Free for humans

Imagine a new viewpoint only when the picture is not enough

Multimodal models still miss spatial questions when the right answer depends on a viewpoint the picture does not show. This paper treats imagined views as a cost the model should pay only after it decides the current evidence is not enough.

arXiv:2602.082365 min readScore 93/100 · editorial triage · not peer reviewPaper hub2026-W48

The 30-second take

  • Imagining extra viewpoints can help spatial reasoning, but doing it every time can waste computation and add misleading evidence.
  • AVIC decides whether the current picture is sufficient before it calls a world model and scales how much imagination to use.
  • AVIC-R learns that gate from answer correctness and an imagination-cost penalty, without labels for when or how much to imagine.

What the paper actually did

The authors analyze test-time visual imagination as a controllable resource for spatial reasoning. They study when static visual evidence is already sufficient, when imagining an unseen or alternative viewpoint improves the answer, and how excessive or unnecessary imagination affects accuracy and efficiency. They introduce AVIC, an adaptive test-time framework that reasons about the sufficiency of the current picture before it selectively invokes and scales a world model. AVIC-R then learns that gating and planning behavior with GRPO, using rewards for question-answering correctness and penalties for imagination cost, without any annotation of when or how much to imagine. They report results on the spatial reasoning benchmarks SAT and MMSI and the embodied navigation benchmark R2R.

What makes this disruptive

Recent systems add world models so a multimodal model can imagine views it has not been shown. This paper argues that imagination is not automatically helpful. Indiscriminate imagination can increase computation and degrade performance by introducing misleading evidence. The disruptive piece is the control layer: the model has to judge that more imagination is necessary, then decide how much to run. AVIC-R learns that policy from answer quality and cost rather than from human labels of the right imagination schedule. The abstract says selective control can match or outperform fixed imagination strategies with substantially fewer world-model calls and language tokens, and that AVIC-R surpasses proprietary baselines including GPT-4o and GPT-4.1 while invoking the world model less often.

Why it matters (outside the lab)

Spatial questions show up whenever a system must act from a partial view. A navigation model may need to imagine what a scene would look like from another angle, or the picture it already has may be enough. Generating extra views on every question spends computation and can steer the answer toward a scene the world model invented. The practical claim is that test-time imagination should be analyzed and controlled if spatial reasoning is going to be both efficient and reliable. The same choice matters for any assistant that manufactures extra evidence while it is answering, rather than only while it is training.

Limitations & open questions

The abstract does not show accuracy figures, token counts, or how often imagination was critical, marginal, or detrimental on each benchmark. It does not describe the world model, the GRPO training setup, or the cases in which selective control still loses to a fixed strategy. It does not show results outside SAT, MMSI, and R2R. The statement that AVIC-R surpasses GPT-4o and GPT-4.1 does not include the scores or the evaluation protocol. Whether the same gate works on questions those benchmarks do not cover is not shown.

Explain ladder

Default article depth

A multimodal model can be asked how a scene would look from a viewpoint it has not seen. This paper says the model should first decide whether the picture it already has is enough. Only then should it spend computation asking a world model to imagine more. Imagining too often can invent misleading evidence. Their adaptive method learned that habit without labels for when to imagine, and the abstract says it beat named proprietary models while calling the world model less often.

Key terms

World model
In this paper, a model used to imagine how a scene would appear from an unseen or alternative viewpoint so the reasoning system can use that imagined view as extra evidence.
AVIC
The adaptive test-time framework that judges whether current visual evidence is sufficient before it selectively invokes and scales visual imagination.
AVIC-R
The variant that learns the imagination gate and plan with GRPO from answer-correctness rewards and imagination-cost penalties, without annotations of when or how much to imagine.
GRPO
The training method the abstract names for learning the imagination policy. The abstract does not spell out the acronym or the training hyperparameters.
Test-time imagination
Extra visual reasoning done while the model is answering, by imagining viewpoints, rather than only during training. The paper treats it as a controllable resource that can help or hurt.

Sources

Related explainers

Same topic and week first — keep exploring the scarcity → abundance map.

Editorial explainer · not peer review · always read the primary paper.

Byline: Disruptive Concepts editorial.