Claude Skills by ADu2021

Optimize test-time computation through graph-based collaborative architecture where nodes represent models, edges represent information flow, and topology itself is optimizable via reinforcement learning to discover ideal model assignments and configurations.

Train robot controllers via five-stage curriculum progressing from base vision-language models to embodiment-specific RL-refined policies. Unified action space enables cross-embodiment transfer with minimal performance loss.

Improve ranking quality via groupwise reranking with RL—process document groups jointly for within-group comparisons using GRPO with composite rewards (recall, ranking metrics, distribution alignment).

Enables efficient MoE architectures through heterogeneous expert sizing and dynamic activation mechanisms that adjust parameter count based on input complexity.

Build language models layer-by-layer on frozen embeddings, training new Transformer blocks sequentially while keeping lower layers frozen. Achieves 5% improvement over monolithic baselines on MMLU while fitting 740M trainable parameters per stage on single GPUs, enabling resource-efficient incremental scaling to 2.3B parameters.

Stabilize and accelerate chain-of-thought RL training by sampling multiple answers per generated thought. GRPO-MA reduces gradient noise and improves convergence across math, code, vision, and manipulation tasks while cutting computational cost versus naive thought scaling—critical for training reasoning models without explicit value networks.

Eliminate expensive external teacher dependencies in VLM RL training via merged-checkpoint teachers. Uses TIES merging of historical RL checkpoints to create free, stable teacher models for step-level guidance—matching external teacher performance while reducing training time 50% and computational costs 60%.

Recover fine details at soft boundaries (hair, fur) through depth refinement networks and view synthesis. Integrate plug-and-play with existing depth models via adaptive combination across monocular, stereo, and novel view tasks.

Enable training and evaluation of desktop computer-using agents through 1.2M action steps across diverse Windows applications, covering GUI grounding, screen parsing, and action prediction with hybrid GUI+API action space reflecting modern agent designs.

Enable GUI agents to ground actions without generating pixel coordinates by using attention-based patch-level alignment and a verifier for selecting optimal action regions from candidates.

Improve GUI agent planning and action grounding through test-time scaling and reinforcement learning. Sample and evaluate multiple action candidates, then use RL to precisely target visual interface elements.

Implement techniques from Guidelines to Prompt Large Language Models for Code Generation: An Empirical Characterization. Large Language Models (LLMs) are nowadays extensively used for various types of software engineering tasks, primarily code generation

Eliminate fixed tokenization by learning data-dependent segmentation jointly with the model through dynamic chunking, matching BPE-tokenized Transformers at equivalent compute while showing improved robustness and better downstream task performance without vocabulary constraints.

Generate comprehensive test cases for code problems that reliably detect wrong solutions through LLM-based edge case synthesis and test quality ranking.

Evaluate LLM-based document rerankers using hardware-agnostic FLOPs metrics instead of latency, enabling fair comparison of ranking quality per unit of computation across different models and deployment scenarios.

Multi-agent framework that balances safety compliance with task completion through adaptive policy extraction and dual-objective optimization. Achieves 38% improvement in policy compliance while maintaining 20% higher task completion.

Compute step-level credit assignments via hindsight generative verification: condition the LLM on successful outcomes to compute importance ratios that filter credit by causal relevance. Synergizes macro stability with micro precision.

Use KV cache as hierarchical memory for real-time video stream understanding with minimal GPU overhead, achieving 10x faster response times compared to standard methods. Use when processing continuous video streams where latency and memory efficiency are critical.

Scale disaggregated LLM inference (prefill-decode) via topology-aware scheduling and metric-driven policies, achieving 26.6% GPU utilization improvement and conserving hundreds of thousands GPU-hours daily

Reason about long-horizon dynamics by disentangling structure and motion in video VAE latents. Learn continuous latent motion chains that preserve temporal coherence while predicting terminal keyframes, enabling efficient reasoning about multi-step scenarios.

Optimize only high-entropy tokens during RL training to achieve better reasoning performance with 80% fewer gradient updates.

Enable data-dependent higher-order interactions in attention using prefix-sufficient statistics that maintain linear time and constant state, replacing quadratic dot-product attention while preserving expressivity through compact matrix operations.

Solves long-horizon planning problems with global constraints by decoupling planning into strategic (resource allocation) and tactical (execution) levels. Prevents constraint drift through synchronized state tracking and cooperative bargaining.

Novel error categorization framework examining failures across hops (diversity), coverage, and overthinking. Combines human annotation with automated metrics to diagnose why reasoning models hallucinate on multi-step tasks.

A VLM-based preference scoring system trained on 1.17M annotated comparisons to evaluate text-to-image generation quality at scale. Uses uncertainty-aware ranking loss for fine-grained assessment across diverse images and supports iterative quality improvement through chain-of-human-preference sampling.

Improve multimodal reasoning by requiring explicit context understanding before reasoning. Use specialized reward mechanisms and context-aware training to prevent information-skipping shortcuts.

Design hybrid architectures mixing linear and full attention at optimal ratios. Superior standalone linear models don't necessarily excel in hybrids; recall improves significantly with more full attention layers at ratios below 3:1, enabling efficient long-sequence processing.

Combine sparse verifier rewards with dense reward model scores using stratified normalization to overcome limitations of either approach alone.

Enable AI systems to recursively improve themselves by making the meta-level modification procedure itself editable, achieving open-ended capability growth.

Build hypergraph-structured memory systems for multi-step RAG that capture high-order relationships between facts, enabling stronger reasoning across long contexts. Use when combining multiple retrieved documents in complex reasoning chains that require understanding connections between pieces of information.

Improve LLM reasoning through iterative refinement where the model refines its best previous attempts. Two-stage training: exploratory draft generation, then conditioned refinement using GRPO. Dynamic conditioning signals evolve with policy, enabling state-of-the-art math reasoning on AIME (85.62%).

Enable robust image editing by decomposing free-form instructions into sequential atomic operations executed through a neural program interpreter.

Enhances RLVR through cooperative-adversarial flywheel, intent verification, and trap instructions. Detects reward hacking and improves training efficiency, achieving 87.43% on IFEval.

Evaluate reasoning model capabilities by analyzing three complexity-dependent behavioral regimes and identifying fundamental limitations in symbolic manipulation rather than computational budgets.

Upscale any degraded image to 4K using an agentic framework that analyzes image quality, selects appropriate restoration tools, and iteratively improves results through reasoning and reflection.

Imagine-then-Plan: Agent Learning from Adaptive Lookahead with World Models. From arXiv:2601.08955

Accelerate sparse attention by reusing token selection indices across layers. Partition layers into full indexer (F) and shared (S) types using greedy search or multi-layer distillation to eliminate 75% of indexer computation.

Introduce IMDD-1M dataset with 1M aligned image-text pairs spanning 60+ material categories and 400+ defect types. Train diffusion-based vision-language models requiring 5% task-specific data vs. dedicated expert models for manufacturing quality control.

Implement techniques from Inference-Time Scaling of Verification: Self-Evolving Deep Research Agents via Test-Time Rubric-Guided Verification. While the majority of existing efforts focus on enhancing policy capabilities via post-training, we propose an alternative paradigm: self-evolving the agent's ability by iteratively verifying the policy model's outputs, guided by meticulously crafted rubrics

Combines SFT and DPO with robust data selection pipeline using multidimensional quality metrics. Achieves DeepSeek-R1 performance with 12% training data, enabling efficient reasoning model alignment.

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

Merge sliding window and linear attention (Gated DeltaNet) for unlimited VLM inputs with 3.6× speedup. InfiniteVL handles video understanding at 24 FPS with constant memory—ideal when context length must scale without quadratic overhead.

Implement dense-sparse switchable attention enabling LLMs to scale from short to long sequences with 4× speedup and 98-99.7% performance retention, requiring no extra parameters by reusing pretrained attention weights through trainable sparse pattern selection.

Optimize multi-turn agent policies by measuring turn-level information gain via counterfactual reasoning. Provide dense reward signals identifying which clarifying questions and observations improve the agent's decision distribution, then adaptively blend information rewards with outcome rewards.

Enable more precise LLM error correction by having users directly edit the model's previous response, conditioning the next response on this corrected version. This approach reduces token overhead by 79% compared to traditional separate-feedback methods while fixing more errors in complex reasoning tasks.

Enable VLMs to perform generalized visual search—locating relational, fuzzy, and conceptual regions from free-form language descriptions. Introduces O3-Bench benchmark with high-density composite charts/maps, uses RL-trained vSearcher for spatial localization, improving frontier models (GPT-5-mini 39%→61.5%) without architecture changes.

Convert pre-trained latent video diffusion into real-time autoregressive generators using adversarial post-training, achieving 24fps streaming on single H100.

Enhance multimodal models through cascade RL for reasoning improvement and visual resolution routing for inference efficiency, achieving 16% reasoning gains and 4.05x speedup.

Map text embeddings into visual representation space for multimodal fusion, eliminating expensive image-text alignment pre-training while improving reasoning-heavy tasks by up to 27.2%.

Connect LLM agents to IoT sensors and microcontrollers through MCP standardization. Use to build monitoring systems and smart home automation where LLMs reason over real-world sensor data.