Definition
Long-horizon tasks are tasks whose solution requires many sequential decisions, often with delayed feedback — planning a research project, refactoring a large codebase, navigating a multi-day workflow. They expose every weakness of current agents because errors compound.
Episodes covering this
Worth reading next
Papers we haven't done a deep dive on yet, but would recommend on this topic.
- Autonomous Chemical Research with Large Language Models
- Let's Verify Step by Step
- OSWorld: Benchmarking Multimodal Agents for Open-Ended Tasks in Real Computer Environments
- FunSearch: Making new discoveries in mathematics using large language models
- OpenEvolve: Open-Source Implementation of AlphaEvolve
- OpenAI o3 System Card
- World Models
- OpenDevin: An Open Platform for AI Software Developers as Generalist Agents
- AI Scientist: Towards Fully Automated Open-Ended Scientific Discovery
- SWE-agent: Agent-Computer Interfaces Enable Automated Software Engineering
- Chain of Thought Empowers Transformers to be Expressive
- Verified Multi-Step Synthesis using Large Language Models and Monte Carlo Tree Search
- SWE-bench: Can Language Models Resolve Real-World GitHub Issues?
- MLAgentBench: Evaluating Language Agents on Machine Learning Experimentation
- Sagas
- VideoAgent: Long-form Video Understanding with Large Language Model as Agent
- DAgger: A Reduction of Imitation Learning and Structured Prediction to No-Regret Online Learning
- RL²: Fast Reinforcement Learning via Slow Reinforcement Learning
- AMAGO: Scalable In-Context Reinforcement Learning for Adaptive Agents
- GiGPO: Group-in-Group Policy Optimization for LLM Agent Training
- RAGEN: Understanding Self-Evolution in LLM Agents via Multi-Turn Reinforcement Learning
- Reflexion: Language Agents with Verbal Reinforcement Learning
- Voyager: An Open-Ended Embodied Agent with Large Language Models
- Cognitive Architectures for Language Agents
- MemGPT: Towards LLMs as Operating Systems