Research

Control policy entropy dynamics in RL by reweighting gradients from clipped tokens. CE-GPPO preserves out-of-clip gradients with beta parameters to stabilize exploration-exploitation balance, preventing entropy collapse while maintaining training stability in LLM fine-tuning.

Implement techniques from Can LLMs Clean Up Your Mess? A Survey of Application-Ready Data Preparation with LLMs. Data preparation aims to denoise raw datasets, uncover cross-dataset relationships, and extract valuable insights from them, which is essential for a wide range of data-centric applications

Adapt large reasoning models for optimization tasks using expert-guided hint correction. Generate high-quality training data with minimal expert intervention (<2.6% token modification). Trigger: fine-tune reasoning models on domain-specific tasks without large supervised datasets.

Align diffusion models to hierarchical fine-grained criteria rather than binary preferences. Decompose expert knowledge into attribute hierarchies and apply Complex Preference Optimization to simultaneously maximize positive attributes while minimizing negative ones.

We introduce Being-H0.5, a foundational Vision-Language-Action (VLA) model designed for robust cross-embodiment generalization across diverse robotic platforms. While existing VLAs often struggle with morphological heterogeneity and data scarcity, we propose a human-centric learning paradigm that treats human interaction traces as a universal 'mother tongue' for physical interaction. To support this, we present UniHand-2.0, the largest embodied pre-training recipe to date, comprising over 35,...

Reinforcement learning (RL) is central to post-training, particularly for agentic models that require specialized reasoning behaviors. In this setting, model merging offers a practical mechanism for integrating multiple RL-trained agents from different tasks into a single generalist model. However, existing merging methods are designed for supervised fine-tuning (SFT), and they are suboptimal to preserve task-specific capabilities on RL-trained agentic models. The root is a task-vector mismat...

Build autonomous research agents using pre-computed knowledge graphs instead of online reasoning. Extract methodological patterns from literature, organize them into structured knowledge, and enable agents to align user research intents with established paradigms for efficient, grounded research planning and execution.

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...

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.

Enable research agents to interleave evidence-based drafting with reasoning-driven deepening, automatically expanding outlines based on discovered gaps, using trajectory pruning for efficient RL training.

Large Language Models (LLMs) based autonomous agents demonstrate multifaceted capabilities to contribute substantially to economic production. However, existing benchmarks remain focused on single agentic capability, failing to capture long-horizon real-world scenarios. Moreover, the reliance on human-in-the-loop feedback for realistic tasks creates a scalability bottleneck, hindering automated rollout collection and evaluation. To bridge this gap, we introduce AgencyBench, a comprehensive be...

Train a single LLM to decompose complex queries into subquestions and integrate retrieved contexts through two-stage supervised and preference-based reinforcement fine-tuning, achieving 7.6% average improvement and matching 685B models with 32B parameters.

The evolution of Large Language Models (LLMs) into autonomous agents has expanded the scope of AI coding from localized code generation to complex, repository-level, and execution-driven problem solving. However, current benchmarks predominantly evaluate code logic in static contexts, neglecting the dynamic, full-process requirements of real-world engineering, particularly in backend development which demands rigorous environment configuration and service deployment. To address this gap, we i...

Convert arXiv papers into ready-to-use agent skills using category-aware extraction. First classifies the paper into one or more of 11 research categories, then applies a specialized extraction pipeline for each category — because different types of papers produce different types of usable knowledge. A single paper can yield multiple skills if it spans categories. Use this skill whenever the user wants to turn a paper into a skill, extract practical techniques from research, build a skill lib...

Convert survey and synthesis papers into field navigation guides. Extracts taxonomies, method selection decision trees, literature navigation heuristics, and open problems. Use this skill when extracting skills from Category 10 (Survey and Synthesis) papers — comprehensive reviews, position papers, tutorials, or roadmaps that organize a research landscape.

Convert papers that disprove conventional wisdom into paradigm-challenge skills. Extracts the prior belief, the falsifying experiment, and the revised principle. Use this skill when extracting skills from Category 3 (Paradigm Challenge) papers — papers that say 'rethinking', 'revisiting', or 'do we really need X', where the core move is adversarial (proving the community wrong).

Convert mechanistic analysis papers into transferable analytical methodology skills. Extracts the research question, analytical instrument, controlled confounds, and practitioner implications. Use this skill when extracting skills from Category 9 (Mechanistic Analysis) papers — Network Dissection-style interpretability work or any paper whose goal is exploratory understanding of why something works.

Convert foundational papers that create new subfields into conceptual framework skills. Extracts problem definitions, vocabulary, founding experiments, and opened research directions. Use this skill when extracting skills from Category 8 (Field Foundation) papers — MAML-style paradigm-creating papers or 'Deep Learning' review-style papers that define entire research communities.

Convert dataset and benchmark papers into evaluation infrastructure skills. For datasets: extracts collection protocol, annotation design, quality control. For benchmarks: extracts task definition, metric selection, leaderboard design. Use this skill when extracting skills from Category 2 (Evaluation Infrastructure) papers — ImageNet-style dataset papers, SWE-bench-style benchmark papers, or any paper whose primary contribution is evaluation methodology.

Categorize ML/AI research papers into 11 types based on their title and abstract. Returns structured JSON with a primary category, optional secondary categories, extractability rating, and rationale. Designed for the SkillXiv paper2skill pipeline as a triage step. Use this skill whenever the user wants to classify, categorize, sort, or triage research papers — whether a single paper or a batch. Also trigger when someone asks "what kind of paper is this?", wants to filter papers by type, or ne...

[Research] Use when a workflow step or the user asks for web research. Gathers and triages candidate sources; `--chain=deep-dive` also deep-dives them into an evidence base.

[Process] Use when a developer needs to understand and judge scoped work: a change set, subsystem, decision, plan or concept — flow, rationale, trade-offs, testing. Bugs: investigate --mode=debug; one feature''s code flow: investigate.

[Documentation] Use when generating the DERIVED technical spec view over code and tests, or reporting §8 TC drift. Generator only, never authors business content. generate|audit|sync.

[Research] Use when a workflow step or the user asks for a research synthesis: findings into a structured report.