Claude Skills by ADu2021

Scale information-seeking agents via nested architecture separating outer reasoning from inner page exploration. Minimal toolkit (search, visit, click, fill) handles dynamic web interactions, inner loop filters content before context injection, multi-task learning trains outer/inner jointly—enabling 4B models to match larger competitors.

Discover that large pretrained models have dense neighborhoods of task-specific experts—random weight perturbations improve performance. Use RandOpt: sample perturbations, select top performers, ensemble via voting for multi-task adaptation.

Simulate GUI behavior by predicting screen frames in response to user inputs. NeuralOS combines hierarchical RNNs for state tracking with diffusion-based rendering, capturing mouse interactions and application state transitions. Trains on synthetic demonstrations plus random exploration; achieves 50-61% human indistinguishability on basic operations while maintaining 18 fps inference on single H100.

Build unified decoder-only transformers for multimodal tasks using 6 trillion interleaved text-image tokens with next-scale prediction for visual content—enabling fast 1024x1024 image generation (5 seconds), image editing, and video generation while rivaling specialized diffusion models.

Train a unified autoregressive model to generate images and text by directly handling continuous image tokens with flow matching, eliminating the need for quantization or separate diffusion models.

Open-source foundation model trained for diverse agent interactions across gaming and interactive environments, supporting generalist agent development with robust performance foundations.

Scale distributed LLM training without all-reduce synchronization using dynamic pipeline routing and modified Nesterov momentum, achieving 4% faster convergence than DiLoCo with exponentially lower communication.

Agentic systems have recently become the dominant paradigm for formal theorem proving, achieving strong performance by coordinating multiple models and tools. However, existing approaches often rely on task-specific pipelines and trained formal provers, limiting their flexibility and reproducibility. In this paper, we propose the paradigm that directly uses a general coding agent as a formal math reasoner. This paradigm is motivated by (1) A general coding agent provides a natural interface f...

Enable vision-language models to self-correct by synthesizing dense training examples from existing rollouts, creating n² correction pairs from n original trajectories.

Solve IMO-level problems through multi-stage hierarchical reasoning with lemma-based memory compression. Intern-S1-MO achieves 26/35 on IMO2025 geometry-excluded problems—ideal when complex reasoning exceeds single-pass context.

Scale mixture-of-experts models efficiently by decomposing experts into atomic vector pairs with Cartesian product routing and expert-centric scheduling. Achieves 10.9× speedup and 50% fewer parameters versus fine-grained baselines through system-algorithm codesign that converts scattered memory access into contiguous batched operations.

Builds generalist reward models evaluating outputs across 5 modalities (text, image, video, audio, 3D) using free-form preference data. Combines discriminative and generative reward modeling approaches. Covers 9 tasks with 317K preference examples, enabling preference-based alignment for diverse output modalities.

Videos convey richer information than images or text, capturing both spatial and temporal dynamics. However, most existing video customization methods rely on reference images or task-specific temporal priors, failing to fully exploit the rich spatio-temporal information inherent in videos, thereby limiting flexibility and generalization in video generation. To address these limitations, we propose OmniTransfer, a unified framework for spatio-temporal video transfer. It leverages multi-view i...

Compress verbose reasoning by conditioning the same model on a conciseness instruction to create a teacher. Minimize KL divergence between student outputs and concise teacher without ground truth, enabling implicit reward learning that improves accuracy while reducing tokens.

Demonstrate that a single strategically engineered training sample can improve reasoning across multiple domains. Polymath learning shows sample quality and multidisciplinary design matter more than quantity, enabling extreme data efficiency in RL training.

Improve deployed language models by learning from real-world user interactions. Extract transferable knowledge from interaction trajectories and consolidate via on-policy context distillation without needing environment access.

Improve parallel reasoning by explicitly generating diverse outlines before executing solution paths. Overcomes mode collapse where independent samples converge on same (often wrong) answer. Generates unique answers (27.6 vs 23.5) with focused reasoning (10% shorter correct paths).

Scales computer-use agent capabilities through reflective Chain-of-Thought reasoning in large-scale annotated datasets spanning multiple operating systems and 200+ applications.

Transfer reasoning behaviors learned in language models to visual domains through two-stage training: cold-start linguistic fine-tuning followed by multimodal RL. Open-Vision-Reasoner achieves 95.3% on MATH500 and 54.6% on MathVerse by learning visual analogs of backtracking, verification, and subgoal decomposition using rule-based rewards.

Designs terminal-based AI coding agents through workload-specialized model routing, where distinct models handle planning, thinking, critique, and execution tasks. Implements extended ReAct loop with context compaction and approval gates for safe command execution.

Build agentic systems for transparent, evidence-based novelty analysis of research submissions through four-phase pipelines: contribution extraction, prior work retrieval, hierarchical comparison, and structured reporting with explicit citations—enabling fair peer review at scale.

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

Enable open-ended mathematical learning through co-evolutionary teacher-student self-play where a single model alternates roles to generate progressively harder problems and solve them, creating a virtuous cycle without external data.

Design data generation pipelines for reasoning models through systematic experimentation with answer sampling, teacher selection, and source quality optimization.

Implements OpenTinker from arXiv:2601.07376

Learn a single visual representation supporting both image understanding and generation by combining VAE-based generative training with contrastive learning objectives. Use when building multimodal systems that need unified image representations for both comprehension and generation tasks.

Train vision-language models for complex reasoning by alternating SFT (supervised fine-tuning via text-only reasoning models) and curriculum RL (Group Relative Policy Optimization). Progressively improve through iterative cycles where each iteration generates better training data.

Verify long reasoning chains by combining outcome and process verification through active learning. OPV achieves 83.1 F1 on verification—crucial when both final answers and reasoning paths must be validated with minimal annotation.

Compact Mentalese symbolic format trained via SFT, then refined with SLPO (Shorter Length Preference Optimization) to reward concise correct solutions without over-penalizing necessarily long reasoning. Compress reasoning while maintaining accuracy.

Detect safety violations in mobile GUI agents using dual mechanisms: formal verification rules for system-level risks (privilege escalation, credential exposure) and VLM-based contextual judges evaluating action appropriateness, achieving 10-30% safety improvement over baselines.

Enable reasoning models to adaptively switch between fast direct responses and slow detailed reasoning, reducing token consumption by 15-40% while maintaining accuracy through dual-mode fine-tuning.

Diagnose and prevent over-searching—unnecessary search invocations that degrade LLM performance and waste tokens. Framework introduces Tokens Per Correctness (TPC) metric and OverSearchQA dataset to train models that know when NOT to search. Teaches agents to recognize answerable vs. unanswerable queries and selectively invoke search only when needed, improving both accuracy and efficiency.

Extract and manipulate internal progress vectors in reasoning models to accelerate thinking phases while maintaining answer quality, achieving 80%+ improvements in token efficiency.

A 3B unified model combining image understanding, text-to-image generation, and image editing end-to-end rather than as separate frozen components. Use when you need a single efficient model for multiple vision-language tasks without the overhead of separate specialized systems.

Weight distillation loss by student pass rate using Beta kernel: suppress mastered (high pass rate) and intractable (low pass rate) problems, prioritize intermediate zone where learning signal is richest.

Improves LLM-based evolutionary search by addressing context pollution, mode collapse, and weak collaboration through hierarchical context management, momentum-based backtracking, and adaptive sampling policies.

PaCoRe: Learning to Scale Test-Time Compute with Parallel Coordinated Reasoning. From arXiv:2601.05593

Verify solution quality through pairwise comparison rather than pointwise scoring. Implement topology coverage and Swiss refinement to allocate verification compute to uncertain pairs, improving calibration and reducing verification overhead.

Writing effective rebuttals is a high-stakes task that demands more than linguistic fluency, as it requires precise alignment between reviewer intent and manuscript details. Current solutions typically treat this as a direct-to-text generation problem, suffering from hallucination, overlooked critiques, and a lack of verifiable grounding. To address these limitations, we introduce $\textbf{RebuttalAgent}$, the first multi-agents framework that reframes rebuttal generation as an evidence-centr...

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

Monitor parallel reasoning branches via 2D probing—periodically extracting intermediate answers to detect consensus and prune divergent branches. Reduces token cost by 25%+ while maintaining accuracy without requiring model retraining.

Train language models to explore multiple reasoning paths simultaneously via reinforcement learning. Uses progressive curriculum learning to address cold-start problems, enabling 8.4% accuracy gains over sequential reasoning on complex mathematical tasks.

Linear RNN architecture enabling efficient parallel computation over directed acyclic graphs with stabilization modes for long-range dependencies without sequential traversal.

Generate 3D objects with explicit part structures enabling compositional editing and animation. Decouples structure planning from geometry synthesis using two stages: autoregressive bounding box generation and part-aware refinement.

Generates semantically-meaningful 3D parts from single images via compositional diffusion transformers with part-level identity and local-global attention.

Use Pass@k as the reward metric in RL training to balance exploration and exploitation, enabling models to learn diverse sampling strategies while maintaining correctness.

Improve post-RL performance by reweighting SFT loss using importance sampling. Prioritize training examples that match the target policy distribution, not the behavior policy. Achieves 14.6% Pass@8 gains on AIME.

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

Comprehensively evaluate 12+ parameter-efficient fine-tuning methods for RL with Verifiable Rewards (RLVR). Show DoRA/AdaLoRA outperform LoRA, SVD-based methods fail on RL, extreme reduction creates bottlenecks—providing empirical evidence that geometric-aware adapters align better with RL's off-principal update dynamics.

Enable language models to actively manage their context using memory tools (read, index, note-take, delete). Models receive agency to self-engineer context rather than passively consume pre-staged information, maintaining compact high-quality reasoning states through strategic deletion of irrelevant information while preserving distilled notes.