A language-guided decision layer for autonomous driving showed a narrow result in the study’s main test. In the rural-road replay, the configuration with intent gating had repaired plans that stayed inside a defined corridor in every deterministic run under v1, v2a and v2b, while the baseline exited the corridor. The document is an arXiv version-1 preprint dated 26 August 2026.
The layer reads structured descriptors and calculates a smoothed intent-geometry divergence score, a signal of disagreement between interaction intent and the vehicle’s geometry. It uses that score to gate the planned maneuver before commitment, upstream of a separate corridor envelope. In the main replay, the envelope-only configuration flagged the excursion but did not modify the plan.
Milliseconds before the corridor exit
The timing comparison used drift onset and corridor exit for the gate, and t0 for the envelope. Under v1, the gate fired 72.2 ± 2.0 milliseconds after drift onset and 161.1 ± 2.0 milliseconds before corridor exit. Under v2b, those figures were 67.7 ± 3.7 milliseconds and 165.6 ± 3.7 milliseconds. The ± values describe wall-clock jitter across deterministic runs. The envelope-only configuration flagged the excursion 100.0 ± 0.0 milliseconds after t0, but it acted after commitment without modifying the plan.
The main test used a 45.75-second monocular rural-road video recorded at 30 frames per second. Each configuration was replayed 10 times. Drift onset was set at frame 983, and corridor exit at cycle 330. The paper reports two experimental rounds against one frozen, disclosed implementation.
Nominal triggers fell, but one criterion was missed
Version 2 treated uncertain outputs as abstention: they contributed no disagreement. It also added hysteresis, the gap between the threshold that turns a response on and the one that turns it off. The off threshold was set 0.10 below the on threshold, while the (1 - c_t) term was left unchanged.
On the identical nominal set of 703 cycles over 5.86 minutes, v1 generated nine triggers, or 1.537 per minute. V2a and v2b each generated two, or 0.341 per minute. All nine v1 triggers shared the same pattern: uncertainty had been scored as half a conflict.
The lower trigger count did not make every preregistered test pass. For v2b, the false-trigger and outcome criteria passed, but the all-case pre-onset criterion failed: only three of five cases were flagged before drift onset. In a descriptive merge-window check, Red flags fell from six of 10 under v1 to four of 10 under v2b.
The failed criterion was not redefined after the result. In clip 000228, the flag came 1.1 seconds after onset; the main-case v2b flag came 67.7 ± 3.7 milliseconds after onset. A different mixed-anchor formulation would have passed four of five cases, but the reported analysis retained the original rule.
The broader comparisons were mixed
Four additional Car Crash Dataset clips passed screening. Clips 000754 and 000871 were typed as intent-level failures, while 000676 and 000228 were boundary checks. Under v1, the uncorrected planner exited the corridor in all four clips, with a median detection lead of 750 milliseconds. Under v2b, the 000676 boundary signal remained at an 800-millisecond lead, and every intent-level failure had a lead equal to or better than v1.
The clips were five-second pre-crash cuts. They could show detection in short replay windows, but they could not establish whether a longer closed-loop system could recover.
After correction of the replay comparison, the full score was the fastest detector on four of five failures under the deployed eligibility rule and on three of five under the unvetoed rule. Geometry-only fired on all five under deployed eligibility, although it was later in selected cases. In the d-only comparison, two of five detections were absent; that comparison omitted the (1 - c_t) term. Under pure model-free eligibility, geometry-only produced four nominal false triggers, compared with two for the full score.
An earlier ablation run silently ignored the variant switch and ran the full score three times, producing identical firing cycles. A one-line fix and an independent cycle-by-cycle recomputation were reported in the artifact bundle.
On 394 positive and 1,863 negative nuScenes interaction tracks, the two scores were compared at a matched false-positive rate of 0.093. At that rate, the full score detected 23.4% of conflicts and anticipated 21 tracks, with a median lead of 1.5 seconds. Geometry-only detected 80.0% and anticipated 19 tracks, with a 2.5-second median lead.
Highway probe showed weak transfer
The implementation behind the tests was rebuilt in August 2026 after the original was lost. It used Qwen2.5-0.5B with LoRA on a single RTX 5090 laptop GPU and was fine-tuned on 55,433 interaction windows labeled by deterministic rules, without human annotation.
On an out-of-domain comma2k19 highway probe covering eight segments, 4,792 cycles and eight minutes, v1 produced 25.4 false triggers per minute and v2b produced 7.5. Both exceeded the study’s usability bound because of out-of-domain perception failures.
Taken together, the evidence points to a narrow engineering result: the full score was fastest on most of the tested failure comparisons, while nominal model-free trigger counts were lower with the uncertainty veto. At a matched false-positive rate, geometry-only scoring detected more nuScenes conflicts. The evidence consists of deterministic replays and rule-labelled benchmark data, with v2b still missing one preregistered criterion. It is not a production validation or a demonstrated reduction in real-world crash risk.
Paper data and sources
Original title: Gating Before Commitment: Anticipating Intent Divergence to Prevent Post-Interaction Decision Failures in Autonomous Driving
Authors: Cong Xu, Ravi Sankar
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-26
DOI: Not available
Original paper · Full text