MADS-Bench
MADS-Bench: Multi-Agent Dialog Simulation Benchmark
A synthetic multi-turn function-calling benchmark with 1095 conversations across 6 domains and 55 personas, focusing on adversarial and failure-prone scenarios.
How to run
$ npm i -g lettokenburn-cli$ ltb bench info mads-bench
External harness. No Harbor adapter has been generated for this benchmark yet. `ltb bench info` prints the known links.
Runs happen on your machine with your own subscription; only the score summary is pushed. CLI docs
Scores
2 sourced entries| Harness | Source | Confidence | Link | |||
|---|---|---|---|---|---|---|
| Qwen2.5 32B Instruct | 54.0 | N/A | Independent | medium | Oct 8, 2026 | arxiv.org |
| Qwen2.5-7B-Instruct | 39.8 | N/A | Independent | medium | Oct 8, 2026 | arxiv.org |
Every number links to where it was published. Low-confidence rows are muted and need per-cell verification.
Reproduced by the community
All runsAbout
MADS-Bench comprises 1095 synthetic dialogues generated via a multi-agent pipeline, spanning six domains, 72 tools, and 55 personas. It specifically targets non-cooperative user behaviors, including unhappy paths and impossible scenarios.
The benchmark evaluates agent robustness using a two-stage training approach (SFT and RLVR). It measures performance in handling conversational and system execution adversities, with approximately 66% of conversations designed to be failure-prone.