Claude Skills by ADu2021

Scale GUI automation by organizing trajectories into tree structures for reuse and branching exploration, reducing data cost while maximizing step-level diversity through adaptive topology.

Achieve 2.4× faster RL training for diffusion models by restructuring denoising as tree search with shared computation. TreeGRPO computes fine-grained step advantages instead of trajectory-level rewards—crucial for efficient diffusion model optimization.

Train vision-language models to produce visually grounded reasoning by enforcing traceable evidence via bounding box localization, using a novel benchmark (TreeBench) and RL-based training with dual IoU rewards for both recall and precision.

T-PPO improves training efficiency via truncated rollouts and extended GAE, enabling batch continuity without waiting for full sequence completion.

Improve credit assignment in retrieval-augmented reasoning by truncating trajectories at single decision points. Generate k samples sharing a common prefix, differing only at the next step to isolate variation and reduce gradient variance by T-fold on T-step tasks.

Convert natural language to SQL for unknown database schemas by formulating the task as a partially observable MDP. Use dual-track GRPO (token-level masked advantages) to learn schema discovery and query generation jointly.

Train LLMs to reduce hallucinations by 28.9% using a ternary reward scheme that explicitly incentivizes abstention (+0) over false claims (-1) while rewarding correct answers (+1). Apply when improving factual reliability is critical and verification signals are available.

Reduce LLM hallucinations by training with a ternary reward signal that distinguishes correct answers, hallucinations, and abstentions. This technique incentivizes truthfulness over accuracy-only metrics, enabling safer, more calibrated language models through GRPO-based optimization.

Improve model performance at test time by dynamically synthesizing curriculum of problem variants. Co-evolving synthesizer and solver agents create reinforcing feedback for continuous improvement without external labels.

Adapt vision-language models at test time without labels by extracting implicit reward signals (prediction frequency and entropy) and optimizing via GRPO.

Enable long-context modeling via test-time training with meta-learning. Inner loop continues training on context, compressing information into weights rather than KV cache, outer loop optimizes initialization—maintaining full-attention quality with RNN-like constant inference latency across 8K-128K token contexts.

Cascaded VAE+SigLIP encoders creating single continuous representation space supporting both vision understanding and generation, trained jointly on both tasks without format mismatches. Deploy for unified multimodal models where understanding and generation enhance each other.

Achieve 100-200× video generation speedup via algorithm-system co-optimization. Combines sparse attention acceleration (SageAttention + trainable Sparse-Linear Attention), step distillation, W8A8 quantization, and custom CUDA kernels—maintaining quality through modular optimizations that compose for cumulative gains.

Stabilize multi-turn agent RL by shifting from token-level to turn-level MDPs. Reformulates states and actions at conversation-turn granularity, uses learned turn-level critics, and applies Generalized Advantage Estimation—eliminating misalignment that destabilizes GRPO training on long-horizon agentic tasks.

Mixture-of-Transformers jointly learning language modeling and video flow matching, enabling interleaved text-video generation where semantic decisions happen in language, pixel generation in video, and users can intervene textually at any step.

Implement techniques from TwinBrainVLA: Unleashing the Potential of Generalist VLMs for Embodied Tasks via Asymmetric Mixture-of-Transformers. The fundamental premise of Vision-Language-Action (VLA) models is to harness the extensive general capabilities of pre-trained Vision-Language Models (VLMs) for generalized embodied intelligence

Train single-step image generators without teacher models or standard adversarial networks. Achieves 0.83 GenEval score at 1-NFE with 100× computational efficiency gains—when you need real-time image synthesis from pre-trained diffusion models.

Reduce GRPO training cost by 87.5% using only 2 rollouts instead of 16 while achieving 98.1% of baseline performance. Leverage the insight that GRPO's group mechanism serves contrastive learning rather than advantage estimation.

Document extraction is a core component of digital workflows, yet existing vision-language models (VLMs) predominantly favor high-resource languages. Thai presents additional challenges due to script complexity from non-latin letters, the absence of explicit word boundaries, and the prevalence of highly unstructured real-world documents, limiting the effectiveness of current open-source models. This paper presents Typhoon OCR, an open VLM for document extraction tailored for Thai and English....

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

Scale memory networks to 120B parameters with improved long-context learning through integrated memory layers, simplified value projection, and optimized parameter ratios for superior memory-intensive tasks.

Transform uncertainty quantification in LLMs from passive reliability measurement into active control signals for reasoning optimization, autonomous agent decision-making, and reinforcement learning. Use when building systems where uncertainty drives real-time behavior modification and improved reliability.

Compress visual embeddings into compact latent space for unified image understanding and generation. Combines attention-based compression with diffusion decoding to bridge comprehension and generation through a shared semantic bottleneck.

Advanced computer vision technique for robust spatial understanding in complex scenes, supporting agent navigation and environmental reasoning capabilities.

Advanced reasoning approach for optimizing inference efficiency through meta-cognitive planning, enabling agents to make better decisions with reduced computational overhead.

Single cloud-side weight-sorting and fine-tuning supporting multiple on-device pruning rates via efficient SVD and MLP decomposition, achieving 4-5.7× memory reduction and 2.7-3.4× throughput across Transformers, SSMs, and hybrid architectures.

Improves LLM reasoning by rewarding correct solutions that exhibit rare high-level strategies, preventing exploration collapse and discovering more diverse solution approaches across mathematics, physics, and medical reasoning.

Enhance Universal Transformers for complex reasoning through ConvSwiGLU modules integrating depthwise convolution into feed-forward blocks and truncated backpropagation through loops (TBPTL) restricting gradient computation to final iterations. Achieve state-of-the-art on ARC-AGI: 53.8% on ARC-1, 16.0% on ARC-2.

Combine semantic encoders from multimodal LLMs with contrastive learning to create unified high-resolution encoders for both visual understanding and generation tasks without relying on VAE compression.

Learn which tokens to unmask during diffusion sampling via reinforcement learning instead of heuristics. Policies eliminate manual tuning and scale across block sizes—crucial when semi-autoregressive generation needs dynamic, learned unmasking strategies.

Scale computer-use agents from 30% to 72% success rate by generating parallel rollouts and selecting best trajectories through behavior narrative evaluation. Use when deploying desktop agents on complex, high-variance task scenarios.

Unified MLLM processing four urban data types simultaneously: geospatial structures, trajectory information, satellite imagery, and street-view photos. Outperforms general-purpose models on 12-task urban benchmark with 31-375% improvements. Use for urban planning, traffic analysis, location intelligence, and smart city applications requiring integrated spatial reasoning.

Enable vision-centric interactive reasoning by synthesizing diverse reasoning datasets through co-evolutionary loops, then training models with progressive curriculum that starts with perception and advances to tool-based problem solving.

Align autoregressive image models with pixel-space quality via variational optimization. Formulates alignment as ELBO combining reconstruction (pixel supervision) and prior regularization (token distribution), treating model as RL policy with tokenizer reconstruction as reward—achieving 86.6% cost reduction vs standard RL fine-tuning.

VCRL improves LLM mathematical reasoning by dynamically adjusting training sample difficulty based on group reward variance. Uses variance-based sampling with memory replay to focus on moderately-difficult samples where models succeed ~50% of the time, achieving 4.67-point improvement over GRPO on 8B models.

Generate safety policies as executable code with formal verification, enabling provably-correct agent behavior through offline policy generation and online runtime enforcement.

Validate LLM multi-step reasoning chains by converting natural language steps to first-order logic and checking logical consistency against established facts and sourced premises—enabling error detection and self-correction for more reliable reasoning.

Use model confidence and prediction consistency to select high-quality actions at test time without external verifiers, enabling reliable action selection in VLA systems.

Train agents to leverage external tools across domains using VerlTool's unified RL framework. Coordinate code execution, search, SQL queries, and vision utilities in multi-turn interactions without domain-specific redesign. 2× faster asynchronous rollouts on mathematical reasoning, knowledge QA, and software engineering tasks.

Enable flexible control over video diffusion models through multi-modal control signals (edges, masks, poses) without retraining. Apply lightweight Transformer-based auxiliary modules to add Canny edge, segmentation, and pose constraints to frozen pre-trained generators.

Enable parameter-efficient computation through dual-pathway feed-forward networks. Create virtual experts via hidden subspace slicing (width-versatile) and recursive weight application (depth-versatile). Use difficulty-aware gating to balance pathways, adding computation not memory to fixed parameter budget.

Control video generation via 4D geometric representation combining static background point clouds and per-object 3D Gaussian trajectories. Enable category-agnostic control over camera and multi-object motion in realistic video synthesis.

Optimize vision-language models for document tasks via embedding visualization and clustering-guided data generation. Identify error-prone regions in visual space and synthetically augment training data targeting weak areas.

Stabilize off-policy RL training for LLMs by deriving principled importance weight reshaping from variational optimization. Instead of heuristic clamping, VESPO uses closed-form exponential weighting W^c1 * exp(c2*(1-W)) to suppress overweighted samples while maintaining smooth gradients. Enables stable training at 64× policy staleness and under fully asynchronous execution with sequence-level operations that avoid length-dependent biases.

Use frozen vision foundation models like DINOv2 and CLIP as image tokenizers for autoregressive generation. Region-adaptive quantization identifies semantically coherent areas and reduces redundancy. Achieves 256-token encoding (vs. 576), 3× AR model speedup, state-of-the-art 1.36 gFID on ImageNet while eliminating classifier-free guidance.

Implements video deep research for multi-hop reasoning combining video analysis, web search, and evidence synthesis. Evaluates workflow vs agentic paradigms with 100-sample benchmark across 6 semantic domains, revealing goal drift and long-horizon consistency as core bottlenecks.

Use pre-trained video generation models (VGMs) as efficient reward models by evaluating video quality directly in latent space at any denoising timestep, enabling process-reward learning across the entire generation trajectory without expensive VAE decoding to RGB.

Ground video reasoning in explicit visual evidence by highlighting timestamps, objects, and bounding boxes, making reasoning verifiable and improving accuracy through RL rewards for spatio-temporal alignment.

Train video QA models on 6K examples using pure RL instead of costly supervised fine-tuning, then adaptively sample video frames at inference by monitoring answer consensus across multiple reasoning trajectories.

Diagnose and improve temporal pattern recognition in video-language models using SpookyBench, which isolates temporal information from spatial cues.