Reason, Then Re-reason: Cross-view Revisiting Improves Spatial Reasoning

ICML 2026

Chaofan Ma1* Zhenjie Mao1* Yuhuan Yang1 Fanqin Zeng1
Yue Shi2 Yingjie Zhou2 Xiaofeng Cao3 Jiangchao Yao1
1Cooperative Medianet Innovation Center, Shanghai Jiao Tong University
2Shanghai Jiao Tong University 3Tongji University
Paper

* Equal contribution. Co-first authors are listed alphabetically.

ReRe teaser cases showing egocentric video, synthesized novel view, and corrected reasoning

Spatial reasoning should be revisitable. Given an egocentric video, an MLLM commits to plausible but wrong answers when the camera trajectory leaves key evidence occluded. ReRe forms an initial hypothesis (Reason), then revisits it under a synthesized novel view (Re-reason) that exposes the complementary geometry, flipping wrong answers to right, as shown below for object counting and route planning.

Abstract

Spatial reasoning from egocentric videos is inherently challenging because the observable evidence is constrained by the camera trajectory. Existing methods rely on single-turn inference, forcing models to resolve geometric ambiguity through semantic priors rather than verifiable evidence. We argue that spatial reasoning should be revisitable: conclusions formed under limited evidence should remain open to revision when complementary viewpoints become available. Building on this insight, we propose Reason, then Re-reason (ReRe), a training-free, inference-time framework with two phases. In the Reason Phase, an MLLM forms a spatial hypothesis from the original video. In the Re-reason Phase, it verifies or revises the hypothesis by observing a synthesized novel-view video. To enable effective cross-view revisiting, we design a Geometry-to-Video pipeline that renders strategically complementary novel views from predicted 3D geometry. Extensive evaluations on VSI-Bench and STI-Bench demonstrate that ReRe substantially boosts open-source MLLMs to rival proprietary state-of-the-art performance.

Revisitable Reasoning Reframes spatial reasoning as a revisitable process: the first answer is a provisional hypothesis, open to revision rather than a final conclusion.
Training-free & Inference-time Works entirely at inference time on frozen MLLMs: no fine-tuning, no extra supervision, and no architectural changes.
Open-source Rivals Proprietary Lifts open-source MLLMs to rival proprietary state-of-the-art models on both VSI-Bench and STI-Bench.

Motivation

Why Single-turn Spatial Reasoning Is Fragile

Egocentric videos are trajectory-conditioned: the visual evidence is limited to what the recorded camera path happens to reveal, and the temporal order of frames rarely aligns with the scene's actual spatial topology. Compounding this, general-purpose MLLMs lack explicit 3D world modeling and only implicitly enforce cross-frame correspondence. So when a model must answer in one pass, it fills missing geometry with semantic priors and produces a plausible but wrong conclusion.

ReRe changes the interaction pattern. Instead of asking the MLLM to directly finalize an answer, it first surfaces a traceable spatial hypothesis and then explicitly checks that hypothesis against new cross-view evidence. This converts spatial reasoning from one-shot perception into hypothesis-driven verification.

Framework

Reason, Then Re-reason

Overview of the ReRe framework

Overview of the ReRe framework. Given an egocentric video and a spatial query, ReRe operates in two phases: the Reason Phase forms an initial hypothesis from the original view, and the Re-reason Phase verifies or revises it against a synthesized allocentric view. A Geometry-to-Video pipeline produces this complementary view from recovered 3D geometry via trajectory planning and view rendering.

Phase I

Reason

The MLLM observes the original egocentric video, identifies objects and spatial cues, infers relations, and outputs a provisional answer with an explicit thinking trace.

Phase II

Re-reason

The MLLM compares a synthesized allocentric video with its prior hypothesis, reflects on inconsistencies, and confirms or revises the final answer.

Geometry-to-Video

Making 3D Evidence Visible to MLLMs

Geometry-to-Video pipeline

Geometry-to-Video pipeline. ReRe predicts a 3D point cloud from the egocentric video via VGGT, plans a scene-spanning Oblique Sweep trajectory, and renders the point cloud into temporally coherent video frames via point-based rasterization. This makes the geometric evidence natively consumable by frozen MLLMs without requiring architectural modifications or additional encoders.

The synthesized view is designed around two principles. Geometric complementarity requires a viewpoint that reduces inter-object occlusion and maximizes spatial coverage. Native compatibility means the geometric evidence must be presented in a familiar visual format rather than raw 3D representations like point clouds, so frozen MLLMs can consume it without architectural modifications.

Comparison of allocentric trajectory designs

Allocentric trajectory design. The Oblique Sweep follows a diagonal path through the scene center with an elevated tilt. The Mid-level Traverse moves horizontally along the diameter at a fixed elevation, while the Bird's-eye Orbit circles the scene center from a top-down perspective.

Experiments

Does ReRe Actually Work

ReRe is evaluated zero-shot on VSI-Bench and the Static Understanding subset of STI-Bench, consistently improving frozen open-source models across the Qwen and InternVL families with no additional training. The largest gains appear in tasks that benefit from extra geometric evidence, such as object size, distance, relative direction, and spatial relation reasoning. For each benchmark we report the full subtask breakdown.

VSI-Bench Results

Model Avg. Obj. Count Abs. Dist. Obj. Size Room Size Rel. Dist. Rel. Dir. Route Plan Appr. Order
GPT-4o 34.0 46.2 5.3 43.8 38.2 37.0 41.3 31.5 28.5
Qwen2.5-VL-3B 26.4 15.8 23.9 33.4 27.8 20.5 34.4 29.9 21.5
+ ReRe 28.2 +1.8 16.7 25.0 35.3 25.3 31.4 35.7 28.9 17.3
Qwen2.5-VL-7B 24.8 11.2 12.2 22.8 29.9 33.8 36.0 32.0 24.9
+ ReRe 29.5 +4.7 18.0 18.8 40.0 31.1 34.1 35.3 32.5 22.0
Qwen3-VL-2B 22.5 14.2 14.7 29.8 10.8 19.9 34.1 23.7 19.4
+ ReRe 31.0 +8.5 17.4 23.4 50.5 21.0 25.8 36.7 27.8 26.2
Qwen3-VL-4B 30.7 14.8 22.0 41.6 33.5 30.1 38.4 23.7 29.5
+ ReRe 36.5 +5.8 21.1 26.7 50.2 35.8 37.2 43.1 28.9 34.1
Qwen3-VL-8B 30.5 16.4 20.7 43.0 28.0 36.8 35.1 22.7 26.9
+ ReRe 35.8 +5.2 19.5 25.8 49.8 31.3 39.7 41.0 29.4 33.8
InternVL2.5-4B 31.3 37.6 24.5 37.4 22.7 32.0 33.2 27.8 26.9
+ ReRe 32.3 +1.0 35.8 23.3 36.9 23.7 32.5 42.1 30.9 22.8
InternVL2.5-8B 35.5 22.5 28.4 45.6 35.3 35.6 43.3 32.5 29.9
+ ReRe 36.7 +1.2 19.1 29.7 46.7 37.9 36.5 46.9 31.4 32.0
InternVL3-2B 26.5 42.2 22.8 26.4 17.6 25.2 34.3 25.8 10.8
+ ReRe 29.9 +3.4 37.9 24.3 27.2 16.6 32.0 43.2 26.8 18.3
InternVL3-8B 32.6 40.4 23.3 43.4 30.1 34.9 33.4 30.9 19.3
+ ReRe 35.5 +2.9 38.8 31.1 44.4 30.0 37.2 37.3 30.9 23.6

STI-Bench (Static Subset) Results

Model Avg. Dim. Meas. Spatial Rel. 3D Video Grounding
GPT-4o 31.0 24.9 49.6 28.1
Gemini-2.0-Flash 36.9 33.7 50.0 33.7
Gemini-2.5-Pro 37.1 34.2 53.4 32.3
Qwen3-VL-2B 22.2 18.0 31.5 21.8
+ ReRe 30.2 +8.0 24.6 50.0 26.2
Qwen3-VL-4B 29.7 29.8 41.8 24.3
+ ReRe 34.4 +4.7 33.6 48.6 28.7
Qwen3-VL-8B 27.9 29.4 43.8 19.2
+ ReRe 30.9 +3.0 29.1 44.5 26.2
InternVL2.5-4B 30.5 25.9 43.1 29.0
+ ReRe 30.6 +0.1 22.2 43.2 32.5
InternVL2.5-8B 32.1 30.1 45.9 27.8
+ ReRe 34.8 +2.7 33.2 49.3 29.7
InternVL3-2B 22.6 20.1 28.8 22.1
+ ReRe 26.7 +4.1 22.5 43.8 22.7
InternVL3-8B 24.5 20.1 44.5 19.2
+ ReRe 27.8 +3.3 28.0 37.0 23.3

Ablations

What Makes ReRe Work

The revisiting process is what matters. Naively combining the two views barely helps: concatenation disrupts temporal coherence, and interleaving still lacks a focused verification objective. It is the structured two-phase process, not merely having an extra view available, that drives the gain.
New evidence, not just thinking twice. Re-reasoning over the original video again performs even worse than the baseline: without new information, a second pass only amplifies the initial hallucination. The bottleneck is partial observability, not insufficient deliberation. ReRe works because the second phase brings in complementary geometry.
The two views are complementary. Reasoning on the synthesized view alone falls short of the full pipeline: it captures geometry but loses fine visual detail. ReRe keeps the egocentric video as the semantic anchor (what objects are) and uses the synthesized view for structural disambiguation (where they are).
Oblique Sweep is the right trajectory. A mid-level traverse fails to resolve occlusions, while a strict bird's-eye orbit drifts too far from the pre-training distribution and causes recognition failures. The Oblique Sweep stays elevated enough to expose hidden layouts, yet oblique enough to remain visually recognizable.

Qualitative Results

Cross-view Revisiting Corrects Spatial Errors

Qualitative ReRe results on VSI-Bench

Qualitative results on VSI-Bench. ReRe resolves spatial ambiguities across four representative cases: (a)-(b) object counting, where synthesized novel views reveal previously unobserved objects such as a second monitor or a second bed; (c) absolute distance, where the expanded view clarifies the separation between the sofa and bed; and (d) relative direction, where the synthesized view disambiguates the window, door, and lamp configuration, correcting a right-side prediction to the correct left. In each case, the Re-reason Phase uses newly synthesized geometric evidence to revise an erroneous initial judgment.

Discussion

How ReRe Stays Robust and Practical

What if the 3D reconstruction is imperfect? ReRe does not need metric-perfect geometry. VGGT fuses the whole video into one point cloud, so regions occluded in one frame are recovered from others; low-confidence points are left blank, which frozen MLLMs read as blind spots rather than fabricated objects; and the egocentric video stays the semantic anchor, so the synthesized view only ever serves as verification.
What does the extra step cost? The overhead is concentrated in a single 3D reconstruction step, while rendering and the second reasoning pass add little on top. This cost is orthogonal to our contribution: the reasoning protocol stays fixed while 3D backbones keep getting faster, so ReRe inherits those speedups. And since geometry depends only on the scene, it can be reconstructed once and reused across every query about the same environment.

BibTeX

@inproceedings{ma2026reason,
  title={Reason, Then Re-reason: Cross-view Revisiting Improves Spatial Reasoning},
  author={Ma, Chaofan and Mao, Zhenjie and Yang, Yuhuan and Zeng, Fanqin and Shi, Yue and Zhou, Yingjie and Cao, Xiaofeng and Yao, Jiangchao},
  booktitle={International Conference on Machine Learning},
  year={2026}
}