Papers

Retrieval and Memory

• Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks - foundational RAG architecture for combining parametric models with retrieved knowledge.

Tool Use and Agent Behavior

• ReAct: Synergizing Reasoning and Acting in Language Models - combines reasoning traces with action selection.
• Toolformer: Language Models Can Teach Themselves to Use Tools - explores self-supervised tool-use learning.
• MRKL Systems: A Modular, Neuro-Symbolic Architecture - early modular architecture for routing between models and tools.
• Reflexion: Language Agents with Verbal Reinforcement Learning - agent improvement through self-reflection and feedback.

Benchmarks and Environments

• SWE-bench - benchmark for software engineering tasks from real GitHub issues.
• WebArena - benchmark environment for autonomous web agents.