- Rating
- 2
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
The paper proposes Vantage, a training-free pipeline for multi-view spatial VQA. A VLM analyzes the question relative to a reference view, picks one of 17 discrete camera actions (three magnitudes each), and writes guidance text. G3T reconstructs a gravity-aligned point cloud from the input views, the action is converted to a pose, and the cloud is reprojected into a new view. The VLM then answers from the original views, the rendered view and its own text. Experiments cover six VLMs on subsets of MindCube-tiny, MMSI-Bench, BLINK, OmniSpatial and SPINBench, with ablations and a 600-case failure analysis.
Critical strengths
- The method is simple, needs no training, and is documented well enough to reproduce: the paper gives the action table, the scene-scale unit and all prompts. It raises the five-benchmark average for all six backbones. The clearest gains are on MindCube-tiny (Gemma-4-31B Rotation 51.0 to 82.0, GPT-5.4 overall 44.2 to 54.3) and on OmniSpatial perspective taking (Qwen3.6-27B 52.6 to 64.7). This shows that a feed-forward reconstructor that only re-renders observed pixels can still help cross-view alignment, even for GPT-5.4.
- The discrete action interface and the gravity alignment are backed by evidence. In Table 4 the action interface beats direct 6-DoF prediction for all four models tested, and Figure 4 shows how actions drift without gravity alignment.
Critical weaknesses
- Novelty is overstated and the closest methods are not compared. Think3D (Zhang et al., 2026h) already lets a VLM choose an anchor camera and a rotation, renders a Pi3 point cloud from that pose, runs training-free with GPT-4.1 and Gemini-2.5-Pro, and reports results on BLINK Multi-view and MindCube. Section 2 describes Think3D as producing generic panoramic or top-down views, which is not what that paper does. pySpatial (ICLR 2026) renders novel views from VGGT reconstructions inside generated programs and reports on MindCube and MMSI-Bench; the paper does not cite it. The new parts here are the action vocabulary, gravity alignment and passing the Stage-1 text to the answer step. None of them is tested against Think3D or pySpatial with the same backbone and splits. Table 1 does not fill the gap: on the same InternVL3-8B backbone, SenseNova-SI scores about 53 points higher on MindCube-tiny and 15 points higher on MMSI, so I do not see support for the claim of "comparable or superior" results.
- The gains are not shown to come from view seeking, and many are within noise. In Table 5, text alone (analysis plus guidance, no rendered view) improves the average by 4.6 points for Qwen3-VL-4B and 2.1 for Gemma-4-31B, at least as much as the rendered view alone (3.3 and 1.6). Table 2 transfers analysis from a larger model without a no-view control. For Qwen3-VL-8B, random actions and view interpolation both score higher than the planned action in Table 4. No result is repeated across seeds. BLINK Multi-view has 133 binary questions, so three of the six models gain exactly one question. Most BLINK baselines are below the 50 percent chance level (Qwen3.6-27B scores 28.6), which suggests a parsing or left/right bias. All benchmarks are restricted to selected subsets. The abstract reports a "mean accuracy gain of 6.7%", but Table 1 gives a mean absolute gain of 3.7 points and a mean relative gain of 8.1 percent. Qwen3-VL-8B improves by only 0.3 points, so "consistent" overstates the result.
Minor points that do not affect my score: the failure annotation protocol lacks annotator and agreement details; and there are typos such as "then then" and "an VLM".
Questions for authors
- Can you compare with Think3D and pySpatial using the same backbones and splits? A re-implementation of Think3D's anchor-camera and rotation interface on your G3T reconstruction would also answer this.
- Do the random-action and interpolation rows in Table 4 still receive the analysis and guidance text? If so, was that text written for the planned action?
- How do three baselines perform: text-only with the same number of VLM calls, a fixed bird's-eye or panoramic render of the same reconstruction, and an oracle that picks the best of the 17 actions?
- Which decoding settings did you use, and how much do the Table 1 numbers vary across seeds? Why are most BLINK baselines below chance, and how are answers parsed?