Claude Skills by ADu2021

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

Reduce reasoning model inference cost by selectively applying latent iterations—use a lightweight neural decider to identify hard tokens and refine only those with focused LoRA modules, achieving 94% token exemption with 8-11% accuracy gain.

Dynamically allocate reasoning budgets per-task: 36.8% length reduction while improving accuracy 8.4% by learning which problems warrant deep reasoning vs. quick answers. Use when optimizing reasoning efficiency across variable-difficulty problems.

Route reasoning between discrete token space (when uncertain) and latent soft embeddings (when confident). Use maximum next-token probability as a routing threshold to dynamically select the reasoning space, improving accuracy under latent reasoning while reducing computational cost through selective discrete sampling.

Control LLM reasoning effort through discrete modes (High/Medium/Low) using budget-aware supervised fine-tuning and adaptive reward shaping, enabling compression-performance tradeoffs.

Improve agent reasoning for spatial tasks by augmenting LLMs with map tools and parallel test-time exploration. Framework uses reinforcement learning to train agents to iteratively refine hypotheses using map feedback. Parallel exploration enables agents to test multiple candidate locations before committing to answers. Introduces MAPBench benchmark for evaluating spatial reasoning in image geolocalization tasks.

Demonstrates that chain-of-thought reasoning improves LLM factual retrieval through computational buffering and self-priming. Improves single-hop factual accuracy by enabling models to generate contextual bridge facts before recalling answers.

Leverage video generation models as unified multimodal reasoning engines that bridge text and vision by embedding reasoning as dynamic visual processes, enabling models to tackle both spatial puzzles and mathematical problems within a single temporal framework.

Train unified models to generate interleaved reasoning steps combining text and image thoughts as complementary modalities, enabling adaptive behavior like autonomous mode-switching and superior test-time scaling for vision-centric tasks without requiring external guidance.

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

Align reasoning models to be safe without external supervision by prepending refusal instructions to harmful queries and training on self-generated safe responses. Unlocks latent safety knowledge within the model's native distribution.

Generates high-quality audio through three-stage CoT reasoning pipeline: foundational foley synthesis, object-focused refinement, and instruction-guided editing. Uses fine-tuned VideoLLaMA for reasoning and flow-matching audio foundation model. Apply for professional audio design workflows or video-to-audio applications.

Train LLMs for tool-integrated mathematical reasoning via hierarchical RL combining episode-level problem correctness with step-level code execution quality. Addresses sparse rewards in reasoning chains through TIRGen data construction and self-correcting inference with dynamic backtracking.

Enable parallel reasoning threads on standard autoregressive inference engines without modifications to position embeddings or KV cache. ThreadWeaver achieves 1.53× speedup while maintaining chain-of-thought quality—ideal when you need faster reasoning without special hardware.

Combine diffusion-based parallel drafting with autoregressive sampling in a single forward pass using structured attention masks—achieving 5x+ token throughput while maintaining autoregressive-level output quality through hybrid generation.

Guarantee LLM inference completes within time budgets via response length prediction, execution time estimation, and adaptive KV cache eviction. Three-stage pipeline predicts response length, estimates end-to-end time with 1.22% accuracy, adjusts cache eviction ratio—enabling time-critical deployment in robotics and autonomous systems.

Automate univariate time series forecasting through a four-agent system orchestrating preprocessing, model selection, validation, and reporting. Use when reducing manual effort in forecasting pipelines and improving reproducibility.

Evaluate multimodal models on agentic visual reasoning through 13 diverse tasks requiring novel tool use for image processing and manipulation in chain-of-thought, revealing that strong performance requires genuine thinking-with-images capabilities.

Compress reasoning models by dynamically re-weighting short-CoT (System-1) and long-CoT (System-2) training data, achieving 40% token reduction while maintaining accuracy.

Autonomously synthesize task-specific agent planning architectures by optimizing for correctness, stability, and efficiency via impedance-guided preference optimization.

Learn to compress entire scenes into a single bottleneck token that captures temporal dynamics. Enables efficient visual tracking and robotic manipulation by forcing reconstruction from minimal target hints, achieving superior performance with training costs comparable to standard autoencoders.

Improve LM performance with token order prediction (TOP) auxiliary loss using learning-to-rank instead of exact multi-token prediction, achieving gains across math, code, and NLP tasks

Dynamically select important tokens at the attention head level, performing dense attention only on selected tokens and scattering results back. Achieves 3.23x attention speedup at 128K context with 1% accuracy loss through layer-wise representation stability analysis.

Build autoregressive image generators using post-training quantization that bridges continuous VAE tokens with discrete vocabulary modeling. Achieves state-of-the-art visual quality via dimension-wise token prediction without training instability.

The Confidence Dichotomy: Analyzing and Mitigating Miscalibration in Tool-Use Agents. From arXiv:2601.07264

Implements ET-Agent from arXiv:2601.06860

Two-stage post-training framework combining rejection-sampled SFT and RL for LLM-guided repository code search and issue localization.

Prove that tool-augmented learning unboundedly scales factual knowledge recall compared to parameter-constrained memorization, enabling efficient knowledge retrieval via external tools

8B parameter orchestrator trained with end-to-end RL balancing outcome, efficiency, and user preference rewards to strategically coordinate diverse tools and models. Generate realistic tool-use data via ToolScale synthetic pipeline for verifiable multi-turn scenarios.

Tool-R0 trains tool-use agents through self-play between Generator and Solver, creating curriculum-like progression without manual task curation.

Enable non-English language models to reliably generate function calls and tool invocations through continued training on bilingual function-calling datasets. Achieves 28% improvement in parsing accuracy while preserving language understanding.

Tool Verification stabilizes self-improving reasoning models by using external tool execution as ground-truth evidence to prevent spurious consensus from becoming reinforced training signals.

Reward-guided search methods have demonstrated strong potential in enhancing tool-using agents by effectively guiding sampling and exploration over complex action spaces. As a core design, those search methods utilize process reward models (PRMs) to provide step-level rewards, enabling more fine-grained monitoring. However, there is a lack of systematic and reliable evaluation benchmarks for PRMs in tool-using settings. In this paper, we introduce ToolPRMBench, a large-scale benchmark specifi...

Specialized outcome reward models (1.7B-14B) for evaluating tool-calling performance in LLMs. Addresses the critical gap in reward modeling where general-purpose models miss key signals of effective tool use. Enables better Best-of-N sampling, data filtering, and RL-based policy training through FC-RewardBench evaluation framework.

Detects and prevents unsafe tool invocations in LLM agents through proactive step-level guardrails, reducing harmful tool calls by 65% while improving task success rates by 10% against prompt injection attacks.

Generate 1.5M realistic agent training trajectories from 495 real-world MCP servers without human annotation. Use to build tool-agentic training datasets with authentic tool execution, multi-turn conversations, and error handling.

Synthesize large-scale training data for LLM agents by generating diverse tool-use trajectories from real Model Context Protocol (MCP) environments. A 1.5M-example dataset created via multi-stage pipeline: query generation, quality filtering, trajectory creation with real tool execution, validation, and diversification.

Implements TourPlanner from arXiv:2601.04698

Recent years have witnessed increasing interest in extending large language models into agentic systems. While the effectiveness of agents has continued to improve, efficiency, which is crucial for real-world deployment, has often been overlooked. This paper therefore investigates efficiency from three core components of agents: memory, tool learning, and planning, considering costs such as latency, tokens, steps, etc. Aimed at conducting comprehensive research addressing the efficiency of th...

Linguistic expressions of emotions such as depression, anxiety, and trauma-related states are pervasive in clinical notes, counseling dialogues, and online mental health communities, and accurate recognition of these emotions is essential for clinical triage, risk assessment, and timely intervention. Although large language models (LLMs) have demonstrated strong generalization ability in emotion analysis tasks, their diagnostic reliability in high-stakes, context-intensive medical settings re...

Implement techniques from Towards Pixel-Level VLM Perception via Simple Points Prediction. We present SimpleSeg, a strikingly simple yet highly effective approach to endow Multimodal Large Language Models (MLLMs) with native pixel-level perception

Evaluate LLM agent capabilities using tower defense game environment with multimodal observations (pixel, text, structured state). Benchmark reveals critical agent limitations: inadequate planning validation, inflexible decision-making, and inefficient action use. Demonstrates significant performance gap between current LLMs and human experts, providing structured framework for measuring agent planning, adaptation, and hallucination tendencies.

Build multilingual LLMs that excel at machine translation while preserving general-purpose capabilities through a staged training pipeline combining continued pretraining, SFT, preference optimization, and RL with verifiable rewards.

Trace unsafe LLM outputs back to conflicting beliefs in training data using suffix-array matching and Belief Conflict Index.

Improve LLM performance at test time through in-context learning and experience libraries, eliminating the need for parameter updates while maintaining competitive results.

Improve LLM outputs without parameter updates using learned token priors that guide inference. Trigger: optimize agent behavior in deployment without model retraining or fine-tuning.

Infer complete object structure despite occlusion using multi-camera video. Enables training models to predict hidden object appearance by combining temporal and spatial context from multiple viewpoints.

Select best reasoning trajectories from multiple samples using step-level scoring from a 0.6B lightweight verifier that exploits hidden states, outperforming external reward models by 4-12% without massive annotations.

Bridge labeled and unlabeled data through trajectory similarity in reinforcement learning. Select reliable unlabeled samples by comparing pass-rate evolution trajectories against labeled data. Achieve 42.6% accuracy with 1K labeled + 3K unlabeled samples, surpassing fully-supervised training on 45K labels with 10% annotation budget.

Train LLM agents via tree-search rollout sampling and step-wise advantage estimation. Achieve 1.5x more rollouts within fixed token budgets and implicit step-level preference learning through dual-level advantage computation on tree structures.