TRIAL: Trajectory-Relative Hindsight Distillation
What it is
TRIAL is a research implementation for agentic reinforcement learning that uses trajectory-relative hindsight distillation to improve LLM agents. It provides scripts to train and evaluate models on ALFWorld and WebShop benchmarks using Qwen2.5-3B-Instruct.
The problem it solves
Standard reinforcement learning for LLM agents often suffers from sparse rewards and inefficient sample usage; TRIAL addresses this by distilling from hindsight trajectories.
What you could use it for
- Training LLM agents on ALFWorld household tasks
- Training LLM agents on WebShop e-commerce tasks
- Researching hindsight distillation methods for reinforcement learning
Read the source ↗Back to the library →
Summary written by an AI model from the source; check the source before you rely on it.