Don't Lose the Thread: Empowering Long-Horizon LLM Agents with Cognitive Resource Self-Allocation
CORAL lets agents checkpoint verified facts and restart from a clean context. An 8B model reaches 42.8% on GAIA with 6-7x fewer input tokens than ReAct.
The claim
Agents that checkpoint facts and reset context stay accurate on long tasks.
The problem
Long-horizon tasks fill working memory with expanding, irrelevant information.
Cluttered context dilutes attention and weakens planning.
How it works
CORAL gives the agent a working memory toolset it can call.
The agent checkpoints task progress and verified facts.
It starts a new episode with a clean context.
Only checkpointed information survives the reset.
Multi-episode policy optimization trains the checkpointing skill.
Figures
What changes
An 8B model reaches 42.8% average accuracy on GAIA.
It uses 6-7x fewer input tokens than a ReAct agent.
On BrowseComp it lifts a strong proprietary agent by 8.0 points.
Limits
The results rest on GAIA and BrowseComp benchmarks.
Checkpointing adds a tool the model must learn to use.
The paper reports no latency figure for the reset cycle.
Source · arXiv