Car Bench Eval

car-bench-eval

CAR-bench: Evaluating the Consistency and Limit-Awareness of LLM Agents under Real-World Uncertainty — Kirmayr et al. (2026) (arXiv:2601.22027, 2026)

What this evaluates

Evaluates LLM agents' ability to resolve uncertainty and adhere to safety policies in automotive environments. It probes limit-awareness, consistency across multiple attempts, and robust multi-turn tool use under incomplete or ambiguous user requests.

Datasets

• CAR-bench — total 240; splits: test (240); repo https://github.com/CAR-bench/car-bench

Metrics

• Passˆ3 (primary) — range: [0, 1]
  • The fraction of tasks where the model produces a correct solution in all k attempts (k=3), averaged across task types. It measures consistency rather than single-attempt success.

Input / output format

Input: Multi-turn conversational prompts simulating automotive user requests, including ambiguous or incomplete scenarios, requiring tool use and policy adherence.

Output: Sequential tool calls (e.g., get operations, actions) and final responses, evaluated for correctness, policy compliance, and successful task completion.

Scoring recipe

def compute_pass_hat_k(predictions, gold, k=3):
    consistent_passes = 0
    for task_id, attempts in predictions.items():
        if all(check_pass(attempt, gold[task_id]) for attempt in attempts[:k]):
            consistent_passes += 1
    return consistent_passes / len(gold)

Common pitfalls

• Confusing Pass@k (success in at least one of k attempts) with Passˆk (success in all k attempts), which measures consistency.
• Overlooking policy violations or tool execution failures when determining if a task 'passes', as the benchmark explicitly tracks these as distinct error modes.
• Failing to account for user-simulation errors, which can artificially lower Passˆk scores if the simulated user provides incorrect or hallucinated information.

Evidence (verbatim from paper)

We report Passˆ3 averaged across task types as our primary metric, prioritizing consistency in this safety-critical automotive domain. We set k=3 as it effectively discriminates model reliability while avoiding score saturation observed at higher k.

Citation

@misc{kirmayr2026carbench,
  title={CAR-bench: Evaluating the Consistency and Limit-Awareness of LLM Agents under Real-World Uncertainty},
  author={Kirmayr et al. (2026)},
  year={2026},
  note={arXiv:2601.22027}
}

• arXiv: 2601.22027