LetTokenBurnAuditing Escalation Reliability in Healthcare LLM Agents: A Perturbation-Controlled Benchmark of Workflow-State Failures
A 400-episode benchmark measuring healthcare LLM agents' escalation reliability under controlled workflow perturbations (missing, stale, conflicting info) across four task classes.
$ npm i -g lettokenburn-cli$ ltb bench info healthcare-escalation-benchmark
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
| Harness | Source | Confidence | Link | |||
|---|---|---|---|---|---|---|
| qwen3.5-9B | 0.8 | N/A | Independent | medium | Oct 15, 2026 | arxiv.org |
| Gemma 3 4B | 0.7 | N/A | Independent | medium | Oct 15, 2026 | arxiv.org |
| Qwen3 4B | 0.6 | N/A | Independent | medium | Oct 15, 2026 | arxiv.org |
| Llama-3.2 3B | 0.5 | N/A | Independent | medium | Oct 15, 2026 | arxiv.org |
| Qwen3 1.7B | 0.4 | N/A | Independent | medium | Oct 15, 2026 | arxiv.org |
Every number links to where it was published. Low-confidence rows are muted and need per-cell verification.
The benchmark comprises 400 synthetic multi-turn healthcare episodes derived from 40 hand-authored seeds, varied across five conditions (clean, missing, stale, conflict, late-risk). It evaluates agents on a four-way action space: Answer, Ask_Clarification, Abstain, and Escalate.
Scoring uses a cost-sensitive metric with a 5:1 penalty for missed versus unnecessary escalations. Results show high clean accuracy but significant escalation failures under perturbation, with verbalized confidence proving inconsistent as a control signal.