Claude Skills by ADu2021

Evaluate and optimize tool-integrated reasoning in LLMs through empirical benchmarking, performance-cost metrics (PAC, AUC-PCC), and measurement frameworks for diverse reasoning tasks.

Convert inference-time feedback into persistent, retrievable guidelines stored as agent memory. Framework enables LLMs to improve performance over time by systematically accumulating and applying learned critiques. Augmented models rapidly match test-time refinement performance while drastically reducing inference cost. Memory-as-tool pattern enables agents to learn from feedback without expensive retraining.

Align latent distributions with arbitrary reference distributions via explicit matching constraints rather than fixed priors. DMVAE achieves gFID 3.2 on ImageNet with 64 epochs—when you need flexibility in latent representation design for image generation.

Synthesize diverse, verifiable training tasks by executing real tools first, then reverse-deriving tasks from execution traces. Ensure diversity across tools and reasoning patterns while maintaining grounding by construction.

Train models to decompose complex problems into subproblems via divide-and-conquer reasoning. Structured approach enables systematic solution assembly and improved long-horizon reasoning compared to end-to-end generation.

Improve multimodal reasoning at test-time through confidence-guided latent optimization without retraining. Iteratively refine learnable latent think tokens via policy gradient using confidence reward. Dynamically select and update relevant image patches based on internal confidence levels. Maintain high efficiency with all optimization in latent space.

Generate multiple tokens simultaneously by having late transformer layers directly predict multiple outputs after early layer processing. No auxiliary parameters or verification needed. Achieve up to 2x generation speedup.

Systematic evaluation toolkit for assessing large language models across multiple dimensions, enabling comprehensive benchmarking of agent capabilities and comparative analysis of model performance.

Build open-source agents for document question-answering by modeling DocQA as information-seeking with explicit tool utilization. DocDancer uses an exploration-then-synthesis pipeline to generate high-quality training data, addressing the scarcity that limits agent-based document understanding systems.

Improve long-context length extrapolation by denoising instabilities in Rotary Position Embeddings (RoPE) through spectral analysis and selective head rewriting—training-free post-hoc intervention for longer context windows.

Compress LLMs by 20-30% in width while preserving functionality through optimal transport-based neuron merging. Instead of discarding neurons, redistribute their signal to retained neurons via learned transport maps. Use when you need to reduce model size with minimal accuracy loss and measurable computational speedup.

Diagnose and fix multi-agent system failures through targeted interventions (message edits, plan changes) rather than static log analysis. DoVer recovers 18-28% of failed trials with 30-60% hypothesis validation—essential for autonomous multi-agent reliability.

Enable Diffusion Language Models to achieve 8.5x inference speedup (24-30 steps vs. 256) through certainty-forcing distillation that trains models to achieve simultaneous high confidence across multiple tokens. Use when optimizing inference latency for dLLM deployments.

Replace PPO's heuristic ratio-based clipping with Divergence Proximal Policy Optimization (DPPO) that directly constrains policy divergence using either Total Variation or KL, enabling lightweight approximations (Binary, Top-K) for vocabulary-scale computations while improving stability and efficiency.

Implements DPWriter from arXiv:2601.09609

Use per-layer routers trained with Monte Carlo Tree Search to dynamically skip, execute, or repeat transformer layers for each token. Saves ~5 layers per example while maintaining accuracy on diverse benchmarks.

Enable stable multi-agent reinforcement learning by normalizing advantages per-agent rather than globally, preventing gradient-norm inflation in heterogeneous multi-agent systems.

Dr. Zero: Self-Evolving Search Agents without Training Data. From arXiv:2601.07055

Convert video generation model outputs into executable robotic manipulation by extracting 3D object flow trajectories as an intermediate representation. Enables zero-shot manipulation of diverse object types (rigid, articulated, deformable, granular) without task-specific training. Use when pre-trained video models capture plausible manipulation patterns but need grounding in low-level robot control.

Scale agent learning by synthesizing diverse experiences using reasoning-based models instead of costly real-world rollouts, maintaining replay buffers with both real and synthetic interactions while using adaptive curriculum to focus on challenging tasks.

Unified system for multi-task video generation combining audio and visual synthesis, demonstrating scalable approaches for content generation that can enhance agent communication capabilities.

DrivingGen: A Comprehensive Benchmark for Generative Video World Models in Autonomous Driving. From arXiv:2601.01528

Implement techniques from DRPG (Decompose, Retrieve, Plan, Generate): An Agentic Framework for Academic Rebuttal. Despite the growing adoption of large language models (LLMs) in scientific research workflows, automated support for academic rebuttal, a crucial step in academic communication and peer review, remains largely underexplored

Improve LLM reasoning by promoting diversity at both trajectory and token levels simultaneously. Global (trajectory) scale rewards distinct correct solutions; local (token) scale applies entropy regularization per decision point. Dual-scale approach couples these via diversity-weighted allocation: solutions that are globally more distinctive receive stronger local regularization, focusing exploration where it matters most among underexplored correct modes.

Implement techniques from DSGym: A Holistic Framework for Evaluating and Training Data Science Agents. Data science agents promise to accelerate discovery and insight-generation by turning data into executable analyses and findings

Optimize disaggregated prefill-decoding LLM serving for multi-turn (agentic) workloads by introducing dual-path KV-cache loading. Traditional approach loads all KV-cache to prefill engines, saturating their storage network. DualPath loads to decoding engines first, then transfers via compute network (lower contention). Adaptive routing selects path based on real-time queue depths. Achieves 1.87× offline throughput and 1.96× online serving improvement.

Generate SVGs through simultaneous image and SVG token generation with internal visual guidance. DuetSVG overcomes text-only limitations by leveraging visual predictions to enhance SVG coherence—ideal when visual quality and geometric correctness matter.

Implement dual preference optimization to generate self-supervised feedback without manual annotation by decomposing tasks into known/unknown components and reconstructing hidden information from model outputs.

Accelerate test-time scaling for diffusion language models by identifying inconsistent tokens, selectively remask and regenerate only uncertain tokens, and aggregate across samples via voting. Achieve 5.5-22× speedup over standard iterative sampling with 6-8% accuracy gains on reasoning tasks.

Automatically construct compact, diverse action spaces for LLM reasoning through corpus-based estimation and submodular optimization—enabling efficient decision-making without manual specification or expensive exhaustive search.

Implement hierarchical language modeling that compresses variable-length token sequences into high-capacity semantic concepts, achieving +2.69% benchmark improvements while reducing inference FLOPs by reallocating compute to concept-level reasoning. Use for efficiency-critical deployments where reasoning quality can be improved while maintaining computational budget.

Minimal modification to SFT that dynamically rescales objectives by token probability. Rectifies implicit reward structure to improve generalization comparable to RL while maintaining SFT simplicity.

Sparse attention mechanism combining content-aware and position-aware sparse patterns through dynamic masking. Achieves 10x speedup while maintaining model quality on long-context benchmarks through hardware-friendly implementation.

Dynamic token pruning framework for VLMs that adapts compression to scene complexity through single-pass selection. Removes 92.6% of visual tokens while maintaining performance and enabling superior fine-tuning.

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

Extends text embedding models to perform listwise reranking through continued training on ranking objectives. Constructs listwise prompts from queries and top-K candidates, leveraging pseudo-relevance feedback while maintaining embedding model efficiency. Unifies retrieval and reranking in single model.

Monitor token-wise entropy to adaptively allocate compute during inference. Branch into multiple paths at high-entropy tokens, reducing token generation by up to 65% while improving accuracy by up to 37% on reasoning tasks.

Train efficient planners for long-horizon agent tasks using homologous consensus filtering to generate synthetic plans from strong LLMs and rule-based RL with executor capability rewards. Reduces training cost by 8x while maintaining state-of-the-art performance.

Enable image editors to handle complex instructions through iterative critique and refinement cycles. A multimodal LLM critiques editing results, reasons about improvements, and refines instructions until satisfactory output—ideal for instruction-following challenges in visual editing.

Systematically optimize agent system costs via empirical analysis of LLM, planning, memory, and search components achieving 28.4% cost reduction.

Systematically convert pretrained autoregressive models into efficient diffusion language models via block-wise attention and position-dependent masking. Efficient-DLM family (1.5B/4B/8B) maintains comparable accuracy to standard AR models while delivering 4.5× higher throughput.

Achieve 10× higher decoding throughput on long prompts by replacing 50% of cross-attention layers with gated memory units (GMUs) combining SSMs and attention. Maintains reasoning capability while reducing memory I/O bottleneck from O(d_kv·N) to O(d_h).

Framework for efficient machine unlearning that reformulates forgetting as inverse learning. Achieves significant computational speedup by replacing expensive Hessian operations with gradient-based optimization, enabling privacy-preserving model updates.

Comprehensive survey of techniques for optimizing large reasoning models. Covers single-model optimization and multi-model collaboration approaches to reduce reasoning path length without sacrificing capability.

Enable autonomous embodied agents to function in 3D communities with structured memory systems. Combines semantic memory (scene graphs, knowledge graphs) and episodic memory (spatiotemporal experiences) for social intelligence and multi-agent coordination.

Improve multimodal embeddings through RL-optimized reasoning that grounds evidence in retrievable visual cues. Frozen embedder provides stable rewards while reasoner generates evidential traceability CoT with text keywords, bounding boxes, and key frames.

Accelerate LLM decoding by predicting multiple future tokens simultaneously using mask-token probing in embedding space, without retraining or auxiliary models.

Bridge vision-to-action gap using pointing as unified intermediate representation, enabling 56.2% success on manipulation tasks without task-specific fine-tuning.

Discover hierarchical temporal abstractions within autoregressive models via internal RL, enabling efficient exploration of sparse-reward tasks. Metacontroller learns abstract action sequences modifying residual streams, switching gates enable quasi-binary patterns, and abstract-space RL achieves many orders-of-magnitude speedup over token-level learning.

Build a single model handling multimodal understanding, generation, and editing tasks efficiently through token compression and intelligent component sharing. EMMA-4B surpasses larger models while reducing computational burden—ideal when you need unified performance across vision and language tasks.