A home robot follows you through an apartment, then gets asked three questions: is the mirror on my left or my right? How far is the sofa from here? When did we pass the plant? One walk, three questions, three answer forms: an option, a number in meters, a moment in time.

Streaming spatial benchmarks now pose all three over one causal history. In practice, though, plenty of post-training recipes supervise only one answer format. A new paper from Westlake University and partners (ZJU, USTC and others) names that failure mode: interface lock-in. The model gets better and better at one way of asking, and transfers weakly, or even degrades, when the question changes shape (https://arxiv.org/abs/2610.34283).

A Failure Mode Nobody Was Measuring

The observation behind the paper is concrete. Training a spatial VLM to answer "which side of me is the object on" does not require it to preserve the metric displacement that determines that answer. Supervision for a distance estimate, meanwhile, may leave categorical relations largely unconstrained. Once post-training pushes the weights toward one readout, gains concentrate on that interface while transfer to the others stays weak.

The authors quantify this with same-data controls. A categorical-only variant reaches 58.03 on OVO-S-Bench but drops to 47.48 on OST-Bench. A trajectory-only variant posts 52.03 / 51.03 / 49.22. The full recipe reaches 53.89 / 63.93 / 52.77, for a three-benchmark mean of 56.86, the highest among the supervision settings compared in the paper, against 48.28 for the official base.

Look at the shape of that result. The full recipe actually scores below the categorical-only arm on OVO-S. It does not win every column; it trades a few points on OVO-S for a large recovery on OST.

Three Ingredients: Native Targets, Chronological Replay, Family Balance

The first stage is causal replay construction. Every answer keeps the native format of the interface it came from: categorical targets keep their candidate structure, metric targets keep magnitude, sign, units and annotation precision, temporal targets keep timestamps and event indices. Trajectory turns are replayed in their original chronological order, with each target attached to its own legal causal prefix. The authors call this temporal-origin alignment, and it is explicitly not a matter of flattening multi-round annotations into independent question-answer pairs.

The second stage is family-balanced learning. The preservation family holds 33,180 single-turn examples (30,635 categorical and 2,545 numerical). The trajectory family holds 6,381 ordered episodes, counting 45,006 supervised metric, counting, judgement and temporal turns, so the preservation side carries more than five times as many examples. The approach normalizes risk inside each family and then fixes their ratio explicitly (alpha = 0.5), so the larger family cannot dominate the shared parameters. The project page puts the full replay pool at 39,561 sequences (https://yihaohu0118.github.io/ChronoReplay/).

The base model is Qwen3-VL-8B-Instruct with the visual encoder frozen; only the language model and the visual merger are trained. The paper adds no map, no memory module and no auxiliary state loss. The trajectory stays implicit inside the VLM.

Three Benchmarks Rise Together

Under matched observation access (OVO-S at Uniform-128, OST under the complete scene-disjoint protocol, UCS covering all 8,114 timestamp-causal questions):

  • OVO-S-Bench: 48.98 to 53.89, a 4.91-point gain, with all twelve task families improving;
  • OST-Bench: 50.52 to 63.93, a 13.41-point gain, positive across all fifteen subtypes, led by Direction Estimation at +26.86, Order Judgement at +20.57 and Distance Temporal Localization at +18.36;
  • UCS-Bench: 45.33 to 52.77, a 7.44-point gain, positive across all eight subcategories, led by quantity at +10.73, object trajectory at +8.46 and observer trajectory at +7.38.

The largest gains land on exactly the items that must be composed from accumulated multi-turn state rather than from the current view.

The 13.41 Points Do Not Belong to Chronology Alone

This is the most transferable move in the paper. The 13.41 points belong to the complete recipe, not to chronological replay as a single ingredient. The authors isolate the latter with a shifted-order control that keeps every turn block, every target and every family weight, and only cyclically rotates the turn blocks: 61.81 against 64.22, an average gap of 2.41 OST points across two paired seeds (2.42 and 2.41).

The same discipline shows up in the answer-format comparison. Native format is 2.92 points higher on UCS, while Canonical format is 1.48 points higher on OST. The results cross, so no encoding wins everywhere, and the paper explicitly declines to read native format as universally superior.

Against ordinary joint SFT (51.8 / 60.7 / 49.8) and natural turn weighting (52.9 / 62.3 / 51.5), the full recipe leads on all three endpoints, by 3.23 and 1.63 points respectively on OST.

The Other Half Lives at Inference Time

Adaptive frame selection changes how evidence is accessed while the trained weights stay fixed. It averages 80.50 frames instead of 128 and lifts OVO-S from 53.89 to 57.98. The paper separates this into a training effect (+4.91) and an access effect (+4.09), noting that the access effect concentrates on L1 (+8.66). The project page frames it as 37.11% fewer frames. L4 remains the hardest level throughout, moving 35.16 to 40.17 to 42.30.

The Cold Water

First, these are author-reported numbers. Code, configurations, the frozen training-selection manifest and the scoring scripts are public (https://github.com/yihaohu0118/ChronoReplay), but the fine-tuned weights, the benchmark media and the saved predictions are not bundled, and the project page states outright that this release ships no saved predictions.

Second, cross-model comparison runs into protocol trouble. The published reference rows in the paper's tables were produced under each model's own reported protocol, and the authors themselves warn that those rows cannot be read as same-condition conclusions. The 8B score of 53.89 is numerically close to the published Qwen3-VL-235B-A22B reference at 53.64, but Gemini-3.1-Pro at 59.19 still sits ahead, and no such claim can be made against it. On UCS, humans still sit at 91.0.

Third, interface lock-in is an empirical pattern the paper repeatedly flags, not an isolated internal mechanism. The conclusion says so directly. The ethics statement is equally plain: the authors have not audited high-stakes uses such as surveillance or autonomous navigation, and do not claim suitability for them.

So What

The real contribution here may not be the 13.41 points. It is the demonstration of how to separate a recipe total from a mechanism contribution: 13.41 belongs to the full recipe, chronology is worth 2.41, and the remainder has to be attributed to family balancing and target format. Lumping them together is how evaluation goes wrong.

The more directly useful lesson is simpler. When a team notices that a model "got better" at some task, it should go measure whether the model dropped on a different way of asking the same question. Interface lock-in never shows up on the loss curve. It only surfaces on the test set where somebody changed the phrasing.