- Rating
- 1
- Confidence
- 3
- Informativeness
- 1
- Presentation
- 2
- Soundness
- 1
Read review
Paper summary
The paper is based on the hypothesis that when video language models answer spatial questions (like whether an object is to the left or right of another), simply remembering past video frames isn't enough. As the camera moves, relative left-and-right visual positions shift depending on where the observer stands and faces. The authors calls this the Observation-to-Query (O2Q) Bottleneck.
And propose a solution they call QUASAR.
Instead of forcing the model to first generate a separate, 3D scene representation before answering, QUASAR takes a direct route: it uses estimated camera positions and the question's reference view to apply small, bounded mathematical corrections directly onto the original visual features. The LLM then answers the question straight from these adjusted visual features.
The paper contributes
- A spatial latent transformation module that calculates norm-bounded spatial tweaks using camera geometry and explicit reference coordinates, writing them directly into the visual features.
- An answer-path spatial supervision: A training technique using four spatial summary tokens and geometric supervision to keep spatial context accessible along the answer path.
And the results report
- Direct visual tweaks win: Adjusting visual features directly outperforms visual-only baselines (in most cases) without requiring explicit 3D scene reconstruction.
- Benchmark gains: Built on Qwen3-VL-8B, QUASAR achieves +5.54 points on VSI-Bench and +19.88 points on VSTI-Bench over the base model.
Critical strengths
The main strength of this paper, I think, is its direct visual adjustment approach.
Instead of forcing the model to build a complex 3D map of a room before answering a question, which may be slow, and memory-intensive, QUASAR applies small mathematical tweaks straight onto the raw visual features in place.
Also, great seeing the code and scripts to run it.
Critical weaknesses
There are 2 weakness of this paper I believe are critical, and one stylistic comment
-
Tables 1 and 2 are not apples-to-apples comparisons. They compare QUASAR (built on Qwen3-VL-8B) against a mix of models using different backbones, input resolutions, and training recipes. There are no controlled, equal-budget target comparisons. More importantly, QUASAR is not the top performer; external memory models like Q-GeoMem and VLM^2 outperform QUASAR on VSI-Bench. Yes, they use external spatial memory, and I don’t think that disqualifies them.
-
The method does not seem to generalize well. While it helps with left/right spatial relative questions, it degrades performance on temporal tasks like tracking object appearance order (Table 12) and causes negative transfer on external benchmarks like CV-Bench (Table 23) and ViewSpatial-Bench (Table 20). An analysis of why, at least, would be nice.
-
[Style] Table 4 is supposedly very important, but it’s confusing and verbose (so is Figure 2). The Appendix reads like a dump of all experiments the authors made to get to the paper and is quite challenging to browse and grasp
Questions for authors
Please prioritize the weaknesses I raised above.
Other questions that could affect my assessment
- Why 4 latent updates and 4 summary tokens? Any ablations on that? Could be useful for people looking to adopt such an approach to understand the trade-offs.
- Why do the visual feature edits actively break basic temporal tasks like tracking object appearance order and require manually disabling writeback to avoid trailing baselines on CV-Bench?
- Why were QUASAR and the visual-only baseline evaluated at different SFT step checkpoints before GRPO? How much of the performance gap is just down to unequal training exposure?