Deferred Exposure of Future Trajectories for Verifiable Reasoning in Autonomous Driving VLMs

Everything the paper does sits in the clause after “yet”. Note where the defect is located: not in the model, not in the objective, but in the annotation pipeline — the shared substrate every downstream student inherits. The operative verb is “expose … to”, and it is precise. The ground-truth trajectory is not wrong information; it is correct information delivered at the wrong moment. Hold on to that distinction, because the entire method is a change of timing and nothing else.

The contribution is a diagnosis, and the evidence is human scoring rather than a benchmark number — expensive, hence only 100 scenes, which in turn means the effect had to be large to be reportable at all. The load-bearing opposition is “rationalize” against “infer”: identical prose, opposite direction of causation, which is exactly why no automatic metric catches it. The sharpest word is “especially”. An across-the-board drop would be tolerable; a drop concentrated in the scenes that actually require reasoning is not.

The obvious fix gets one clause before the objection arrives. What the objection actually says is worth restating plainly: free-form trajectory generation asks the VLM to decide, to draw precise geometry, and to respect vehicle dynamics all at once, so a wrong output no longer tells you which of the three failed. That unattributability, not accuracy, is what motivates the multiple-choice reformulation two sentences later — easy to miss if you read the reformulation as a convenience.

The “To …, we introduce …” frame carries the design constraints, and both halves are needed: verifiable, and without open-ended synthesis. Anything satisfying only one of them already existed. Be sceptical in the right place — multiple choice buys exact verification by handing the model the answer set, which smuggles the paper's own problem back in. The authors say so themselves in §1, and DEFT is the patch. Note “explicit”: candidates are coordinate sequences, not labels like “slow down”, so the decision stays at trajectory granularity.

One sentence, and it is the whole method: the data is unchanged, only its position in the sequence moves. Notice how cheap that is — no new supervision signal, no extra model at inference, nothing to retrain from scratch. The phrase “pre-decision anchor → post-decision verification target” also names the experiment the paper owes you: evidence that the deferral, rather than the extra tokens a two-turn format inevitably generates, is what helps. §5 pays that debt with a token-budget-matched ablation.

“While preserving or even enhancing general visual capabilities” is aimed at the standard objection to any domain fine-tune: that the driving gain was bought with forgetting. The hedge is honest — the general-benchmark gain turns out to be about one point in §5, and the paper's own distillation variants do lose ground. The closing triple is positioning, not result. The substantive claims are “VLM-only inference” (nothing extra is needed to reproduce it) and “controllable difficulty” (the similarity band on distractors is a dial you can turn).

A shared-premises paragraph with no novelty in it yet — but read it for scope, which is narrower than the title implies. The action expert is never touched and nothing here is measured end-to-end on the road. “Making its AD-specific reasoning capability critical to downstream planning” is the bridge the authors need: it licenses treating the VLM's reasoning as a worthwhile target in isolation. That is plausible, convenient, and not demonstrated within this paper.

The mechanism claim, stated before any evidence for it. Ask what would falsify it: if the logged future were merely a hint the annotator still had to justify from the frames, faithfulness would go up, not down. §2 tests exactly that. The appeal to cognitive psychology is doing quiet rhetorical work — it licenses the coined term “trajectory anchoring bias” without the authors having to argue the phenomenon is novel, only that it shows up here.

The two arms are named here — GT-conditioned and causal-planning — and nothing else is used for the remaining twenty pages, so fix them now. The payload is the second sentence, and it is an escalation: an unfaithful rationale is a nuisance, but a fabricated object written into training data propagates into every student trained on it. That step, from “less faithful” to “poisoned supervision”, is what makes this a paper rather than a footnote.

A concession paragraph: the authors state the fix a reader would propose and grant that it works, so the objection in the next paragraph has something to land on. Note “restores”. Solve-then-verify is presented as the existing norm in reasoning-model distillation, which makes driving the outlier that needs explaining — a stronger and more defensible position than claiming the idea is new.

This is a requirements list written as prose: whatever comes next must keep decisions at trajectory level, live entirely in text, and be checkable. Stating the requirements before the solution makes the design look derived rather than picked, and is worth imitating. “Emerge from the VLM's general reasoning capability” is also a bet — that driving competence is mostly general reasoning applied to a scene — and the general-benchmark results in §5 are the evidence offered for it.

The authors attack their own contribution one paragraph after introducing it, and this is the hinge of the paper: multiple choice does not remove the anchor, it swaps one anchor for another. The tell is “comparing the relative quality of trajectories” — a model can rank six candidates against each other without ever describing the scene, and still be right most of the time. If you keep one paragraph from the introduction, keep this one. It is what makes §4 necessary rather than decorative.