using agents spanning reward design, data synthesis, model scaling, and algorithm selection. Seven ranked findings: scale-dependent rewards (curriculum for 1.5B–3B; dense for 7B), semi-sparse 'Macro' rewards balance specialization/transfer, 1K-sample sweet spot with 4:3:3 difficulty mix. Achieves SOTA on TravelPlanner with smaller models than leading proprietary systems.

Evaluate web agents on tasks grounded in first-person video perception with Ego2Web benchmark. Bridges egocentric video understanding and web agent execution across e-commerce and knowledge lookup domains. Includes Ego2WebJudge LLM-based evaluator achieving 84% human agreement, replacing unreliable domain-specific metrics. Reveals weak agent performance on physically-grounded tasks, highlighting the necessity of accurate multimodal visual understanding.

Enable spatially adaptive Gaussian allocation in 3D Gaussian Splatting without iterative optimization. Reduces primitive count by 70-90% while maintaining reconstruction quality through learned densification score prediction.

Introduce Quasicryth, a text compressor using Fibonacci quasicrystal tilings for phrase-level compression. Prove that aperiodic structures never structurally collapse at depth, enabling compression at arbitrary hierarchy levels. Bridge quasicrystal mathematics with practical compression, achieving 22.59% ratio on enwik9 with unbounded scaling advantages.

Replace self-attention world models with reaction-diffusion PDEs as the predictive substrate. Demonstrate that aperiodic PDE dynamics achieve superior multi-step rollout stability and inherent error correction through Laplacian diffusion smoothing, while maintaining O(N) complexity and enabling autonomous corruption recovery.

Unify discrete diffusion for language by replacing token-wise uniform masking with generalized noising via rate matrices and snapshot latents, enabling semantic-aware forward processes. Demonstrates 1.16 BPC on Text8 and 7.65 perplexity on OpenWebText, beating autoregressive baselines; enables generalized noising processes over arbitrary vocabularies with efficient training aligned to standard architectures.

Replace traditional VAE latent spaces with frozen geometric foundation model encoders (e.g., Depth Anything 3) as diffusion latent space. Leverage strong cross-view geometric correspondences for multi-view consistency. Use cascaded generation up to optimal boundary layer, then deterministically derive deeper features from frozen encoder, resulting in 4.4× faster convergence with superior geometric consistency compared to VAE-based approaches.

Enforce semantic compatibility constraints directly into instance construction for 3D object detection. Uses MLLM-driven semantic grouping to partition object vocabulary into plausible cross-view category equivalence clusters, preventing geometry-driven over-merging.

Extend visual reasoning to spatial-temporal sequences via two-agent reasoning+summary pipeline with ST-GRPO (temporal alignment) and J-GRPO (evaluative robustness) algorithms. Achieves +8.1% on image and +6.9% on video benchmarks through autonomous data generation and self-evolving rewards; enables continuous improvement loops for visual reasoning without human annotation.

Navigate LLM agent workflow design by understanding the taxonomy of static vs dynamic methods and the dimensions that organize them. Agentic Computation Graphs (ACGs) framework distinguishes when structure is determined (before vs during execution), which components optimize, and what signals guide optimization. Provides structure-aware evaluation criteria combining downstream metrics with graph properties and robustness. Use when designing or optimizing agent workflows, choosing between fixe...

Integrate agentic tool interaction (Lean4 compiler, syntax checkers) with curriculum-based RL for formal reasoning. Replace standard importance sampling with Hierarchical Importance Sampling Policy Optimization (HisPO): sequence-level masking removes train-inference discrepancies, token-level masking filters inconsistent tokens, staleness control manages policy drift. Achieves 97.1% auto-formalization (vs 83% baseline), 95.5% MiniF2F-Test (72 attempts vs 1,024+), 70.8% ProverBench.

Add tool-calling interface for on-demand high-resolution crop retrieval, enabling VLMs to first observe low-resolution global view then selectively request detailed crops. Achieves 80.3% of full-res performance (vs 80.46%) with only 36% of visual tokens, reducing wall-clock latency 4.4× (from 2.71s to 0.61s) through KV-cache reuse without architectural changes.

Constrain video GRPO policy updates to stay within pre-trained model's data manifold using dual-control exploration. Implement precise manifold-aware SDE with logarithmic noise variance correction (captures geometric signal decay standard methods miss). Apply gradient norm equalizer to balance learning across diffusion timesteps (mitigate vanishing/exploding gradients). Use dual trust region combining position control (anchored exploration) and velocity control (KL constraints) for stability-...

Enhance diffusion language model performance on long-context tasks by embedding simulated denoising into training via bi-level optimization. Fast weights capture local trajectory experience; base model optimized with accumulated parametric memory. Achieves +17.0% on RULER Variable Tracking (8K) and +9.6% on BABILong with gains primarily from training-stage improvements.

Replace autoregressive token-by-token OCR decoding with block-wise diffusion decoding to achieve 3.2× speedup while maintaining 99.9% accuracy. Works best for document layout parsing where parallel generation is feasible. Trigger: When optimizing OCR systems and want faster inference without accuracy loss.

Empirical analysis revealing that LLMs produce post-conventional moral reasoning (Kohlberg Stages 5-6) regardless of size or prompting—inverse of human developmental patterns (Stage 4 dominant). Finds moral ventriloquism: models acquire rhetorical conventions of mature moral reasoning without developmental trajectory. Key evidence: action-justification decoupling (models produce Stage 5+ vocabulary while selecting Stage 2-3 actions), identical responses to semantically distinct dilemmas (ICC ...

Identify three ranked findings on multi-task SFT: (1) heterogeneous overfitting—sub-datasets peak at different training points (contradicts uniform duration practice); (2) parameter divergence—excluding 1/10 of data shifts optimal points 0.91 epochs for remaining tasks; (3) SFT compute negligible (0.01% of training). Implement mSFT: iterative roll-out/roll-back search per-dataset. Robust across 0.5B-8B models, 9K-27K samples, 5-15 tasks, achieving +3.4% improvement with reduced FLOPs.

Evaluate multi-reference image generation fidelity using MultiBind's dimension-wise confusion framework. Detects cross-subject attribute errors that holistic metrics (FID, CLIP) miss, including drift (degradation), swap (permutation), dominance (interference), and blending (averaging). Protocol uses specialist models for face identity, appearance, pose, and expression; achieves reproducible failure diagnosis revealing severe binding failures in models appearing competitive on aggregate quality.

Evaluate world models on faithful interactive response capability through three hierarchical interaction levels (single object, localized, multi-object effects) and four complementary metrics measuring video quality, controllability, and physical plausibility. Identifies the gap between visual fidelity and true interactive state transition modeling.

Replace live web API-dependent research with offline corpus-based trajectory synthesis. Decouples answer-guided document retrieval (10K gold + 15M FineWeb) from synthesis via local search engine, eliminating $5,760 Serper costs while enabling reproducible, analyzable reasoning chains through three primitives: Search (ranked retrieval), Open (full document fetch), Find (intra-document verification).

A single insight eliminates multi-view requirements for novel-view synthesis: monocular depth acts as a training-time geometric scaffold to generate synthetic view pairs from unpaired internet images, but can be discarded at inference. This reframes the problem from needing paired multi-view data to leveraging abundant 2D internet imagery. Trigger: When limited to monocular video or single-image novel-view synthesis, use depth as training scaffold on unpaired data—the model learns geometry wi...

Defines Personalized Streaming Video Understanding (PSVU) as a new task bridging static image personalization and video intelligence. PEARL-Bench (132 videos, 2173 annotations) establishes founding experiments. PEARL framework provides training-free plug-and-play strategy using dual-grained memory (concept + streaming) and concept-aware retrieval. Trigger: When building AI assistants that interact with streaming video and personalize on user-defined concepts, apply this problem definition and...

Replace uniform token-level advantages with perception-exploration gating that weights tokens by visual grounding strength. Adds 3.67 points to geometry reasoning and 5.32 to few-shot classification with <1% compute overhead. Works best for multimodal CoT where visual grounding anchors reasoning. Trigger: When using token-level RL on VLMs and want to emphasize visually-grounded reasoning steps.

Enhance VLM spatial grounding by enforcing explicit intermediate spatial reasoning before text generation. Generate segmentation and depth tokens as perception pathways, anchoring responses in 2D and 3D geometric reasoning for improved referring expressions and spatial tasks.

Achieve high-accuracy agentic post-training with 4x fewer rollout turns and 5.5x less wall-clock time than end-to-end RL. Identify informative intermediate turns via pivot filtering and use verifier-based functional rewards for precise credit assignment.

Eliminate hallucination via four-stage progressive training: multilingual adaptation → English dialogue SFT → bilingual SFT → GRPO alignment. Achieve 0.0% hallucination rate for encoder-decoder models using structured citation markers and knowledge-source attribution, with automatic transfer of citation format across languages.

A single insight reframes sim-to-real video generation as decoupled structure-and-appearance transformation: use geometric conditioning to preserve structural fidelity while allowing free appearance transformation. This enables IC-LoRA training on pseudo-paired synthetic-real data constructed via sparse-to-dense propagation, eliminating the need for aligned real-world video capture. Trigger: When converting rendered 3D output to photorealistic video, apply structure-aware appearance transform...

Understand the policy challenges of governing autonomous AI agents under existing frameworks like the EU AI Act. Identifies three primary governance failures: performance failures during execution, misuse risks from malicious deployment, and economic inequality in agent access. Argues that regulations designed for static AI systems inadequately address agent autonomy. Use when understanding regulatory implications of agent deployment, advocating for policy changes, evaluating governance readi...

Enable LLM agents to autonomously improve on research-code tasks through reflective learning from execution trajectories. Distill recurring failure patterns into actionable heuristics applied via targeted prompt edits, improving performance 3.51%-4.89% while maintaining 10x better cost-efficiency.

Analyze reinforcement learning with verifiable rewards using signed log-probability differences to identify reasoning-critical tokens. Reveals that RLVR updates concentrate on low-probability tokens, enabling test-time amplification and training-time reweighting techniques.