Summary
Meituan and Nanjing University build interactive tool-use environments and verifiable tasks entirely from scratch — from a domain name alone, not tool documentation. An LLM writes tool and database schemas; a code agent implements them while a test agent writes unit tests and matching database instances, and a debug agent iterates until every “procedural test” passes. A tool-dependency agent links verified tools into a graph. Tasks start from an executable seed tool chain, an initial state built to support it and padded with schema-valid distractor records whose density scales with task complexity, and are then expanded into non-linear subgraphs while an oracle agent checks they stay solvable. Reward is a check of the final database state, not an LLM judge. Training Qwen3 with GRPO on 16 synthetic domains improves zero-shot results on τ²-Bench and VitaBench, and performance rises monotonically as training domains grow from 2 to 16 at a fixed task count.
Key points
- From scratch, not from docs. The pipeline starts from a domain keyword (e.g. “Job Seeking”); the authors argue document-based synthesis like AutoForge has limited scalability and that environments must be “grounded in verified, executable code rather than probabilistic text generation.”
- Procedural testing. A Code Agent implements each tool while a Test Agent synthesizes unit tests and matched database instances; any unexpected failure or state mismatch sends a Debug Agent back to fix the tool or the instance until tests pass. Database code is generated and tested first, as the prerequisite for tools.
- Distractor injection. The initial state is populated with extra records that satisfy every schema constraint but are “functionally orthogonal to the ground-truth trajectory,” with density scaled to the target task complexity.
- Graph expansion with an oracle gate. Seed chains are expanded into larger subgraphs; a Qwen3-235B oracle with best-of-16 search estimates whether the expanded task is still solvable, balancing complexity against solvability.
- Rule-based reward. The final database state is compared field by field (dynamically generated IDs and optional columns exempt), avoiding the cost and hackability of LLM-as-judge rewards.
- Scale. 16 domains, each about 50 tools and 5–20 database tables; Qwen2.5-72B as user simulator; GRPO for 48 steps.
- Results. Qwen3-SE-8B improves on every reported τ²-Bench and VitaBench domain over Qwen3-8B; Qwen3-SE-32B roughly doubles VitaBench cross-domain, 5.3 → 10.8. Evaluation is strictly out of distribution — test domains are excluded from training.
- Domain scaling curve. At a fixed 1,024 tasks, training on 2, 4, 8 and 16 domains improves Pass@4 monotonically on both benchmarks and has not plateaued at 16.
- Executability verification matters. Removing execution-based verification degrades performance across all τ²-Bench domains: the data then contains tool calls that look plausible but fail at runtime (e.g. refunding a non-existent order), which gives conflicting reward signal.
Relevance
- WorldSmith — a construct pipeline with built-in testing agents (code, test, debug) and an oracle solvability gate; the closest published analogue to an agent team building verified environments. Its domain-scaling curve is direct evidence for diversity over volume.
- Environment Generation — distractor injection as a cheap difficulty lever, and a fixed-budget domain-scaling experiment.
- Tau2-Bench and VitaBench — the out-of-distribution targets.
Notes
From paper club, 7 April 2026: ScaleEnv takes the AutoForge approach further by creating a Distractor that injects the DB with data that’s close, but not quite the right answer. If the task is to “book the cheapest flight from SFO to LAX”, the Distractor inserts multiple flights from SFO to LAX on the same dates with similar prices. This ensures the agent actually reasons with logic before moving on to the next tool call.