Claude Skills by ADu2021

Allow attention heads to reference features from multiple depths by accessing both current-layer and depth key-value pairs. Prevent signal degradation in deep models while maintaining computational efficiency.

Optimize Mixture-of-Experts efficiency by decoupling token-level expert activation from layer architecture. Use dynamic threshold routing where expert count per token varies by input complexity, and apply layer-wise capacity scheduling to match representational diversity patterns.

Enable LLMs to autonomously select and apply diverse reasoning strategies without prompt engineering. Trains models with diverse thought templates covering 50-500 distinct reasoning approaches, achieving 2-13% improvements over baseline prompting methods.

Build parameter-efficient models that assign different computation depths per token via learned routing, combining weight sharing with dynamic computation complexity. Use when you need to maximize model capacity within compute budgets or create models that allocate compute adaptively based on token complexity.

Train multimodal models for long-chain reflective reasoning (iterative thinking, backtracking) using Adaptive Hybrid Policy Optimization. Trigger: improve VLM performance on tasks requiring iterative refinement and error correction.

Deep Research Agents (DRAs) generate citation-rich reports via multi-step search and synthesis, yet existing benchmarks mainly target text-only settings or short-form multimodal QA, missing end-to-end multimodal evidence use. We introduce MMDeepResearch-Bench (MMDR-Bench), a benchmark of 140 expert-crafted tasks across 21 domains, where each task provides an image-text bundle to evaluate multimodal understanding and citation-grounded report generation. Compared to prior setups, MMDR-Bench emp...

Evaluate whether generative models possess foundational reasoning capabilities. Develop five-ability framework (physical, logical, 3D spatial, 2D spatial, temporal reasoning) across abstract reasoning, embodied navigation, and physical commonsense benchmarks. Use structured rubric requiring simultaneous satisfaction of all sub-metrics.

Compresses MoE language models through shared basis factorization of expert weight matrices, achieving 24-30% parameter reduction with minimal accuracy loss.

Build GUI automation agents using self-evolving trajectory generation, trajectory-aware policy optimization, and integrated action semantics for cross-platform interaction.

Transform pre-trained vision-language models into powerful bidirectional multimodal embeddings through modality-aware continual pre-training and heterogeneous contrastive fine-tuning. 3B model matches 7B baselines.

Enable small vision-language models to reason over diverse data types by converting latent embeddings into visual representations, achieving specialized performance without domain-specific training.

Merges multiple fine-tuned models by operating in input-representation space rather than parameter space. Creates synthetic inputs whose gradients align with task vectors, bridging joint training and post-hoc merging for robust multi-task model combination.

Optimize data mixtures across diverse vision-language domains when applying RL with verifiable rewards to multimodal LLMs, balancing task-specific performance with generalization.

Train large models efficiently across heterogeneous hardware (GPUs, TPUs, Trainium) using strict encapsulation principles, achieving constant code complexity when adding features across hundreds of modules.

Prevents MoE router instability during RL training by recording and replaying inference-phase routing distributions back into training. Reduces training-inference routing divergence and KL divergence, enabling stable MoE RL scaling without sacrificing training speed.

Determine optimal MoE sparsity by separating memorization and reasoning trade-offs: active FLOPs improve reasoning while total parameters improve memorization, requiring joint optimization

Improve agent reasoning by designing thought structures that balance deep analysis, self-reflection, and exploratory thinking. Framework discovers that effective long-form reasoning exhibits molecular-like interaction patterns—specific bonds between reasoning components that enable fast entropy convergence. Method synthesizes improved reasoning trajectories using distribution-transfer, improving both model performance and RL training stability.

Enable multimodal LLMs to generate and reason with latent visual embeddings as intermediate thoughts: implement supervised fine-tuning to produce continuous visual representations, then optimize via VLPO to treat embeddings as learnable actions in reinforcement learning.

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

Generate videos at 29 FPS with interactive motion control through teacher-student distillation of motion-conditioned video models, using sliding-window causal attention and attention sinks to maintain constant latency for indefinite-length generation.

Improve factuality in large reasoning models by analyzing reasoning state transitions and reweighting preference optimization signals, suppressing defective reasoning segments while amplifying patterns that lead to factual outputs.

Improve agent search through meta-RL: generate multiple episodes sequentially, each building on prior attempts with explicit self-reflection. Use turn-level RLOO advantage estimation to provide dense credit without value models.

Prevent unrecoverable gradient explosions in LLM training by periodically restoring weight matrix stable rank through SVD-based matrix sign operations, eliminating sudden training failures without computational burden.

Apply μ-parametrization to Mixture-of-Experts architectures to enable reliable hyperparameter transfer across model sizes, eliminating costly retuning when scaling to trillion-parameter systems.

Enables LLM self-improvement without external verification through multi-agent co-evolution. Proposer generates questions, Solver attempts solutions, Judge evaluates both. All agents evolve together via RL, achieving 4.54% improvement on reasoning benchmarks without human supervision.

Design multi-agent systems with brain-inspired memory mechanisms that enable efficient information sharing and coordination. Implement hierarchical memory structures (working memory, episodic memory, semantic memory) similar to neuroscience models to improve multi-agent reasoning, planning, and task completion.

Build persistent, structured memory systems for LLM agents that remember user context across sessions, organize information semantically, and retrieve relevant knowledge automatically before responding. Achieves 35% accuracy gains over RAG baselines with 99.9% smaller storage overhead.

Train planner and worker agent roles within a single LLM via role-specific prompts and RL, avoiding multi-instance overhead while preserving specialization. Trigger: improve tool-use planning robustness to noisy outputs without deploying separate models.

Accelerate autoregressive image generation via multi-resolution drafting with spatially-informed verification. Local rejection and resampling enable efficient error correction focusing on spatial neighborhoods, achieving 1.7× speedup over baselines.

Enable balanced multi-task GRPO training via robustness-aware optimization and improvement-aware task reweighting, dynamically adjusting task weights based on both reward and loss trajectory improvement, achieving 6-28% worst-task improvements while maintaining competitive average accuracy.

Enable agents to communicate through shared latent thoughts rather than natural language, recovering both shared and private latent representations with theoretical guarantees for more efficient collaboration.

Train specialized agents in pipelines using dense per-action process rewards from AI coaching. Solves credit assignment in sequential workflows, enabling better generalization and faster convergence than outcome-only training.

Improve text-image alignment in diffusion transformers through Temperature-Adjusted Cross-modal Attention (TACA), addressing token imbalance and timestep-dependent weighting with parameter-efficient LoRA fine-tuning.

Generate unified embeddings for videos, images, and visual documents enabling semantic similarity, retrieval, and clustering across heterogeneous visual content types.

Enable native parallel token generation in language models by implementing adaptive task decomposition and merge strategies, achieving 2x speedup with 1.87% performance gains.

Improve LLM training efficiency by selectively applying Muon optimizer to Value-Output attention weights and FFN layers, which function as associative memories. Use when training data exhibits heavy-tailed distributions requiring robust rare-fact learning.

Enhanced language model pre-training methodology improving linguistic competence across languages, strengthening foundational capabilities for multilingual agent systems.

Teach multimodal LLMs to read and understand sheet music through a synthetic QA dataset with kern+ symbolic notation. Enables models to handle music sheet OCR, symbol recognition, and chord estimation at 8× better performance than GPT-4o.

Improves LLM reasoning quality at inference time by optimizing token logits using gradient descent, combining reward model signals with KL-regularization. Bridges parametric training-time and non-parametric test-time scaling through token-level optimization.

Apply training-free negative guidance in diffusion models by extrapolating in attention space with L1-based normalization, restoring suppression of unwanted attributes across architectures and modalities.

Enable LLMs to develop genuine parallel reasoning without external supervision through progressive self-distilled training. Transform models from sequential reasoning to native parallel cognition with 4.6× speedup—ideal when latency and reasoning quality both matter.

Convert auto-regressive language models to efficient diffusion-based generators through gradual block size increments. NBDiff-7B inherits long-context capabilities from AR predecessors while achieving state-of-the-art parallel generation—ideal when you need efficiency without sacrificing reasoning.

Efficient agentic reasoning via sparse MoE activating 50% parameters per token. Combines Mamba-Transformer hybrid with 6-of-128 expert routing, three-stage post-training (SFT, verifiable RL, RLHF), and Group Relative Length Control—achieving 3.3× inference throughput of competitors while maintaining 1M token context support and superior reasoning.

Build efficient open-source LLMs via hybrid Mamba-Transformer MoE architecture with LatentMoE expert design, multi-token prediction training, FP4 precision, and multi-environment RL post-training—achieving 3.3× higher throughput than equivalently-sized models while maintaining state-of-the-art reasoning, coding, and tool-use capabilities.

Train language models through sequential, domain-wise RL stages (RLHF → Instruction-Following → Math → Code → SWE) without catastrophic forgetting. Exploit policy-dependent training data distribution where previous behaviors persist when reward-relevant. 14B model surpasses DeepSeek-R1-0528 (671B) on LiveCodeBench.

Deploy multiple reasoning model sizes efficiently by embedding nested submodels within a single parent—use end-to-end trained routers to select submodels at inference, achieving 360× cost reduction vs training families separately.

Evolutionary architecture search discovering optimal depth-width ratios and operator combinations under deployment latency constraints using augmented scaling laws. Deploy when you need fastest inference per latency target with mixed attention mechanisms.

Create 7.5M long-form mathematical solution traces with multi-mode supervision (high/medium/low reasoning depths, with/without Python). Integrate 85K competition problems (AoPS) and 262K community questions (StackExchange). Implement sequential bucketing training achieving 2-3× speedup while maintaining accuracy.

Build hybrid Mamba-Transformer models combining efficient Mamba-2 layers with standard attention to achieve 6x higher inference throughput while maintaining reasoning accuracy on long-context tasks.

Train vision transformers through autoregressive next-embedding prediction without pixel reconstruction, tokenizers, or contrastive losses. Apply causal masking and stop-gradient on target embeddings. Achieve 83.8% (ViT-B) and 85.3% (ViT-L) ImageNet-1K accuracy with strong transfer to downstream tasks.