Reaches 51.7% on the competition-level MATH benchmark with a 7B model and no tools or voting, rivaling Gemini-Ultra and GPT-4. Built on a 120B-token math corpus mined from Common Crawl, and introduces GRPO, a memory-efficient PPO variant for reasoning.

Explains when language models benefit from external programs, comparing tool selection, use, and creation. Evaluates performance gains against compute costs and identifies gaps in tool-use benchmarks.

Tests whether agents can finish 369 tasks in real desktop applications, using executable checks of the final state. The original study found a wide gap between agent and human success rates.

Extracts reusable workflows from past task trajectories to guide future web actions. Supports offline demonstrations and online experience, improving long-horizon navigation without manually written routines.

Enforces output grammars during LLM decoding while reducing CPU validation overhead. Precomputed token masks and persistent stacks support nested formats without checking the full vocabulary at every step.

Separates web-agent planning from action execution and trains the planner on plans inferred from successful trajectories. Reports 57.58% success on WebArena-Lite and 81.36% on text-only WebVoyager.

Explains how base-model behavior and GRPO normalization distort claims about reasoning gains. Introduces Dr. GRPO to reduce length bias, with a 7B training recipe reaching 43.3% on AIME 2024.

Turns successful web-agent trajectories into executable, verified Python skills that expand the action space. On WebArena, success reaches 40.4%, versus 32.7% for a static agent and 36.3% for textual skills.

Tests whether browsing agents can uncover obscure facts across the web through 1,266 short-answer questions. Answers are easy to check even when finding them requires sustained, creative search.

Trains web agents to gather information through longer interaction, using an online RL curriculum that expands rollout horizons. Tests on WebVoyager and WebArena show how exploration can complement longer reasoning before each action.

Tests research agents on 100 expert-written tasks across 22 fields. Separates report quality from citation reliability, using adaptive reference-based grading and checks that cited pages support individual claims.

Turns human desktop demonstrations into training data for agents that reason over screenshots and control mouse and keyboard. Releases a cross-OS collection tool, AgentNet data, and models; the paper reports 45.0% on OSWorld-Verified for its 72B model.