research
The Agent Checkpoints the Facts and Clears the Clutter
CORAL used six to seven times fewer input tokens than ReAct and reached 42.8 percent average GAIA accuracy with an 8-billion-parameter model.
Summary
CORAL used six to seven times fewer input tokens than ReAct and reached 42.8 percent average GAIA accuracy with an 8-billion-parameter model.
The toolset lets an agent preserve verified facts, reset an overloaded context and resume in a clean episode. Reported gains extend to multi-hop QA and BrowseComp, where a proprietary agent improved by eight points.
Why it matters
CORAL used six to seven times fewer input tokens than ReAct and reached 42.8 percent average GAIA accuracy with an 8-billion-parameter model.
Limits and context
No additional limitation was separately recorded.
Key claims
CORAL used six to seven times fewer input tokens than ReAct and reached 42.8 percent average GAIA accuracy with an 8-billion-parameter model.
Evidence: source-2026-11-01-018
Sources
- arXiv preprint 2610.40350, original v1arXiv · primary research
Corrections
No corrections have been recorded for this story.