Reasoning agent: use LLM to reason to act eg. ReAct, AutoGPT
ReAct reasoning is an internal action for agent
short-term memory: context window append-only; limited context and attention; do not persist over new tasks
Reflextion--a new way of learning learn via text feedback learn by updating language(a long-term memory of task knowledge) Voyager: procedural memory (eg. how to build a swore) Generative Agent: episodic memory(eg. all events), semantic memory(eg. Tony like reading)
Cognitive architectures for language agents(CoALA) Memory + Action space + Decision making = Agent
LLM Agent: thinking has no scale limit; generalizable, not task-specific practical and scalable
Applications Webshop(2022) large-scale complex environment automatic reward WebArena(2023) SWE-Bench(2023) ChemCrow
Lessons for research simplicity and generality think in abstraction and familiar with tasks learning history and other subjects
Future work: training, interface, robustness, human, benchmark
FireAct: Training LLM for agents use agent to generate data -->finetune
Agent-computer interface LLM should have their own best interface
Tau-Bench human: finish 1000 times over 1000 times