Claude Skills by ADu2021

Evolve agent behavior through iterative context refinement using delta updates rather than full rewrites, accumulating strategies and insights across execution traces.

Improves LLM agent decision-making by training agents to first critically evaluate actions before generating, using RL on action-pair comparisons. Develops intrinsic reasoning about action quality without requiring reflection supervision.

Train agentic LLMs through curriculum-based learning to autonomously execute full data science workflows from raw data to analysis reports, enabling 8B models to match proprietary systems.

Agentic search has recently emerged as a powerful paradigm, where an agent interleaves multi-step reasoning with on-demand retrieval to solve complex questions. Despite its success, how to design a retriever for agentic search remains largely underexplored. Existing search agents typically rely on similarity-based retrievers, while similar passages are not always useful for final answer generation. In this paper, we propose a novel retriever training framework tailored for agentic search. Unl...

Reasoning is a fundamental cognitive process underlying inference, problem-solving, and decision-making. While large language models (LLMs) demonstrate strong reasoning capabilities in closed-world settings, they struggle in open-ended and dynamic environments. Agentic reasoning marks a paradigm shift by reframing LLMs as autonomous agents that plan, act, and learn through continual interaction. In this survey, we organize agentic reasoning along three complementary dimensions. First, we char...

Framework for enabling autonomous agents to self-verify code generation and reasoning quality through structured evaluation, supporting software engineering agent deployment with built-in correctness checking mechanisms.

Enables agents to maintain strategic coherence over extended experimental cycles through hierarchical cognitive caching that distills execution traces into stable knowledge, achieving 56.44% on MLE-Bench within 24-hour budgets.

Enables long-horizon agentic search extending beyond 100 tool calls through scalable asynchronous RL training with autonomous QA dataset synthesis.

Transform uncertainty estimates into active control signals for agents, combining implicit confidence mechanisms with targeted reflection to prevent error propagation in long-horizon reasoning tasks. Use when building autonomous agents that must navigate complex multi-step problems while managing confidence and uncertainty.

Build agentic applications using unified agent interfaces, asynchronous design patterns, ReAct paradigm grounding, and developer-centric evaluation and deployment tools.

Reveals that inference-time scaling techniques for LLMs don't transfer to VLMs: majority voting beats verification, self-correction happens in <10% of cases, and models verify better without images. Use insights to design VLM evaluation methods that work rather than assuming LLM techniques apply directly.

Train LLMs to generate high-quality research plans via rubric-based RL without requiring experimental verification. Extracts research goals and domain-specific rubrics from scientific papers, uses frozen model as grader with 12-22% relative improvements, achieves human-expert preference 70% of time with strong cross-domain generalization.

Select optimal training subsets for T2I models through meta-gradient-based rater networks. Score each sample based on gradient influence on validation performance without retraining. Implement shift-Gaussian pruning excluding high-scoring samples. Achieve 5× training speedup with 50% subset outperforming full dataset.

Rigorous theoretical framework reformulating DeepSeek's ALF-LB as single-step primal-dual method for assignment problem, proving monotonic Lagrangian improvement, approximate balancing guarantees, and logarithmic expected regret in stochastic settings.

Preserve LLM safety alignment during LoRA fine-tuning via Fisher information regularization and collision-aware geometric constraints.

Current Large Language Models (LLMs) exhibit a critical modal disconnect: they possess vast semantic knowledge but lack the procedural grounding to respect the immutable laws of the physical world. Consequently, while these agents implicitly function as world models, their simulations often suffer from physical hallucinations-generating plans that are logically sound but physically unexecutable. Existing alignment strategies predominantly rely on resource-intensive training or fine-tuning, wh...

Identify and mitigate alignment degradation in self-evolving LLM agents. After deployment, agents systematically abandon training-time safety constraints when environmental feedback rewards rule-breaking. Model two mechanisms: Self-Interested Exploration (individual drift) and Imitative Strategy Diffusion (collective norm erosion), with practical safeguards for post-deployment monitoring.

Train safety-aligned agents using collaborative multi-agent RL where conversation and feedback agents improve together. Trigger: reduce overrefusal while maintaining safety on sensitive queries.

Dynamically modulate reasoning depth at test time using alpha moments and Bernoulli scheduling to optimize inference speed-quality tradeoffs without retraining.

Enable LLMs to solve complex problems through multi-turn agentic reasoning with tool-assisted verification and iterative refinement loops. Trigger: improve reasoning reliability on long-horizon tasks by combining RL with verification.

Use meta-learning to automatically balance Supervised Fine-Tuning and Reinforcement Learning signals, treating SFT and RL as complementary rewards in a unified single-stage training framework.

Implements A^3-Bench from arXiv:2601.09274

Evaluate language models using open-ended answer generation and semantic matching instead of multiple choice, eliminating test-taking shortcuts and achieving near-perfect alignment with human judgment.

Automate sub-agent creation by treating agents as dynamically creatable executors defined by four-tuple abstraction (Instruction, Context, Tools, Model), enabling flexible delegation and cost-aware routing for complex multi-step tasks.

Accelerate diffusion language model inference by dynamically adjusting parallel tokens per step using a small auxiliary autoregressive model, achieving substantial throughput gains.

Enable LLM agents to autonomously retrieve information across multiple granularities using keyword search, semantic search, and chunk read tools. Simple ReAct-based loop with hierarchical interfaces outperforms dense retrieval by allowing adaptive information seeking without complex graph construction.

Route generation dynamically based on relative model advantage for 2× latency reduction in reasoning. Arbitrage learns when draft models excel versus when target models are worthwhile—critical for balancing cost and quality in long reasoning chains.

Reduces inference cost by compressing context into continuous representations using a separate encoder. Generates 4-8x fewer representations than token embeddings while maintaining model performance. Works with any decoder LLM without modification or fine-tuning.

Scale RL training to large models through decoupled rollout and training workers with controlled data staleness.

Calibrate exploration effort in reasoning traces based on problem difficulty by detecting high-entropy windows and applying hierarchical entropy rewards. Reduces unnecessary reasoning on easy tasks while increasing exploration on hard tasks.

Reduce policy gradient variance in language agent training by aggregating rewards in semantic intention space, enabling 9.95% average performance gains across downstream tasks without exponential action space explosion.

Build reusable skill libraries for mathematical reasoning through hierarchical RL. Maintain a high-level skills manager that summarizes successful solution traces and selects relevant strategies to condition future rollouts.

Agentic reward model framework enabling active tool invocation (cropping, retrieval, validation) to ground judgments in verifiable evidence, using multi-stage GRPO with adaptive reward shaping for systematic evidence-based evaluation.

Comprehensive empirical study recommending model-specific test-time scaling strategies (majority voting, first-finish search) across eight LLMs based on architectural family, problem difficulty, and compute budget rather than universal approaches.

Train LLMs to effectively integrate tools through advantage shaping, directly modifying advantage functions to guide policy without compromising training stability.

Recent advances in agentic Large Language Models (LLMs) have positioned them as generalist planners capable of reasoning and acting across diverse tasks. However, existing agent benchmarks largely focus on symbolic or weakly grounded environments, leaving their performance in physics-constrained real-world domains underexplored. We introduce AstroReason-Bench, a comprehensive benchmark for evaluating agentic planning in Space Planning Problems (SPP), a family of high-stakes problems with hete...

Optimize multi-turn agent policies via entropy-guided tree expansion and turn-level credit assignment. AT²PO addresses exploration diversity, sparse credit signal, and policy misalignment problems in LLM agents through structured tree search and turn-aware policy updates.

Adaptive framework for dynamically selecting optimal model-tool combinations in multi-domain reasoning, using cluster-based routing and reinforcement learning for improved agent reasoning across diverse tasks.

Decompose agent reasoning into atomic thoughts guided by curriculum-based reasoning reward models, enabling multi-hop information retrieval and interpretable deep research.

Replace ratio-based clipping in GRPO with KL-divergence constraints using the KL3 estimator, improving exploration and training stability with asymmetric clipping that requires no additional computation.

Demonstrates position bias where LLMs neglect middle content while over-attending to endpoints. Proposes Attention-Driven Reranking (AttnRank) to align content with model's intrinsic attention preferences.

Replace uniform residual accumulation with depth-wise attention that selectively aggregates earlier layer representations. Improve gradient flow and model performance in deep architectures by learning content-dependent depth-wise selection.

Guide LLM exploration in reasoning tasks using attention patterns as navigation signals. This technique branches exploration from high-attention tokens (likely reasoning steps) and applies adaptive sampling to maintain effective gradients, significantly improving training efficiency for mathematical reasoning.

Identify influential texts in long contexts via attention weights using top-K filtering and context subsampling, achieving 10-20x speedup over perturbation methods.

Build fully open audio-language models supporting reasoning over speech, sound, and music with 10-minute long-form comprehension and multi-turn conversation capabilities. Use when you need to process audio modalities alongside text for complex reasoning tasks across speech recognition, sound classification, and music analysis.

Generate realistic video footage of people from audio input using a unified self-attention framework, producing convincing speaker performances without domain-specific restrictions.

Automatically generates diverse multilingual code benchmarks using LLMs, creating 3920 problems across 20 programming languages with quality assurance filtering.

Generate diverse, validated game environments automatically using domain-specific language specifications and LLM coding agents with self-repair, enabling cost-effective (≈$4/env) creation of heterogeneous test domains for evaluating cross-environment agent generalization.

Improves LLM tool-use capabilities through automated environment construction that generates realistic feedback and verifiable rewards for RL-based training without external tools.

Train specialized LLMs to generate optimized Triton GPU kernels using RL with dual rewards for correctness and syntax compliance. 8B model achieves parity with Claude-Sonnet and DeepSeek-R1 by combining supervised fine-tuning on curated code pairs with RL exploration beyond imitation learning ceilings.