Claude Skills by ADu2021

State-of-the-art multimodal model advancing vision-language understanding and generation capabilities through improved visual encoders, dense token representations, and unified reasoning over images and text.

Systematically post-train models for long-context reasoning through multi-hop data synthesis, stabilized RL with adaptive entropy control, and memory-augmented architecture supporting 4M+ token sequences. Achieves performance comparable to GPT-5 and Gemini-2.5-Pro on long-context benchmarks.

Construct multi-step reasoning benchmarks with interdependent problems to evaluate and improve long-horizon reasoning in large reasoning models. Enables evaluation of reasoning depth and breadth beyond single-step tasks.

Ground LLM world models with retrieved current knowledge from tutorials and documentation. Reduce hallucination in environment prediction and improve long-horizon planning by 16-23% on web agent benchmarks.

Co-evolutionary framework where Challenger generates tasks and Solver solves them. Models evolve autonomously from scratch without human annotations. Improves math reasoning +6.49pts and general reasoning +7.54pts.

Accelerate video diffusion models using sparse radial attention that exploits energy decay patterns. Achieves 3.7× speedup on long videos while maintaining quality through O(n log n) complexity instead of O(n²).

Optimize where multimodal models attend by treating attention weights as a learnable policy, using policy gradients with advantage weighting to improve visual grounding and perception without changing model architecture.

Learn optimal per-layer bit-width assignments for LLM quantization via RL, generalizing across models without retraining. Achieves superior compression under fixed bit budgets.

Apply rank-one weight modifications to amplify model safety via residual stream steering, requiring no fine-tuning and preserving utility on standard benchmarks

Bridge image understanding-generation gap via In-Context Chain-of-Thought reasoning and RL training with surrogate rewards. Improve faithful execution of mixed image-text prompts in generation and editing tasks.

Restructure latent representations in pretrained audio autoencoders without full retraining. Apply three variants—ordered, semantic, and equivariant—to enforce structure like channel ordering, semantic alignment, or filter correspondence. Achieves 20-60% semantic gains in under 48 GPU hours versus 14.5K hours for full retraining.

Compress search trajectories into structured states capturing partial answers, evidence, and uncertainties. Recursive execution leverages compressed states to avoid redundant exploration, improving resource efficiency by 50%.

Implements RealMem from arXiv:2601.06966

Enhances passage ranking through reasoning capabilities via synthesized training data and multi-stage training combining supervised fine-tuning with reinforcement learning for improved ranking accuracy.

Enable image generators to reason explicitly through text before creating images using supervised fine-tuning and reinforcement learning optimization.

Optimize chain-of-thought reasoning under computational budgets using information-theoretic compression principles, improving reasoning efficiency without accuracy loss.

Reasoning Core procedurally generates verifiable symbolic reasoning datasets across formal domains (planning, logic, parsing), with external solvers and curriculum control.

Create infinite training environments for reasoning with automatic verification using procedural generation and domain-specific evaluators.

Large language models have achieved remarkable capabilities across domains, yet mechanisms underlying sophisticated reasoning remain elusive. Recent reasoning models outperform comparable instruction-tuned models on complex cognitive tasks, attributed to extended computation through longer chains of thought. Here we show that enhanced reasoning emerges not from extended computation alone, but from simulating multi-agent-like interactions -- a society of thought -- which enables diversificatio...

Overcome token-level randomness limitations in RL by shifting exploration to latent reasoning strategies. Train a VAE encoding diverse reasoning patterns, sample latents during RL, decode to prefix embeddings steering internal reasoning—enabling structured exploration across math, coding, and QA with interpretable, controllable behavior.

Reduce LLM sampling costs by 50% while maintaining reasoning performance through Reasoning Path Confidence (RPC), which combines perplexity-guided pruning with self-consistency sampling.

Extract and transfer reasoning capabilities between language models using task vectors derived from supervised fine-tuning and reinforcement learning weight differences. Apply reasoning vectors via simple arithmetic to enhance any compatible instruction-tuned model without retraining.

Improves base model reasoning through iterative sampling without training or fine-tuning. Uses MCMC-inspired sampling to extract latent reasoning from pretrained models, achieving RL-comparable gains on math, coding, and QA tasks while preserving diversity.

Diagnose and correct reasoning inefficiencies (overthinking and underthinking) in large reasoning models using confidence-based steering vectors, without retraining. Enables optimal reasoning budgets across model scales.

Implement techniques from RebuttalAgent: Strategic Persuasion in Academic Rebuttal via Theory of Mind. Although artificial intelligence (AI) has become deeply integrated into various stages of the research workflow and achieved remarkable advancements, academic rebuttal remains a significant and underexplored challenge

Prevents catastrophic forgetting in continual learning by merging models using layer-wise hidden representations as similarity proxies. Shallow layers preserve domain-general features while deep layers enable task-specific adaptation, enabling seamless multi-domain integration without task labels or historical data.

Unifies planning and action by treating plans as abstract placeholder functions recursively decomposed to primitive actions. Enables agents to dynamically adjust abstraction levels per task without rigid hierarchies. Improves inference performance and training efficiency through automatic multi-level data generation.

Build self-evolving multi-agent browser systems that combine web reconnaissance with dynamic tool generation and execution. Enables autonomous agents to analyze failed trajectories, generate specialized tools on-the-fly, and adapt to novel web environments without pre-defined action sets, achieving 36.48% success on VisualWebArena.

Enable efficient long-sequence generation by combining block-sparse attention with periodic dense rectification to bound error accumulation.

Extend neural network reasoning capabilities through recurrence (repeated computation cycles), external memory (intermediate state storage), and test-time compute scaling for multi-step reasoning.

Enable test-time compute scaling in vision-language-action models via weight-tied recurrent inference within latent space, with adaptive stopping based on action divergence.

Process prompts exceeding model context windows by recursively decomposing long inputs into manageable chunks and calling the model recursively on snippets—enabling inference on contexts 100x longer than native window while maintaining quality and improving over vanilla baseline approaches.

Recursive Think-Answer Process enables models to iteratively refine reasoning and answers during inference, reducing self-correction errors and improving accuracy without retraining.

Comprehensive evaluation dataset for systematic vulnerability testing of language models, enabling identification and mitigation of failure modes before agent deployment.

Train language models for multi-step information-seeking using dual-constrained task synthesis and cost-efficient staged learning. Generate complex queries by controlling topological complexity and information dispersion, then train atomic reasoning skills before introducing environment interaction. Achieve state-of-the-art on long-horizon search benchmarks with 30B parameter models.

Trains LLMs to autonomously debug and improve code through structured RL-optimized reflection cycles. Internalizes debugging process into model weights rather than relying on external oracles or expensive iterative prompting.

Identify and patch critical safety vulnerabilities in large reasoning models. Via linear probing and causal intervention, locate specific attention heads responsible for alignment degradation at final tokens. Recover safety via 'Cliff-as-a-Judge' data curation targeting examples exhibiting largest refusal decline, achieving comparable improvements using only 1.7% of vanilla safety training data.

ReGFT pre-trains models on hybrid reference-augmented trajectories before RL, enabling them to solve harder problems and accelerate convergence.

Merge softmax and linear attention for video diffusion models using chunk-wise recurrent reformulation with constant memory usage. Enable efficient distillation from existing softmax models, reducing training cost two orders of magnitude to ~160 GPU hours.

Recover learning signals in RL for LLM reasoning by dynamically allocating sampling budget based on prompt difficulty. Use log-objective weighting (1/p for pass rate p) to prioritize challenging examples, achieving 2x convergence speedup versus uniform sampling while maintaining identical compute budgets across math, coding, and general benchmarks.

Extract maximum value from limited reasoning traces by leveraging both successful and failed examples through REINFORCE-style distillation.

Train multimodal LLMs with RL (PIVOT) instead of SFT to produce stronger, precisely-localized visual representations in vision encoders using <1% of standard pretraining cost.

Add an intermediate RL stage between pretraining and post-training using dynamic token budgeting, curriculum sampling, and dual training. Trigger: reduce reasoning steps while maintaining or improving performance in post-training.

Apply reinforcement learning to pre-training by framing next-token prediction as a reasoning task with verifiable rewards, achieving superior scaling compared to standard language modeling.

Reduce inference cost by dynamically switching from large to small LLMs during reasoning generation. Large model handles demanding reasoning phases; small model completes consolidation and answer stages triggered by discourse cues. Achieves 2.2× speedup with minimal accuracy loss.

Enable small language models to dynamically invoke larger models at critical reasoning tokens rather than offloading entire queries. RelayLLM achieves 49.52% accuracy across benchmarks while invoking the large model for only 1.07% of tokens—98.2% cost reduction compared to non-collaborative approaches.

Detect and eliminate data contamination that invalidates RL benchmarks by measuring benchmark reconstruction ability. Implement clean evaluation datasets to distinguish genuine reasoning improvements from memorization. Use when validating RL training results or ensuring benchmark integrity for mathematical reasoning.

Compressed historical latents with camera poses in KV cache (4× compression), extended teacher training (20-second sequences), and replayed back-propagation (block-wise differentiation) enabling real-time interactive video generation with long-range spatial consistency.

Improve LLM reasoning by reweighting pre-training data during mid-training based on discrepancies between RL-tuned and base models, boosting reasoning performance without external teachers or extra data.

Learn to route requests across multiple LoRA adapters using RL-based router training with constant routing weights. Prevents weight collapse and ensures balanced contribution from all selected adapters during inference.