DEFT-RLVR trains driving VLMs to reason without seeing the ground-truth future trajectory first, then verifies against candidate trajectories after the decision is committed, cutting severe hallucination in teacher rationales from 50% to 29% in the annotation study that motivates it.
You’ve shipped an agent that reasons over dashcam video and picks a driving maneuver. To train it, you feed a big teacher model the scene and the logged future trajectory, then ask it to explain why that trajectory is correct. The explanations look great. But the paper shows the teacher is doing something closer to reverse-engineering a known answer than actually reasoning from the scene. When the scene is causally hard (a hard brake, a sharp turn), the teacher invents evidence, such as a mandatory-turn sign that isn’t there. Any student model trained on those rationales inherits fabricated causal chains. The dominant baseline in this space is Chain-of-Thought-supervised vision-language-action training conditioned on the logged trajectory. This paper argues that conditioning is the bug.
The fix has two pieces. First, AD-MCQ reframes trajectory planning as picking one of six candidate future trajectories rather than generating waypoints from scratch. Candidates come from a K-means clustering codebook of 8192 trajectory prototypes clustered from ~489k logged futures; distractors include scale-matched and constant-velocity trajectories to block shortcut heuristics. This gives a deterministic verifier: did the model pick the option matching the logged future?
Second, DEFT (Deferred Exposure of Future Trajectories) splits inference into two turns. Turn 1 shows only the scene, ego state, and navigation intent, and asks the model to commit to a high-level decision. Turn 2 then reveals the six candidates and asks the model to ground its already-committed decision in one of them. Both turns are optimized jointly as a single rollout using Group Relative Policy Optimization (GRPO), with two reward signals multiplied: an exact-match reward on the candidate choice, and a rubric reward that only fires when the answer is correct. The rubric itself is generated once offline per scene by a separate VLM, then graded online by a text-only judge that never sees the images, the candidates, or the ground truth. This keeps grading fast and stops the judge from re-introducing the same anchoring bias.
# per training example
rubric = offline_vlm_rubric(scene) # 6-10 weighted criteria, done once
for rollout in group_of_16:
plan = policy(scene, ego, nav) # turn 1: no candidates visible
choice = policy(plan, candidates) # turn 2: candidates revealed
r_mcq = int(choice == logged_future_option)
r_rub = text_judge(normalize(plan), rubric) if r_mcq else 0
reward = r_mcq * r_rub
grpo_update(rollouts, rewards)
The prevailing recipe for Chain-of-Thought supervision in driving Vision-Language-Action model models is to show the teacher the logged future trajectory so it can explain why that trajectory is right. This paper shows the opposite. Handing the teacher the answer produces confident post-hoc rationalization, not causal reasoning. Trajectories should be verification targets after a decision, never premises before one. The load-bearing evidence is the controlled human study on 100 hard-causal scenes where the only difference between conditions is whether the teacher sees the logged future, not the headline benchmark numbers.
On the controlled annotation study, exposing the ground-truth trajectory to the teacher raises severe hallucination from 29.0% to 50.0% and drops the pairwise-preference win rate from 60.5% to 24.0%. Same model, same scenes, same prompts, only the trajectory block differs. That is the finding that makes the thesis true.
Downstream, on the AD-MCQ-500 test set:
•
Training-free DEFT (just the two-turn inference structure, no training) lifts Qwen3-VL-8B accuracy from 28.1% to 56.6% and Qwen3.5-4B from 34.0% to 65.6%.
•
Under matched reinforcement learning with only the correctness reward, deferred exposure beats joint exposure by 15.3 pp on the 8B model.
•
Full DEFT-RLVR reaches 77.9% on the 8B model with the highest causal-faithfulness score across variants.
•
Distillation from a 397B teacher hits 84.1% accuracy but costs ~5 pp of general visual capability; the reinforcement-learning path gains ~1.3 pp on general benchmarks instead.
•
Cross-domain transfer to nuScenes scenes holds: 39.6% \u2192 49.5% with DEFT-RLVR.
Reach for this pattern when you’re training an agent whose “correct answer” is a logged outcome you’d naturally want to show the teacher. The trap is that any teacher with the answer in-context will confabulate a scene-grounded reason for it, and your student learns the confabulation style. Instead: force the teacher to commit to a decision from evidence alone, then verify against the logged outcome as a separate step. Structure the verification as picking from a small candidate set your verifier can check exactly, not as freeform generation.
The paper doesn’t mention a code release; the benchmark construction is described in enough detail (K-means codebook of 8192 prototypes over ~489k trajectories from Waymo Open E2E plus an internal corpus, structured distractor recipes) that it’s reproducible in principle, but no repo link appears in the text. Models used are Qwen3-VL-8B-Instruct, Qwen3.5-4B, Qwen3.5-397B-A17B as teacher, and Qwen3.6-35B-A3B as rubric generator and judge.
If your teacher can see the answer, it will invent the reasoning. Give it the scene, take its decision, then verify.
•
The framing works because there’s a clean verifier: a discrete candidate set with one correct option. In domains without a deterministic verifier, the outcome-gated rubric reward has nothing to gate on and the whole loop degrades.
•
The candidate set is built from a codebook of logged trajectories. The paper explicitly notes it doesn’t certify that the logged future is the only safe choice, only that it’s the demonstrated one, so “correct” here means “matches demonstration,” not “is safe.”
•
Gains on general visual benchmarks are small (~1 pp) and mixed across sub-benchmarks. Framing this as “preserves general capability” is fair; framing it as “improves general capability” leans on averages that hide per-benchmark regressions like the 7.79 pp drop on one RefSpatial split for the 4B model.