Agent Lightning v1.0: Towards Harnessed Agentic RL
What this work does
Trains agents within their actual runtime harness, explicitly handling prompt reconstruction, token-prefix compatibility, variable sample counts and rollout-level normalization.
Connection to discovery
A direct checklist for converting long scientific-agent traces into correctly aligned training samples.
Evidence
Worked examples show identical text with different token boundaries and how different loss aggregation choices reweight rollouts.
Scope and limitations
Text matching alone is insufficient for merging sampled tokens. These infrastructure fixes do not determine what a scientifically valid reward should be.
THE RESEARCH CONVERSATION
Discussion & reactions
Share a question, an insight, or a replication experience. Sign in with GitHub to join the conversation.