ECCV 2026 Multimodal VLM LLM Reasoning Multi-view spatial reasoning Structured chain-of-thought GRPO Cross-view alignment Multimodal post-training

ViewFusion: Structured Spatial Thinking Chains for Multi-View Reasoning¶

Conference: ECCV 2026
Paper: ECCV 2026
Code: https://github.com/taoxj2001/ViewFusion
Area: Multimodal VLM / LLM Reasoning
Keywords: Multi-view spatial reasoning; Structured chain-of-thought; GRPO; Cross-view alignment; Multimodal post-training

TL;DR¶

ViewFusion splits multi-view spatial reasoning into a two-stage chain-of-thought that aligns views first and answers afterwards (<spatial_thinking> → <thinking> → <answer>), cold-starts it with synthesized reasoning traces via SFT, and then applies GRPO with a composite reward over correctness, format validity and length, raising Qwen3-VL-4B-Instruct from 30.1% to 35.4% on MMSI-Bench (and from 37.0% to 77.0% on MindCube).

Background & Motivation¶

Multi-view spatial reasoning requires a model to establish spatial correspondence across viewpoints — how the camera moved, which objects stay the same under a viewpoint change, and how occlusion evolves as the perspective shifts — and then answer direction, position, or perspective-transformation questions. Current MLLMs are not reliable at this. Work such as Cambrian-1, Spatial-MLLM, SpaceR and ViLaSR improves single-image spatial ability mainly through spatially aware instruction tuning and curated supervision, and Visual Spatial Tuning (VST) markedly improves generalization after training on large-scale spatial data; yet as soon as the input becomes multi-view, models still treat each image as an independent source of evidence and jump straight to answering, using the multi-view context as weak auxiliary information rather than as complementary observations that must be jointly aligned. Even when an intermediate "observation" or description stage is inserted (the observe-first designs of HumanOmniV2, Visionary-R1, Observe-R1), that step is essentially view-local: it states which salient entities appear in each frame but never how those entities transform across viewpoints. The most informative cues are therefore exactly the ones dropped — whether an object disappears because it is occluded or because it left the scene, how the relative ordering of background landmarks changes under camera motion, and how scale and perspective distortion shift apparent positions. The resulting reasoning trace reads coherently but is grounded in an incomplete cross-view spatial model, so predictions become brittle on questions that genuinely require multi-view integration. The failure modes found in error analysis are telling: two views treated as two different locations, left/right confused after a turn, or objects matched to the wrong counterpart — mistakes that additional textual deliberation cannot repair.

The authors further argue that using RL as a post-training strategy does not fix this. Methods such as GRPO do improve task-level performance from rollouts and often encourage longer deliberation, but their reward is defined on the final answer alone: in multi-view settings the signal is sparse and places no constraint whatsoever on how the model should use the available views. In the authors' preliminary study, models trained with vanilla GRPO frequently develop shortcut behaviour — they start solving the question before integrating the full multi-view context, or lean predominantly on one view while treating the rest as incidental. The reasoning looks more detailed, yet the cross-view spatial model underneath remains incomplete: providing more views does not reliably translate into better multi-view reasoning and can even inject noise that amplifies shortcuts and destabilizes intermediate reasoning.

The crux is therefore clear: cross-view alignment is a precondition for answering correctly, yet in existing training paradigms it receives no explicit supervision at all and is expected to fall out of the answer objective for free. This paper's angle is to flip that from an implicit byproduct of answering into a deliberate first step. Core idea: use a structured two-stage chain-of-thought to make cross-view spatial pre-thinking explicit — the first stage (<spatial_thinking>) infers viewpoint relations and cross-view correspondences and forms a textual spatial workspace, the second stage (<thinking>) reasons about the question conditioned on that workspace, and a synthesized-trace SFT cold start plus GRPO with a composite reward (correctness, format validity, length) keeps this two-stage behaviour stable during RL.

Method¶

Overall Architecture¶

The input is 2–8 multi-view images of the same scene plus a multiple-choice question; the output is a fixed three-part text sequence: <spatial_thinking> (cross-view spatial pre-thinking), then <thinking> (question-driven reasoning), and finally <answer> (a single option). Training proceeds in two steps: SFT on 18K synthesized reasoning traces so the model learns the protocol under teacher forcing, followed by GRPO on 16K instances that carry no reasoning traces, where a composite reward pushes the policy towards correct answers while keeping the protocol from collapsing. Nothing in the pipeline adds an alignment module, spatial tokens, or a 3D geometry head — cross-view alignment is carried entirely by the language chain itself.

%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400}}}%%
flowchart TD
    A["Multi-view input + question"] --> B["Structured two-stage chain-of-thought"]
    B --> C["Synthesized reasoning supervision<br/>trace rewriting + strict filtering"]
    C --> D["Composite reward<br/>correctness + format + length"]
    D --> E["Two-stage optimization<br/>SFT cold start + GRPO"]
    E --> F["Final option"]

Three things are worth stating up front, and the three key designs below correspond to them. Where the "structure" actually lies: not in a graph structure, spatial tokens, or quantized coordinates, but in staged decomposition + a fixed tag order + a division of labour — the first thinking segment is responsible only for cross-view alignment, the second is where question-driven reasoning and reference-frame mapping may happen, the two roles may not be swapped or merged, and tag opening/closing and ordering are hard-constrained during training. How multi-view information is fused inside the chain: the method does not encode each view separately and then match them; instead the same autoregressive text chain does the aligning — inside <spatial_thinking> the model explicitly links the appearances of the same element across views, states whether the camera rotated or translated, judges whether an object left the field of view or became occluded, and compares the relative ordering of background landmarks across the two images. That text is itself the intermediate workspace the later reasoning conditions on, so cross-view correspondence is written out rather than computed implicitly in some attention layer. Where the supervision comes from: on the SFT side, Qwen3-32B-VL-Instruct rewrites the original rationales of public datasets into three-part traces that are then strictly filtered; on the RL side only the multi-view input and the answer are kept, deliberately without rewritten traces, so that optimization is driven by outcome rewards.

Key Designs¶

1. Structured two-stage chain-of-thought: turning cross-view alignment from a byproduct into a mandatory first step

What existing paradigms (including observe-first variants) share is that they let the model "solve while observing": the observation step is a within-view description, and the answering step jumps from those descriptions to the answer, leaving cross-view relations implicit throughout. ViewFusion fixes the output into three segments, <spatial_thinking>, <thinking> and <answer>, and assigns them distinct duties. The first segment must answer "what happened between these views" — which visual elements the images share, whether the camera rotated or moved forward, which objects vanish due to occlusion, and how the relative ordering of background landmarks changes — ending in a consistent description of the cross-view spatial situation. Only the second segment performs question-driven reasoning on top of that workspace and maps the target object's position into the reference frame the question specifies (for example, "facing north"). The third segment emits a single option letter.

Two mechanisms explain why this works. First, it writes "reconcile the multi-view evidence before concluding" into the causal order of the generated sequence: the model cannot begin solving before it has written its alignment conclusion, so the single-view shortcut is blocked at the entrance rather than merely discouraged. Second, alignment errors move from being implicit and unattributable to explicit and visible token by token, which makes them directly constrainable during training (the format reward requires this segment's tags to exist and be closed) and easy for a human to inspect to see exactly where the model misaligned. This is also the essential difference from "letting the model think longer": what is constrained is the content and order of the thinking, not its length.

2. Synthesized reasoning supervision: rewriting original rationales into three-part traces and filtering the data strictly

The public multi-view datasets used here (VST-500K and MindCube-Trainset) contain only question-answer pairs, and their original rationales are not written according to the two-stage protocol; SFT on them directly would teach the old "describe first, then answer" behaviour. The construction instead uses Qwen3-32B-VL-Instruct to rewrite each original rationale into a trace containing <spatial_thinking>, <thinking> and <answer>, so that the first part explicitly carries the cross-view pre-thinking. Strict filtering rules are then applied, discarding samples with missing fields, wrong segment order, unclosed tags, or malformed formatting, yielding a clean 18K SFT corpus. A separate 16K-instance RL split is drawn from the same sources but keeps only the multi-view input and the answer used to compute outcome rewards, with no rewritten traces.

The trade-off is specific. SFT needs a structurally consistent, format-controllable cold-start corpus, otherwise the two-stage protocol is washed out by shortcut answers early in training or produces unparseable output (the paper explicitly positions SFT as the cold start that prevents early collapse). Deliberately withholding rationales from the RL split prevents the policy from overfitting to the wording of one particular teacher rationale: what it should learn is which behaviour produces correct answers, not how to paraphrase Qwen3-32B-VL-Instruct. The two splits share their sources but not their supervision, which is also what lets the structured protocol and the RL-driven capability gain be examined separately in the ablations.

3. Composite reward: correctness plus format validity plus length shaping, to control reward sparsity and generation collapse at once

RL rewards in the multi-view setting are inherently sparse, and optimizing correctness alone induces two kinds of degeneration. Behaviourally, the model bets on a salient cue from a single image or starts answering before cross-view consistency is established. Generatively, it may drop the intermediate segment altogether, emit malformed text that is hard to parse, or collapse into very short responses, leaving too little reasoning content for multi-stage inference. The reward therefore has three terms:

\[r(x,y) = r_{\mathrm{ans}}(x,y) + r_{\mathrm{fmt}}(x,y) + r_{\mathrm{len}}(x,y)\]

The correctness term \(r_{\mathrm{ans}}\) is binary: the predicted option is extracted from <answer> and scores 1 if it matches the label (all training instances are multiple-choice). The format term \(r_{\mathrm{fmt}}\) is likewise a binary indicator, equal to 1 only when all three tag pairs appear in the fixed order and each is correctly opened and closed; missing tags, swapped order, duplicated sections or malformed closures are all invalid. It closes off the escape route of skipping pre-thinking and emitting an answer directly, and is what keeps the two-stage protocol from decaying into decoration during RL. The length term \(r_{\mathrm{len}}\) is a shaping term that pays a weight \(\omega\) only when the prediction is correct and the response length falls inside a preferred interval \([\ell_{\min}, \ell_{\max}]\), and is 0 otherwise (equation (6) is garbled by OCR in the cached text and is paraphrased here from the body text; ⚠️ refer to the original paper). Experiments set \(\omega = 0.2\), \(\ell_{\min} = 320\) and \(\ell_{\max} = 512\), which penalizes both under-generation (insufficient reasoning) and over-generation (verbosity). ⚠️ Equation (7) of the paper adds the three terms directly, yet the body text also mentions a \(\lambda \in [0,1]\) (set to 0.02 in experiments) controlling the strength of format regularization; the two do not line up in the cached text, so refer to the original paper for the exact form.

4. Two-stage optimization: an SFT cold start to lock in the protocol, then GRPO to reinforce the trajectories that answer correctly

The two training stages solve two different problems. SFT at a learning rate of \(1\times10^{-5}\) maximizes the likelihood of the target sequence; its job is to engrave the protocol and format into the model so generation is stable and parseable. It is cheap and stable, but it inherits the teacher rationale's biases, and its teacher-forced objective does not match free-form generation at test time, so it cannot directly optimize task-level success. The GRPO stage uses a smaller learning rate of \(1\times10^{-6}\): for each input it samples \(K = 8\) trajectories, computes relative advantages against the group mean reward (no separate value network is needed, and variance is low), and updates the policy with a PPO-style clipped objective plus a KL regularizer against the reference policy. The benefit is visible in the experiments: the SFT model scores 32.4% on MMSI-Bench and reaches 35.4% after GRPO, while the training curves show total and accuracy rewards rising steadily and the KL divergence rising then plateauing — the policy does explore beyond the reference model, but stays under control. The format reward is near-saturated from the start, consistent with the ablation finding that format is largely learned before RL.

A Worked Example¶

Take one living-room example from MMSI-Bench (the first case of Figure 3 in the paper). Two photos show the same living room — a TV on a stand in the middle, bookshelves on both sides, an L-shaped dark sofa — and the question asks: "Suppose I am taking the first picture; in which direction relative to me is the subject of the second picture likely located? A: back left, B: front left, C: back right, D: front right."

ViewFusion aligns first. The first segment writes out the camera position and layout of the first image (facing the TV, with a bookshelf on each side), then describes the second: the camera is lower and closer to the corner where the bookshelf meets the wall, a red rug appears at the bottom edge, and the subject is the base of the bookshelf. From this it infers that the camera has rotated significantly to the left and moved slightly forward relative to the first viewpoint — effectively stepping into the space between the bookshelf and the sofa. Only then does the second segment answer: since the new camera position is to the left and forward of the first, the subject of the second image (the bookshelf base and the corner) lies to the front left of the person taking the first photo. The output is <answer>B.

Contrast the Qwen3-VL-4B-Instruct response: it merely writes "the first picture is of a bookshelf on the left side of the room, and the second picture (per the question) is of the subject at the back right", then reads the premise off directly and answers C. It treats the premise given in the question as an already-aligned conclusion; nowhere in its reasoning is there any cross-view correspondence or camera-motion inference. The comparison is instructive: the baseline does not misdescribe the objects, it simply never establishes the relation between the views.

Loss & Training¶

The SFT stage trains on 18K synthesized traces at a learning rate of \(1\times10^{-5}\) as a cold start. The RL stage runs GRPO on 16K trace-free instances at a learning rate of \(1\times10^{-6}\), sampling \(K=8\) trajectories per instance for group-relative advantages, for 1500 steps (about 18 hours on 8×H100). The reward is the sum of correctness, format validity and length shaping, with \(\omega = 0.2\) and \([\ell_{\min}, \ell_{\max}] = [320, 512]\) (⚠️ the body text additionally mentions a format regularization strength \(\lambda = 0.02\); its relation to equation (7) should be checked against the original paper). At evaluation time the prediction is extracted from the <answer> field and only a single option letter is accepted; outputs with format violations or failed parses count as incorrect. Decoding and inference hyper-parameters follow the recommended settings of Qwen3-VL-4B.

Key Experimental Results¶

Main Results¶

Evaluation covers three multi-view / multi-perspective benchmarks: MMSI-Bench (cross-view alignment, viewpoint transformation, occlusion-sensitive inference), MindCube (building a consistent mental model from limited views and perspective-sensitive reasoning under partial observability), and ViewSpatial-Bench (viewpoint-dependent spatial localization and cross-view reference frames). All questions are multiple-choice, and accuracy is reported.

Model	Size	MMSI-Bench	MindCube	ViewSpatial	Avg.
RandomChoice	–	25.0	33.0	26.3	28.1
GPT-5	–	41.8	56.3	45.5	47.9
Gemini-3-Pro-Preview	–	45.2	70.8	50.3	55.4
Qwen3-VL-4B-Instruct	4B	30.1	37.0	42.5	36.5
Qwen3-VL-8B-Instruct	8B	31.1	29.4	42.2	34.2
SpatialLadder-3B	3B	27.4	43.4	39.8	36.9
Cambrian-S-7B	7B	25.8	39.6	40.9	35.4
ViLaSR-7B	7B	30.2	35.1	35.7	33.7
VST-7B-RL	7B	34.8	39.1	42.4	38.8
ViewFusion (SFT)	4B	32.4	68.5	45.1	48.7
ViewFusion (SFT+RL)	4B	35.4	77.0	45.4	52.6

The headline result: among open-source models at the 4B scale, ViewFusion has the best average (52.6), 16.1 points above Qwen3-VL-4B-Instruct; on MMSI-Bench it gains +5.3 (35.4 vs. 30.1), on ViewSpatial +2.9 (45.4 vs. 42.5), and on MindCube it pulls 37.0 up to 77.0. It also leads the 7B VST-7B-RL (38.8) by 13.8 average points. Note that the SFT model already reaches 68.5 on MindCube, so most of that benchmark's gain comes from the synthesized supervision itself, with RL adding a further +8.5.

Fine-Grained Analysis (MMSI-Bench)¶

Model	Cam.–Cam.	Obj.–Obj.	Reg.–Reg.	Cam.–Obj.	Obj.–Reg.	Cam.–Reg.	Meas.	Appr.	Motion Cam.	Motion Obj.	MSR	Avg.
Qwen3-VL-4B-Thinking	25.8	26.6	34.5	33.7	25.9	36.1	48.4	28.8	21.6	26.3	23.2	29.0
Qwen3-VL-4B-Instruct	30.1	34.0	29.6	34.9	29.4	39.8	45.3	19.7	21.6	23.7	26.7	30.1
ViewFusion	46.2	41.5	30.9	44.2	21.2	53.0	35.9	34.9	40.5	32.9	23.2	35.4

Column names are the MMSI-Bench subcategory abbreviations: Cam./Obj./Reg. stand for camera, object and region, and a pair such as Cam.–Obj. denotes questions about the relation between the two; Meas./Appr. are the measurement and appearance parts of the attribute category; Motion Cam./Obj. are motion questions whose subject is the camera or an object; MSR is multi-step relation reasoning. Bold marks the entries where ViewFusion beats both baselines.

Ablation Study¶

Config	Cam.–Cam.	Cam.–Reg.	Motion Cam.	Avg.	Note
Full method	46.2	53.0	40.5	35.4	structured protocol + SFT + GRPO
Free-form reasoning + RL	37.6	49.4	31.1	33.4	two-stage protocol removed, RL unchanged, −2.0
w/o GRPO	28.0	47.0	32.4	32.4	SFT cold start only, −3.0 (largest drop)
w/o Format Reward	40.9	51.8	37.8	35.0	\(r_{\mathrm{fmt}}\) removed, only −0.4

Key Findings¶

GRPO is the main source of capability gain, while the structured protocol is the inductive bias against shortcuts: removing GRPO (leaving only SFT) costs 3.0 average points, the largest drop; replacing the two-stage protocol with free-form reasoning while keeping RL costs 2.0 points, concentrated in the subcategories that require viewpoint inference (Cam.–Cam. 46.2 → 37.6, Motion Cam. 40.5 → 31.1). Without an explicit pre-thinking segment, the model does fall back on shortcuts.
The format reward is a stabilizer, not a performance source: removing it costs only 0.4 points, because compliance with the three-part format is already high after SFT. The training curves agree — the format reward is near-saturated from the very beginning. Its real value is preventing the protocol from collapsing during RL.
The gains really do come from genuine cross-view alignment: against Qwen3-VL-4B-Instruct the largest improvements are in camera–camera (+16.1), camera–region (+13.2), motion–camera (+18.9) and attribute–appearance (+15.2), i.e. questions that require inferring the viewpoint or relocating objects after a viewpoint change, rather than subcategories that are solvable from a single image.
But there is clear negative transfer in some subcategories: object–region relations drop from 29.4 to 21.2, attribute–measurement from 45.3 to 35.9, and Reg.–Reg. is essentially flat (29.6 → 30.9). The paper highlights only where the gains concentrate and does not discuss these regressions; a plausible reading is that the two-stage protocol tilts capacity towards cross-view direction judgements and crowds out budget for fine-grained within-view measurement questions (this is the note author's speculation; the paper offers no analysis).
Thinking longer is not the same as thinking more correctly: ViewFusion surpasses Qwen3-VL-4B-Thinking on MMSI-Bench (35.4 vs. 29.0), a reasoning model trained on large amounts of high-quality chain-of-thought data. The bottleneck in multi-view spatial reasoning is cross-view consistency, not reasoning length or CoT quality.
Training dynamics are healthy: over 1500 GRPO steps the total and accuracy rewards rise steadily, the format reward saturates at a high level, and the KL divergence climbs then plateaus, indicating that exploration is effectively bounded by the KL regularizer.

Highlights & Insights¶

Turning an implicit latent into an explicit intermediate artifact: cross-view alignment used to be an unsupervised, uninspectable latent inside the model; this paper forces it into a readable piece of text, so it can be conditioned on (consumed by the second-stage reasoning) and constrained by the reward (the format term requires it to exist and be valid). The general lesson — make the intermediate state visible before you can constrain it — transfers to any multi-step task with hidden premises, such as intent disambiguation before tool calls or evidence consolidation before long-document QA.
Clean division of labour in the reward: the correctness term drives capability, the format term enforces a behavioural invariant, and the length term shapes the output distribution. In particular, the finding that the format reward is a stabilizer rather than a performance source is confirmed twice over, by the ablation and by the training curves, which avoids the common overclaim of selling format rewards as accuracy gains.
Deliberately withholding rationales from the RL split: SFT uses rewritten traces while RL uses only question-answer pairs. That is a more restrained choice than feeding both stages the same synthesized traces — it limits RL's objective to task success and keeps the policy from being locked into one teacher's wording. Feeding traces to both would look better on paper but be more brittle.
Zero extra parameters in the protocol: no alignment module, spatial tokens or coordinate regression head — everything is achieved through tag structure and rewards, which makes it easy to port onto any base model that already accepts multi-image input.

Limitations & Future Work¶

Single base model, single scale: validation is limited to Qwen3-VL-4B, with no 2B/8B scaling curve and no cross-base transfer (e.g. InternVL3), so it is hard to tell whether the gains come from the protocol or from this particular base model.
MindCube gains may be contaminated by the training distribution: the training data come from VST-500K and MindCube-Trainset, and MindCube is also an evaluation benchmark; a near-doubling from 37.0 to 77.0 is likely to include a substantial same-distribution component rather than generalized cross-view ability. The MMSI-Bench (+5.3) and ViewSpatial (+2.9) gains are more credible and should carry the main conclusion. The paper does not discuss this potential overlap.
SFT supervision is capped by the teacher: all traces are rewritten once by Qwen3-32B-VL-Instruct and passed through rule-based filtering, with no human verification or quality scoring. If the teacher itself misaligns views, the error is frozen in as a structurally valid but spatially wrong alignment, and the format reward only guarantees structural validity, never spatial correctness.
The reward form is largely heuristic: the length interval [320, 512] is a manually fixed constant with no sensitivity study, and the \(\lambda\) mentioned in the body text is inconsistent with the plain sum over three terms in equation (7) (⚠️ refer to the original paper); there is no evidence that the relative weighting of the format and length terms is optimal. The length reward also gives a bonus to long correct responses on top of the correctness reward, which can mask partially correct but verbose trajectories.
No cost comparison: the two-stage chain is longer, yet the paper reports neither average output length nor inference latency relative to the baselines and the Thinking model; the RL cost (18 hours on 8×H100) is given only for the 4B model, with no discussion of feasibility at smaller or larger scales.
The "workspace" is only natural language: all cross-view alignment is expressed textually, with no geometric constraints or 3D consistency checks, so it can still inherit language priors and common-sense bias (for example inventing a viewpoint relation from a prior such as "kitchens usually have a window"). A natural improvement is to add verifiable geometric or correspondence outputs to <spatial_thinking> and use a consistency check as an extra reward.

vs Observe-R1 / Visionary-R1 / HumanOmniV2: these works also make the model observe before reasoning, but their observation step is a descriptive summary of the visual input (which entities are present, what the scene is) — essentially telling the model not to answer too fast. ViewFusion instead asks for cross-view relation inference: how the camera moved, which elements correspond under a viewpoint change, how occlusion evolves. That content can only be written by integrating the views, whereas a single-view description can be written without looking at the second image at all.
vs GRPO-style reasoning RL such as R1-VL / Insight-V / Video-R1: they improve the quality of multimodal reasoning paths via step-wise dense rewards or self-generated trajectory selection, but the target is still the generic notion of "reasoning quality". ViewFusion decomposes the reward into correctness plus protocol format plus length, where the format term exists specifically to hold a structural constraint — "you must align across views" — in place, designed around one concrete failure mode of multi-view shortcuts.
vs Visual Spatial Tuning (VST): VST is a data-side route, training on large-scale spatial perception and reasoning data, and VST-7B-RL is strong on both MMSI-Bench (34.8) and ViewSpatial (42.4). ViewFusion is a protocol-and-reward-side route that reaches 35.4 / 45.4 from a 4B base, showing that with the same data sources, how the reasoning process is organized also yields substantial gains. The two routes are orthogonal and can be stacked.
vs Qwen3-VL-4B-Thinking: that model gains longer deliberation from large amounts of high-quality CoT data, yet scores only 29.0 on MMSI-Bench and even falls below the Instruct version in several subcategories. This is the paper's most telling control: the bottleneck is not how long the model thinks, but whether its thinking contains a cross-view consistency step.
Transferable insight: replace "cross-view consistency" with other hidden premises in other tasks (temporal consistency, multi-document evidence consistency, coreference consistency across dialogue turns), and the recipe — explicitly write out an intermediate workspace, use a format reward to guarantee that workspace exists, use an outcome reward to optimize correctness — should still hold.

Rating¶

Novelty: ⭐⭐⭐⭐ The two-stage "align first, then answer" protocol is not brand new (observe-first exists), but making cross-view alignment an explicit, reward-constrained intermediate workspace and systematically explaining why it suppresses multi-view shortcuts is a valuable reformulation and realization.
Experimental Thoroughness: ⭐⭐⭐ Three benchmarks do show gains, and the ablations plus training curves largely explain each component's role; however there is only one 4B base model, no scaling or cross-base experiments, the MindCube result is suspiciously same-source as training, and the length reward and \(\lambda\) lack sensitivity analysis.
Writing Quality: ⭐⭐⭐⭐ The failure-mode diagnosis (late fusion, view-local description, RL shortcuts) is concrete and example-backed, and the architecture and reward definitions are clear; the weaknesses are the inconsistency between equation (7) and the \(\lambda\) description, and the silence about the regressing subcategories.
Value: ⭐⭐⭐⭐ The method adds no modules, the recipe is clear and the code is open, so any team doing post-training of multi-image / multi-view MLLMs can reuse it directly; the stable gains on MMSI-Bench and ViewSpatial are more informative than the MindCube jump.