CrossEnv synthesizes verifiable cross-environment agent tasks
Researchers introduced CrossEnv, an automated framework that uses entity-tool graphs to synthesize executable, verifiable workflows spanning multiple environments and services. Paired with a five-level progressive curriculum, it enables LLM agents to master complex, cross-domain dependencies with significantly fewer training examples.
Single-environment synthetic data has hit diminishing returns for real-world agent reliability where tasks inevitably span fragmented APIs and stateful services. By formalizing cross-service data dependencies through entity graphs, CrossEnv solves the hardest part of agent training: producing ground-truth verifiability across messy tool boundaries.
- –Entity-tool graphs bridge isolated service schemas via shared entities, ensuring state changes in one environment stay logically consistent in downstream tools.
- –A five-tier curriculum gradually escalates from isolated tool execution to multi-domain orchestration and interactive user dialogue, proving far more sample-efficient than brute-force task generation.
- –The framework demonstrates strong zero-shot generalization across four external agent benchmarks, transferring successfully to novel environment combinations and longer workflows.
- –Programmatic synthesis eliminates human annotation bottlenecks while avoiding costly, fragile interactions with live production environments during RL training.
DISCOVERED
3h ago
2026-11-19
PUBLISHED
4h ago
2026-11-19
RELEVANCE
AUTHOR
Discover AI