Claude Skills by ADu2021

Improve exploration in LLM-based agents through external memory-augmented RL with hybrid on/off-policy training. Agents generate exploration 'tips' (self-reflections) after trajectories, storing them in memory. During rollouts, policy samples between standard execution and memory-conditioned execution. Off-policy updates distill memory-guided behaviors into base policy via reward-guided knowledge distillation. Achieves 128.6% improvement on ScienceWorld and 11.3% on WebShop vs. GRPO.

Choose optimal pretraining strategy for text encoders: pure MLM, pure CLM, or biphasic CLM-then-MLM training, with empirical guidance on performance across downstream tasks.

Implement techniques from Endless Terminals: Scaling RL Environments for Terminal Agents. Environments are the bottleneck for self-improving agents

Enable step-by-step reasoning in diffusion models through iterative latent state refinement. Condition diffusion on evolving thought states across multiple reasoning steps, grounded with textual supervision to prevent drift.

Enhanced language model pre-training methodology improving linguistic competence across languages, strengthening foundational capabilities for multilingual agent systems.

Train LLMs to disambiguate tool calls in enterprise settings where multiple similar APIs exist and parameters are incomplete. Generates synthetic multi-turn dialogues with realistic ambiguity to improve tool selection accuracy by 27+ percentage points.

Technique for efficient model adaptation that mitigates catastrophic forgetting during fine-tuning, enabling agents to learn new tasks while preserving existing capabilities.

One-line code modification augmenting RL advantage function with clipped entropy term to encourage exploratory reasoning chains while maintaining optimization stability.

Stabilize RL training on reasoning tasks by performing entropy-guided rollouts from uncertain decision points, avoiding policy collapse and premature convergence. Increases fully correct trajectories on math reasoning while maintaining stable entropy throughout training.

Monitor Shannon entropy in LLM token distributions to detect sudden uncertainty spikes, then trigger adaptive prompt consolidation to realign conversation context and improve accuracy by 56.6% and reliability by 35.3%.

Stabilize LLM post-training by constraining global distributional shifts in policy exploration. Entropy Ratio Clipping supplements local clipping mechanisms with global entropy constraints—essential when PPO alone produces unstable gradients and distribution shifts.

Automatically generate diverse, scalable tool-interactive training environments for LLM agents without manual sandbox creation. Uses topic mining and logic modeling to create varied environment architectures with task scenarios, enabling agents to learn complex multi-turn, multi-tool interactions. Synthesis framework tested on 191 environments with ~7,000 scenarios, improving Qwen3 model performance on knowledge-intensive and search tasks.

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

Stabilize multi-turn LLM agent training with entropy-regularized policy optimization that prevents exploration-exploitation cascade failures in sparse-reward environments through trajectory-level entropy regulation, historical smoothing, and adaptive phase-based weighting. Achieve up to 152% performance improvement on scientific reasoning tasks and 19.8% on embodied control by maintaining controlled entropy oscillations across 30+ interaction turns.

Transform vision-language models into embodied agents through two stages: learning embodied priors from trajectory-augmented data with LLM reasoning, then online RL with self-summarization and dense rewards for long-horizon tasks.

24-trillion-token web dataset with 12-category taxonomy labels enabling efficient curation of specialized datasets through SQL filtering and small annotation models.

Chunk-based landmark-guided sparse attention enabling 16M-token context windows with 90%+ retrieval accuracy on in-context tasks. Use when processing ultra-long documents where full attention is prohibitive but context selection must be dynamic.

Train autonomous agents to use computers by generating synthetic task experiences and iterating on them, achieving 56.7% success on OSWorld benchmarks through scalable experience-driven optimization. Use when you need agents that autonomously learn complex computer interaction patterns without manual task curation.

Scale Evolution Strategies to billion-parameter LLMs without backpropagation for superior robustness and stability across diverse models, reward horizons, and evaluation tasks. Outperforms RL methods while eliminating gradient computation overhead.

Generate verifiable synthetic training data (problems + solutions + verification artifacts) through evolutionary synthesis that works across math, code, and agent tasks without task-specific rules.

System for evolving and composing programmatic skills through learning, enabling agents to discover and develop new behavioral capabilities dynamically.

Build unified LLMs supporting both fast standard inference and slow reasoning modes through hybrid attention and adaptive post-training. Combine non-reasoning and reasoning capabilities in a single model with 128K context windows and tool use. Use when you need models balancing speed and reasoning depth across diverse tasks and languages.

Build robust reward models for computer-use agents by processing execution videos with spatiotemporal token pruning (STP+TTP) to remove redundancy while preserving task-relevant UI details. Achieve 84.7% accuracy with adversarial instruction translation.

Improve RLVR training efficiency by selectively replaying trajectories based on correctness and entropy. Medium-difficulty questions and low-entropy solutions are most valuable; selective replay yields +3.5-7.6% improvements.

Improve LLM reasoning efficiency by systematically reusing past rollouts through experience replay. ExGRPO organizes training data by success and diversity, applying a mixed-policy objective that prioritizes high-quality examples while maintaining exploration, achieving 3.5-7.6 point gains over on-policy methods.

Evaluate AI systems' ability to conduct autonomous research experiments using EXP-Bench, a benchmark for multi-step scientific reasoning and iterative experimental workflows.

Dynamically adapt LLM reasoning strategies at inference time by curating episodic memory of past problem solutions—generate task-specific prompts, tool configs, and control logic for up to 111× cost reduction and 14% accuracy gains.

Improve RL training efficiency by embedding explicit reflection and revision loops. Models generate initial responses, receive feedback, produce self-reflections describing improvements, revise their attempts, and distill successful corrections into the base policy. Achieves up to 81% improvement on complex tasks through structured behavioral change.

Improve MoE language model efficiency with causal threshold-based routing that eliminates auxiliary losses and enables dynamic per-token computation.

Investigate exploration-exploitation trade-offs in reinforcement learning with verifiable rewards through theoretical analysis and empirical validation. Derive explicit clipping bias bounds, establish policy-entropy shift formulation, and introduce reward-misalignment framework. Show policy entropy and performance lack direct causal relationships.

Multimodal system for joint audio-video understanding and generation, enhancing agent perception and communication in environments with both audio and visual information.

Unify LLM alignment methods through f-divergence theory. f-GRPO extends GRPO to handle any divergence measure (KL, Jensen-Shannon, Hellinger), enabling tailored alignment objectives. f-HAL combines on-policy and off-policy preference learning to prevent reward hacking while maintaining safety alignment.

Prevent policy collapse onto common solutions during reinforcement learning by applying difficulty-aware focal weighting to gradient contributions, maintaining diversity across solution modes while preserving performance.

F1 is a 4.2B-parameter Vision-Language-Action model that reformulates robotic control as foresight-guided inverse dynamics. Rather than reactive state-to-action mappings, F1 predicts plausible future visual states and derives actions to achieve them, enabling robust planning in dynamic environments with 82.2% real-world task success rate.

Enable extreme 1-2bit quantization of pre-trained LLMs by leveraging complex-valued arithmetic. Convert real-valued linear layers to complex domain losslessly, quantize to fourth roots of unity {±1, ±i}, and apply recursive residual error quantization for near full-precision performance.

Achieving human-level performance in Vision-and-Language Navigation (VLN) requires an embodied agent to jointly understand multimodal instructions and visual-spatial context while reasoning over long action sequences. Recent works, such as NavCoT and NavGPT-2, demonstrate the potential of Chain-of-Thought (CoT) reasoning for improving interpretability and long-horizon planning. Moreover, multimodal extensions like OctoNav-R1 and CoT-VLA further validate CoT as a promising pathway toward human...

Generates high-quality images directly from pixels using flow-matching-based latent sequences. Transforms images via invertible flows into manageable latent sequences, applies autoregressive modeling, and uses classifier-free guidance. Provides exact likelihood estimates and one-step distillation capabilities.

Identify dominant frequency components in RoPE embeddings to determine which attention dimensions are essential, reducing KV cache by 81% while maintaining performance. Training-free approach integrates with existing compression methods for 2.56x speedup on long-context reasoning.

Implements Fast-ThinkAct from arXiv:2601.09708

Train efficient reasoning models using stage-wise context scaling and complexity-aware data selection. Achieves 49.6% accuracy on AIME 2024 while reducing training steps by 50% through alternating compress-extend cycles that progressively refine reasoning quality.

Identify and measure feedback friction in LLM reasoning tasks where models resist high-quality guidance, discovering that confidence predicts feedback receptiveness and revealing mitigation strategies.

Systematically adapt state-of-the-art distillation methods for T2I generation. Compare sCM (stabilized Consistency Models), MeanFlow, and IMM within unified framework. sCM excels at extreme few-step regimes (52.81% GenEval at 2 steps), MeanFlow achieves superior fidelity at 4 NFEs.

Demonstrate that adversarial attacks on vision-language models need not target all tokens equally. Entropy-guided attacks identify high-entropy tokens (critical decision points) where perturbations have maximum impact, achieving comparable attack success with 80% fewer tokens targeted.

Edit 3D faces with flexible mask layouts using only a few training samples. FFaceNeRF employs geometry adapters with feature injection and latent mixing for tri-plane augmentation, enabling rapid NeRF adaptation without fixed segmentation masks. Ideal for personalized medical imaging, creative face editing, and applications requiring user-defined mask control.

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

Active learning (AL) has the potential to drastically reduce annotation costs in 3D biomedical image segmentation, where expert labeling of volumetric data is both time-consuming and expensive. Yet, existing AL methods are unable to consistently outperform improved random sampling baselines adapted to 3D data, leaving the field without a reliable solution. We introduce Class-stratified Scheduled Power Predictive Entropy (ClaSP PE), a simple and effective query strategy that addresses two key ...

Provide continuous confidence scores throughout LLM text generation via supervised learning and backward confidence integration, enabling real-time uncertainty awareness

Financial agents powered by large language models (LLMs) are increasingly deployed for investment analysis, risk assessment, and automated decision-making, where their abilities to plan, invoke tools, and manipulate mutable state introduce new security risks in high-stakes and highly regulated financial environments. However, existing safety evaluations largely focus on language-model-level content compliance or abstract agent settings, failing to capture execution-grounded risks arising from...

Enable video generation customization via first-frame reuse—treat initial frame as a visual entity buffer storing objects for reuse during generation with just 20-50 examples, requiring minimal architectural change.

Train small models to recover from tool execution errors by treating errors as training signals. Use error simulators to generate recovery examples and fission failed trajectories into multiple recovery attempts.