Claude Skills by ADu2021

Improve LLM training efficiency through manifold-based optimization that projects momentum onto tangent spaces and constrains updates on rotational Oblique manifolds. Achieves 1.75× faster convergence than Muon with reduced memory.

Dynamically reweight advantage functions based on trajectory certainty to improve policy optimization in foundation models. Addresses advantage reversion and mirror problems by mixing standardized and mean-normalized advantage formulations. Enables more stable gradient signals across high- and low-certainty samples.

Coordinate specialized agents with distinct personality traits (Openness, Agreeableness, Conscientiousness, Extraversion) to solve complex scientific problems across text and vision, using a Critic agent to apply Socratic questioning for iterative refinement and error correction.

Adapt DPO temperature parameters per sample using reward model margins, amplifying learning signals for hard preference pairs while dampening easy ones.

Enable LLMs to scale reasoning length from O(n²) to O(n) by structuring thinking into fixed-size chunks with learnable cross-chunk summaries. Trigger: train reasoning models with unbounded or expensive chain-of-thought sequences.

Train multi-agent reasoning systems with decoupled reward signals and pipeline parallelism—enable specialized Solver/Verifier/Corrector agents to iteratively refine solutions without waiting for full trajectories, handling extended reasoning up to 320K tokens.

Optimize task-specific prompts using five cooperative agents (Planner, Teacher, Critic, Student, Target) in a POMDP framework, where the Planner generates adaptive trajectories and a Teacher-Critic-Student triad refines prompts through Socratic dialogue, achieving 85.11% accuracy on general tasks and 75.81% on specialized domains.

Improves tool-integrated reasoning by using bipartite matching to assign dense turn-level rewards, enabling credit assignment for individual tool interactions in multi-turn tasks where 4B models outperform 8B competitors.

Understand why math reasoning improvements don't always transfer to general capabilities. Use RL-based training instead of SFT to preserve representation structure and enable broader generalization.

Improve mathematical problem-solving in multimodal models by decoupling visual perception from inference reasoning. A two-stage pipeline extracts essential visual information and reasoned properties before passing enriched text to inference models, dramatically improving accuracy on visual math problems.

Improve multimodal mathematical reasoning through iterative reflection cycles where an outcome reward model provides feedback on reasoning quality, and correct solutions are incorporated back into training—enabling continuous model adaptation beyond static datasets.

Enables LLM-based agent teams to improve reasoning accuracy at inference time through collaborative deliberation and structured experience retrieval, achieving 3-8% accuracy gains without expensive multi-agent training.

Evaluate LLM agents on realistic tool-use tasks via 28 live MCP servers with 250 tools, assessing fuzzy tool discovery, multi-step planning, and cross-domain workflow coordination

Evaluate LLM agents through realistic multi-turn tool-use workflows across 127 complex MCP tasks spanning CRUD operations, state management, and error handling. Use when assessing agent capabilities on real-world tool orchestration beyond shallow read-only interactions.

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

Trains mean-velocity models on rectified couplings from pretrained flow models to dramatically smooth loss landscape, enabling faster convergence and superior one-step generation quality without additional training data.

Implement techniques from Mecellem Models: Turkish Models Trained from Scratch and Continually Pre-trained for the Legal Domain. This paper presents Mecellem models, a framework for developing specialized language models for the Turkish legal domain through domain adaptation strategies

Analysis framework for understanding how large language models perform complex cognitive tasks, revealing internal reasoning mechanisms that inform agent architecture and capability assessment.

Build efficient medical vision-language models that reason about medical images and text simultaneously, achieving competitive performance with much larger models while maintaining 500× lower computational cost.

Promptable segmentation foundation models such as SAM3 have demonstrated strong generalization capabilities through interactive and concept-based prompting. However, their direct applicability to medical image segmentation remains limited by severe domain shifts, the absence of privileged spatial prompts, and the need to reason over complex anatomical and volumetric structures. Here we present Medical SAM3, a foundation model for universal prompt-driven medical image segmentation, obtained by...

Generate accurate, high-quality medical videos for clinical education and documentation by leveraging large-scale annotated medical datasets with domain-specific fine-tuning on video diffusion models.

Implement techniques from MeepleLM: A Virtual Playtester Simulating Diverse Subjective Experiences. Recent advancements have expanded the role of Large Language Models in board games from playing agents to creative co-designers

Enable agents to autonomously design and refine task-specific agents by evolving externalised behavioral skills and prompts without modifying base LLM parameters.

Optimize agent long-term memory by treating it as an information bottleneck problem. Dynamically compress redundant information while preserving task-relevant content through semantic-symbolic-topological hybrid retrieval.

Reduces optical flow GPU memory 3.9× while maintaining state-of-the-art accuracy through correlation volume downsampling and dimension compensation. Enables native FullHD training with 2.09GB inference memory. Use for motion estimation in memory-constrained environments or high-resolution video processing.

MemGovern: Enhancing Code Agents through Learning from Governed Human Experiences. From arXiv:2601.06789

Implements MemoBrain from arXiv:2601.08079

Evaluate memorization in 3D generative models through controlled experiments discovering factors like dataset diversity and guidance scale. Provide simple yet effective strategies like rotation augmentation to reduce memorization without degrading generation quality.

Evaluate and improve memory capabilities in LLM agents across four competencies: accurate retrieval, test-time learning, long-range understanding, and selective forgetting. Identifies critical gaps in how agents store, update, and revise information.

Treat memory management as learnable RL policy actions (delete/insert) rather than fixed mechanisms. Enable models to autonomously decide what to keep, remove, or add to context, reducing average context length by 51% while matching larger models.

Enable long-horizon agents to manage finite context by separating working memory from persistent storage. Use indexed summaries with pointers to archived evidence, treating memory operations as first-class agent actions learned via RL.

Decouple feed-forward networks from self-attention by training FFNs on context-free token embeddings instead of residual streams. Enables pre-computation of FFN outputs as static lookup tables for inference efficiency and improved interpretability.

Treat memory as a manageable system resource for LLMs through unified management of plaintext, activation, and parameter-level memories with dynamic scheduling and lifecycle governance.

Enable agents to accurately identify temporally relevant information in long multi-session dialogues through RL-based memory retrieval. Combines coarse-to-fine candidate selection with multi-level temporal consistency rewards—providing dense supervision that disambiguates time expressions and maintains coherence across 128k-token contexts.

Existing works increasingly adopt memory-centric mechanisms to process long contexts in a segment manner, and effective memory management is one of the key capabilities that enables large language models to effectively propagate information across the entire sequence. Therefore, leveraging reward models (RMs) to automatically and reliably evaluate memory quality is critical. In this work, we introduce MemoryRewardBench, the first benchmark to systematically study the ability of RMs to evaluat...

Build agents that evolve their own memory operations by learning a skill bank of memory transformations and periodically discovering new skills from challenging cases, enabling adaptive memory management that improves with scale.

Improve reasoning models by aligning their meta-predictions with actual rollouts through self-generated training signals. Trigger: accelerate reasoning model training while maintaining performance through better meta-cognitive awareness.

Enable LLM agents to evolve behavioral skills and policies online through skill synthesis from failures and opportunistic gradient-based refinement, without service interruption.

Train LLMs to faithfully express uncertainty through natural language that accurately reflects their actual confidence, improving trustworthiness and reducing overconfidence.

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

Build state-of-the-art 7B multimodal models by combining four-stage vision-language pretraining with mixed on-policy RL integrating verifiable and human feedback rewards.

System enabling intelligent tool integration and selection for enhanced multimodal reasoning, allowing agents to leverage heterogeneous capabilities for complex problem-solving across domains.

Combine sparse attention (25% of layers) and linear attention (75% of layers) via strategic layer placement to handle 1M-token contexts with 75% training cost reduction. Hybrid positional encoding preserves long-range information while maintaining position awareness.

Build ultra-efficient language models for edge devices using sparse attention, high-quality data filtering, and ternary quantization, achieving Qwen3-8B performance with 22% of training tokens.

Hybrid-attention MoE reasoning model supporting 1M token context and 80K token generation, combining lightning attention with CISPO RL algorithm for efficient scaling.

Improve diffusion model alignment with human preferences by handling subjective and conflicting annotations. Adaptive-DPO incorporates minority-instance metrics (intra-annotator confidence and inter-annotator stability) to distinguish majority from minority samples, enhancing performance on both synthetic and real preference data.

Integrates fine-grained visual tokens into mathematical reasoning via Interleave Tokens that dynamically select relevant image regions for each reasoning step.

Run draft and target models in parallel across GPU/NPU using bidirectional speculation: draft predicts forward continuations while target speculates correction paths simultaneously. Achieve 2.8x-5.8x wall-time speedup on 14B-66B models.

Score reasoning paths using PMI to identify question-specific relevant steps without rollout simulations, enabling efficient tree search 12× faster than MCTS while improving reasoning accuracy.

Improve reasoning efficiency by dynamically adjusting elaboration depth within a single response. Difficult steps receive extensive chain-of-thought; straightforward steps use concise inference. Mix reasoning modes to achieve 57-79% token savings under compute budgets while maintaining accuracy on math, STEM, and coding benchmarks.