Claude Skills by ADu2021

Perform robust, low-cost model updates via REPAIR: closed-loop feedback, dynamic memory management, and frequent knowledge fusion. Preserve non-target knowledge while achieving 10-30% improvements in editing accuracy. Enable sequential edits without catastrophic forgetting through strong locality guards and stable integration mechanisms.

Enable language models to dynamically assign continuous position values to tokens based on contextual relevance instead of fixed linear positions. Implement learnable SwiGLU module extracting position information, integrate with RoPE for end-to-end optimization. Improves performance on noisy contexts and long-sequence tasks.

Training-free token importance metric using representation shifts in MLP layers, enabling efficient compression compatible with fused kernels.

Dual-scoring mechanism (Magnitude and Stability scores) enabling process-level rewards that penalize inefficient reasoning patterns like overthinking and backtracking without trained reward models. Improve reasoning efficiency in RL without additional supervision.

Apply software release engineering practices to self-improving LLM agents to achieve stable evolution. AgentDevel uses implementation-blind critique, executable diagnosis, and flip-centered gating to prevent regression while enabling auditable improvement trajectories.

Extract and transfer reasoning abilities using sparse autoencoders (SAE-Tuning) on CoT-free data, achieving RL-equivalent performance at 2000x lower cost and 450x faster training.

Evaluate large reasoning models under stress by asking multiple problems simultaneously, revealing true multi-tasking capacity and context management. Use to identify robustness gaps invisible in single-problem evaluations and discriminate model capabilities beyond traditional benchmarks.

Generative approach for creating complex dynamic scenes and content, supporting agent capabilities in understanding and reasoning about multi-agent environments.

Enable longer reasoning within bounded context through iterative refinement: generate solution → verify → compress → refine. Achieves +11% on AIME with lower per-call latency than naive chain-of-thought.

Train LLM agents to evolve by generating dual intrinsic feedback signals—numerical rewards for capability progress and language lessons for future reuse. Combines hindsight reflection with memory-augmented RL using SimUtil-UCB retrieval.

Convert pretrained non-recurrent language models into depth-recurrent variants through a curriculum of increasing recurrence—decoupling training compute from inference compute and improving performance at given inference budgets on reasoning tasks.

Develop sophisticated multimodal reasoning through text-centric cold-start initialization, prioritized advantage distillation, and staged RL refinement.

Control MLLM output characteristics at inference time using separate reward models for precision and recall, enabling dynamic trade-offs without retraining.

Improve multimodal LLMs on fine-grained visual reasoning tasks (e.g., reading transit maps) by decomposing training into stages: basic perception (VQA) -> simple reasoning -> complex spatial reasoning. Incorporates 'detail rewards' for intermediate visual understanding, bootstrapping models from simple to complex tasks while addressing sparse reward challenges.

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

Enable robot policies to self-improve through imagination using learned dynamics and value models without physical trial-and-error. Compositional world model separates concerns enabling 35-45% performance gains on contact-rich manipulation.

Enable simultaneous optimization of environment difficulty, policy, and reward model. System uses reward model evaluations to guide environment adaptation, creating positive feedback loop for scalable agent improvement.

Correct color distortion in nighttime photos by combining statistical gray-pixel detection with reinforcement learning parameter optimization. Achieves superior cross-camera generalization without extensive labeled nighttime training data through a hybrid architecture that preserves interpretability while gaining adaptive tuning capability.

Scale LLM training using RL on unlabeled pre-training corpora without human annotation. Derive reward signals directly from text segments to optimize both autoregressive generation and in-context reasoning across knowledge and mathematical domains.

Hybrid-policy optimization combining multiple importance sampling and exploration-based advantage functions. Prevents capability boundary collapse while maintaining verification rewards, enabling LLMs to exceed baseline boundaries.

Extend RL to open-ended tasks using structured rubrics as reward anchors, enabling fine-grained evaluation of subjective outputs without requiring binary correctness signals.

RLVR extends reasoning capabilities by proving answer-only rewards implicitly incentivize correct intermediate reasoning via the Logic Prior principle.

RL in vision-language models improves inference-time alignment between vision and reasoning in mid-to-late layers, not vision ability or reasoning separately. Strategic layer freezing enables diagnosis of which components contribute to RL gains.

Train reasoning models to discover diverse solution strategies through two-player RL that jointly optimizes abstraction generation and solution derivation. Use when exploring multiple reasoning approaches is preferable to single-strategy depth.

Train LLMs to discover domain-specific abstractions (concise representations of procedural and factual knowledge) through two-player RL. An abstraction generator proposes key insights, a solution generator uses them to solve problems, and both receive RL rewards, learning structured reasoning that transfers to harder problems.

RLAD performs selective distillation during RL, using trust-region ratio objectives to balance exploration, exploitation, and teacher guidance for efficient reasoning model training.

Use reinforcement learning to identify reasoning-critical attention heads and apply selective KV cache compression, reducing cache by 20-50% while preserving reasoning quality. Maintains speedups up to 1.21x with minimal performance loss.

Improve reasoning during pretraining (not just post-training) by computing rewards from information gain—how much reasoning improves log-likelihood of observed tokens. Works at 1T token scale across diverse corpora.

Integrate reinforcement learning into the pretraining phase by measuring the utility of intermediate reasoning for predicting subsequent tokens. This approach generates dense reward signals during standard pretraining, enabling models to develop reasoning abilities earlier and with significant performance gains (19% improvement on 1.7B, 45% lift on 12B models).

Scale RL training for language models by automatically adapting problem difficulty to match model capabilities using verifiable reward environments—achieving 3.37% absolute improvement on reasoning benchmarks through environment diversity rather than compute scaling alone.

Overcome reward function design challenges via General Reward Model (GRM) for step-wise progress assessment. Uses multi-view observations for occlusion robustness, hop-based progress normalization, Policy-Invariant Reward Shaping—enabling 95% robot task success within 150 interactions with theoretically-grounded dense rewards.

Learn coordinated manipulation behaviors for multi-robot systems using compositional constraints that enforce safe and efficient collaboration. Generate training data through automated collection with task-specific constraint interfaces, then train imitation learning policies adaptable to varying difficulty levels.

Brain-inspired framework integrating spatial, temporal, episodic, and semantic memory systems for embodied agents. Achieves 26.5% performance improvement via dynamic spatial knowledge graphs and closed-loop planning with critic modules.

Generate physically plausible robot manipulation videos by jointly learning RGB generation, temporal depth prediction, and keypoint dynamics. Enables training manipulation policies on synthetic data with strong correlation to simulator performance.

Train vision-language models for robotic manipulation using RL to improve embodied reasoning about spatial relationships and movements, achieving 31% improvement on manipulation benchmarks.

Generate synthetic robot manipulation data via diffusion models using visual identity prompting from exemplar images. Improve multi-view temporal coherence and scalability for robot policy training without extensive real-world data collection.

Enable MLLMs to handle visually degraded images by explicitly reasoning about degradation types and severity. Trains models to perceive degradation parameters, analyze semantic impact, and reconstruct interpretations while adapting reasoning depth to degradation complexity—achieving robust understanding with interpretable explanations.

Compress language models by formulating heterogeneous layer-wise allocation as a constrained knapsack problem. Measure per-layer compression error, solve via dynamic programming to minimize total error within target model size, and avoid pathological solutions where some layers are severely damaged.

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

Train efficient 14B-parameter agents via GRPO with resample-on-correct rollout strategy and Python code execution, achieving state-of-the-art reasoning through autonomous exploration and iterative refinement

Guide LLM exploration through rubric-based scaffolding that gradually diminishes, enabling models to internalize reasoning patterns while maintaining exploration quality for robust RL training.

Train LLM agents to anticipate environment consequences by learning world models through reinforcement learning with embedding-space similarity rewards, avoiding task-specific labels while enabling robust environment adaptation.

Recover safety alignment in fine-tuned LLMs using only a single safety example, without sacrificing utility. Leverages low-rank structure of safety gradients to enable minimal-cost correction even when models have been compromised by extensive harmful training data.

Implement Saffron's multifurcation reward model approach to achieve efficient inference-time safety scaling, improving robustness against prompt injection attacks while reducing computational overhead.

Enable agents to reason over long videos through multi-turn reasoning for complex questions and single-turn answering for simpler queries. Equip orchestrator VLM with multiple tools (web search, speech transcription, temporal grounding). Use synthetic data generation and multi-reward RL achieving 6.1% improvements on open-ended tasks.

Build effective retrieval systems for deep research agents by understanding query-retriever mismatch, preferring BM25 for keyword-oriented queries, and augmenting corpus documents with metadata and LLM-generated keywords to improve information discovery.

Enable agents to continuously improve by accumulating reusable skills across sequential task chains. Train via GRPO across task sequences where skills persist and compound, provide dual rewards for both task completion and skill generation/reuse—improving completion rates 8.9% while reducing token costs by 59% compared to non-skill baselines.

Inject privileged hints during GRPO training to reshape rollout distributions when advantage collapses occur, increasing outcome diversity without changing task rewards. Hints removed at deployment; policy automatically learns when to use hints via online refresh mechanism.

Teach multimodal models to determine when deep reasoning is necessary and how to reason effectively through dual-reward reinforcement learning, preventing both overthinking on simple tasks and underthinking on complex ones while reducing hallucinations.

Implement techniques from SALAD: Achieve High-Sparsity Attention via Efficient Linear Attention Tuning for Video Diffusion Transformer. Diffusion Transformers have recently demonstrated remarkable performance in video generation