- Rating
- 2
- Confidence
- 4
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 1
Read review
Paper summary
The paper presents DrivingBench, a harness that allows vision-language models to drive a real Toyota Corolla using three MCP tools: observe, set_motion, and stop_now. Four model–harness pairs are evaluated on a 127 m cone course, with up to three attempts each. Only GPT-6 Astra completes the course, succeeding on its second attempt in 5:22, while most attempts fail within the first 15 m. The paper identifies several key challenges, including spatial misinterpretation, premature decisions, control limitations, inference latency, and failure to apply lessons from reflection.
Critical strengths
- What makes this work valuable to me is that it studies closed-loop control of a full-size vehicle in a world that keeps moving during deliberation. This setting exposes behaviors that static or turn-based benchmarks cannot: stop-and-go strategies lose steering authority because the EPS can hardly turn the wheel from a standstill (median time to 90%: 4.0 s, versus 1.9 s while rolling; Fig. 8b); models act on frames captured roughly 4 m earlier (Fig. 7b); and accurate self-diagnosis often fails to translate into changed control (§5.4).
- The harness, prompts, per-attempt telemetry, tool calls with the model’s stated reasons, and videos have been released. The authors also openly acknowledge several limitations: the model–harness confound (§4.2), n = 1 (§6), and the inferred centerline (Fig. 3). As a result, the case study is auditable, even though it cannot be reproduced without the specific hardware and course.
Critical weaknesses
-
The empirical claims rest on one trial per model, and the three attempts within a trial share context. Under those conditions the ranking and the in-context-learning claim cannot be separated from run-to-run variance. The failures are also confounded with the actuation stack: from rest, a 1 m/s command produces measured peaks of 1.9–2.8 m/s (Figs. 12–13, Table 3) and the wheel barely turns until the car is rolling (Fig. 8b), yet 8 of 11 failures occur in section A right after a standing start and §5.3 attributes them to perception. Without repeated trials and an analysis that separates the car's response from the model's decisions, the paper's conclusions are not reliable.
-
The work is a case study rather than a benchmark, and lacks the baselines needed to interpret it. There is one course, one vehicle, one trial per system and no baselines: no human teleoperating through the same MCP interface, no scripted or classical controller. Model and harness are confounded (§4.2), and the harnesses differ in system prompt and latency, which §5.2 shows is decisive.
Questions for authors
- How many development or pilot runs were conducted, with which models, and on the same course? Was the final prompt tuned mainly against GPT-6 Astra? Please report the refusal rate of each model before and after the
drivingbench_sandboxrenaming and the other prompt changes. - Quantify per-start launch speed overshoot (commanded vs. measured) and, for the eight Section A failures, determine whether perfect tracking of commanded speed and steering would have kept the car on course—i.e., how much of each failure was due to the model’s decisions versus the actuation stack.
- Have you run any additional independent trials since submission? Even 3–5 per model would materially change how much confidence the ranking and in-context-learning claims can support.
- What was the performance of a human operator driving through the identical MCP interface (same tools, same latency, same images)? This would establish what fraction of the gap to human driving is due to the interface rather than the models.
- Why was the turning reference in the prompt (~60 s for 90°) not corrected after it was found to disagree with telemetry, and were all trials run under the same lighting and weather conditions and on the same day?