DAEDALUS: Bootstrapping Agent Memory from Self-Generated Tasks
DAEDALUS bootstraps reusable agent memory from self-generated practice, lifting success by up to 15.9 points.
DAEDALUS bootstraps reusable memory for LLM agents from self-generated practice, without existing tasks or an oracle verifier. An explorer creates challenging but solvable tasks and a solver attempts them; a heuristic from each failure is accepted only after repeated success with that heuristic, then consolidated into a test-time memory bank. Across AppWorld, tau2-bench, and AutomationBench, it improves mean success by up to 15.9 points and pass^5 by up to 2.2x over a no-memory baseline, at lower inference cost than most methods that use training tasks. Ablations find solver traces are the key source of useful heuristics, and the generated tasks can proxy benchmarks for ranking models.
- An explorer and a solver generate and attempt tasks without an oracle verifier.
- Failure heuristics are kept only after the solver repeatedly succeeds.
- Mean success rose by up to 15.9 points and pass^5 by up to 2.2x.
- Gains appear on a small budget and transfer to other model families.
- Generated tasks can proxy benchmarks when ranking models.
From the original article
1 min read at huggingface.co
LLM agents often lack the operational knowledge to act reliably in new environments, as they must discover specific tool behaviors or environment conventions on their own. Without memory of past attempts, they repeat the same mistakes across tasks, leading to more task failures and longer trajectories.
By Antoine Edy, Max Conti, Victor Xing · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.