Claude Skills by ADu2021

Group Filtered Policy Optimization (GFPO) reduces inference-time computation by sampling larger groups during training and filtering responses based on length and token efficiency to teach models efficient reasoning.

Tokenize any region mask into two special tokens, enabling pixel-wise tasks like segmentation and region understanding in multimodal language models without architectural changes. Use when adapting vision-language models to perform pixel-level understanding and manipulation.

Decomposes math problems into sequential sub-problems, assesses difficulty, and allocates simple ones to fast System 1 reasoning while directing complex ones to deliberate System 2. Save 33-53% tokens while improving accuracy by up to 13.75 points on AIME.

Build scalable open-source agents that autonomously operate graphical user interfaces across Windows, macOS, Linux, Android, iOS, and web platforms using large-scale cross-platform training data and vision-language models for GUI understanding, element grounding, and task completion.

SCALER: Synthetic Scalable Adaptive Learning Environment for Reasoning. From arXiv:2601.04809

Train video game-playing foundation models discovering that increasing training data and network depth enables learning more causal policies. Release 8300+ hours of gameplay data and open-source models for real-time consumer GPU inference.

Scales reasoning depth through internal iteration rather than explicit generation. Ouro models perform repeated computation in latent space with entropy-regularized objectives enabling learned depth allocation. Smaller 1.4B model matches 12B standard models through improved knowledge manipulation.

Establish language-specific scaling laws for multilingual code pre-training across 7 programming languages (0.2B-14B models, 1T tokens). Show interpreted languages scale better than compiled, synergy gains depend on syntax similarity, and parallel pairing improves translation—enabling proportion-dependent token allocation outperforming uniform distribution.

Predict optimal data mixture proportions for multi-domain LLM training using scaling laws that require only 10-20 small experiments. Determine which domains should contribute how much data across model sizes (186M-7B), reducing computational waste in foundation model pretraining.

Accelerate diffusion LLM decoding by halting when confidence stabilizes using progress-dependent thresholds. SchED achieves 3.8-4.0× speedup while retaining 99.8-100% quality—training-free and model-agnostic for any dLLM.

Build community-driven evaluation platforms for scientific tasks using pairwise model comparisons and human voting. Assess foundation models on literature-grounded reasoning without automated metrics.

Implement techniques from Scientific Image Synthesis: Benchmarking, Methodologies, and Downstream Utility. While synthetic data has proven effective for improving scientific reasoning in the text domain, multimodal reasoning remains constrained by the difficulty of synthesizing scientifically rigorous images

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

Framework enabling LLMs to actively manage internal working memory and context through tools for fragmentation, summarization, and semantic search. Mitigates proactive interference and improves reasoning at extended context lengths.

Accelerate diffusion models through spectral-evolution-aware caching. Exploit insight that early timesteps establish low-frequency structure while later timesteps refine high-frequency details. Apply FFT-based frequency filtering to feature cache decisions: preserve content-relevant frequency components while suppressing noise. Plug-and-play training-free enhancement achieving 1.5–2.5× speedup across FLUX, HunyuanVideo, Wan2.1 models.

Vision-based computer use agent that self-improves through experiential learning, curriculum generation, and reward-based RL on diverse software.

Evaluate search-augmented language models on fact-seeking questions with conflicting or unhelpful search results, revealing critical reasoning gaps in frontier models and testing robustness to noisy information.

Decompose search-integrated reasoning into actor (generates trajectories) and refiner (identifies and corrects flawed steps). Refiner performs surgical cut-and-regenerate correction preserving valid prefixes. Trained end-to-end with GRPO; adds minimal overhead (2%) to baseline.

Unify LLM reasoning with embedding generation by training models to output embeddings as direct outputs of step-by-step reasoning. Trigger: improve performance on complex retrieval tasks requiring deep reasoning.

Train multi-turn AI agents with convergence guarantees using sequential backward-induction updates, eliminating the need for separate critic networks while maintaining theoretical optimality. Use for long-horizon agentic reasoning where monotonic improvement and global optimality are required.

Fast parallel code generation using discrete-state diffusion models with curriculum learning and trajectory optimization, achieving 2,146 tokens/second inference speed.

Enable LLM agents to incrementally construct formal proofs through strategic tool orchestration rather than one-shot generation. Combines natural language reasoning, sketch translation, and formal verification in an agentic workflow trained via PPO and Rubric RL, achieving competitive performance on undergraduate and graduate proof problems.

Framework combining formal verification feedback with long chain-of-thought reasoning to enable deep and broad mathematical reasoning for automated theorem proving. Achieves 78.1% on formalized IMO problems through lemma-based refinement and test-time inference strategies.

Achieves single-step video restoration at 1080p resolution with 4x speedup over multi-step diffusion approaches via adversarial training, adaptive window attention, and feature matching loss.

Segment-level credit assignment for RL in LLMs using Monte Carlo advantage estimation, enabling precise reward attribution without critic models for improved reasoning task performance.

Train text-to-image models from scratch for any inference step count via self-evaluation mechanism. Model evaluates its own generated samples using current score estimates as dynamic self-teacher, enabling global distribution matching without external teachers—achieving few-step quality equivalent to many-step models at all budgets.

Enable agents to learn continuously from execution experience through hierarchical memory and autonomous reflection. Trigger: improve agent performance on long-horizon tasks by accumulating and applying experience.

Address exposure bias in video generation by training on self-generated sequences, achieving 17 FPS real-time performance while matching quality of slower baselines.

Enable VLMs to self-assess visual perception accuracy through decomposed two-stage reasoning: perception generation then validation, eliminating external supervision dependency

Add explicit visual supervision to VLMs by training models to autoregressively reconstruct semantic image tokens, achieving 2-3% average gains and 10-point improvements on hallucination robustness.

Accelerate video generation and enable long-video synthesis by decomposing into two diffusion stages: first generate compact semantic features for global planning, then generate VAE latents conditioned on semantics. Includes learnable semantic compression to improve training convergence—enabling minute-long videos with faster convergence than direct VAE modeling.

Encode reasoning steps as hidden embeddings instead of explicit text using contrastively-trained sentence transformers and lightweight distilled models, reducing token generation cost while preserving semantic alignment with ground-truth reasoning.

Build vision-language agents that seamlessly integrate visual reasoning with dynamic tool manipulation (search, cropping) through reinforcement learning, achieving state-of-the-art performance on fine-grained visual understanding tasks—surpassing proprietary models like Gemini and GPT.

Deploy a state-of-the-art binary classifier using ModernBERT to detect prompt injection attacks and protect LLMs from adversarial input manipulation.

Implement adaptive parallel decoding for language models using diffusion-based next-sequence prediction. Enable dynamic block-based token generation with confidence thresholds to achieve 2x+ speedups while maintaining competitive performance. Retrofit existing autoregressive models with minimal additional training data.

Replace parallel self-consistency with sequential reasoning where chains iteratively build on previous attempts, weighted by inverse entropy to prioritize confident solutions, achieving 46.7 pp accuracy gains over parallel approaches.

Accelerate language model generation 3-5x by combining autoregressive and masked token prediction. Works via fine-tuning—no architectural changes needed. Parallel decode non-consecutive tokens with entropy-bounded sampling.

Demonstrate that synthetic CoT traces with incorrect final answers outperform human-written correct solutions for supervised fine-tuning. Distribution proximity between training data and student model's natural output matters more than correctness—validating human traces with model-like distributions improves performance, providing practical guidance for dataset curation.

DeltaLoss sensitivity metric combining gradient and quantization-induced parameter deviation for adaptive bit-width allocation, with lightweight pre-tuning search for scale initialization, enabling competitive accuracy at 4-5 bits in 2.5-6 hours.

Convert videos to language-based representations and leverage LLM reasoning without video-specific training.

Train LLMs to reason implicitly with step-level supervision, stabilizing latent representations while preserving 2.3× inference speedup over explicit chain-of-thought. Addresses training collapse in implicit reasoning by aligning intermediate latent states with explicit reasoning steps through an auxiliary decoder during training, then removing the decoder for efficient inference.

Improve pass@K by using asymmetric probability boosting: increase probabilities of top-K correct solutions while penalizing top-1 incorrect predictions. Focus boosting on high-entropy tokens where exploration helps most.

Place RMSNorm immediately after every linear layer to stabilize activation scales at O(sqrt(d)) and reduce Hessian spectral norm. Enables 3-10x larger learning rates and faster convergence without architectural changes; improves loss by 0.08 on 7B models.

Implement efficient memory systems for long-term LLM agent interactions using semantic compression, achieving 30-fold inference token reduction while improving F1 scores by 26.4%—enabling agents to learn from extended interaction histories without prohibitive context costs.

Train LLMs for multi-turn tool-integrated reasoning end-to-end using RL without supervised pretraining. SimpleTIR stabilizes training by filtering void turns (responses lacking code blocks or final answers) to prevent gradient explosion from distributional drift, enabling discovery of emergent reasoning patterns like self-correction and cross-validation on mathematical benchmarks.

Apply reinforcement learning to Vision-Language-Action models for robotic control, achieving 99% LIBERO task success and discovering novel manipulation strategies (pushcut) without task-specific reward engineering. Scales efficiently via parallelized trajectory sampling and outcome-based rewards.

Evaluating whether multimodal large language models truly understand long-form scientific papers remains challenging: answer-only metrics and synthetic 'Needle-In-A-Haystack' tests often reward answer matching without requiring a causal, evidence-linked reasoning trace in the document. We propose the 'Fish-in-the-Ocean' (FITO) paradigm, which requires models to construct explicit cross-modal evidence chains within native scientific documents. To operationalize FITO, we build SIN-Data, a scien...

Personalize image generation and editing from a single reference image through inference-time LoRA optimization. Iteratively update model parameters based on visual similarity scores without training encoders or fine-tuning on multiple images.

Replace LoRA's two-matrix decomposition with a single learnable matrix (AA⊤) to eliminate scale imbalances and improve training stability. Reduces parameters by ~50% while maintaining or exceeding LoRA performance.

Improve agent performance by autonomously distilling behavioral patterns from trajectories into reusable skills, then using these skills to guide future decisions. Achieves 89.9% success on ALFWorld through differential processing of success vs failure episodes and dynamic skill library evolution.