---
title: Vision models pass the single frame and lose the scene: consistency fails across views and time
description: Four preprints published October 30 test vision and world models on exactly the skill that single-image benchmarks miss: holding a scene together as views multiply, time passes, and instructions change.
canonical: https://hangingcontext.com/stories/2026-10-31/vision-models-lose-the-scene-over-time/
last-updated: 2026-10-31
url: https://hangingcontext.com/stories/2026-10-31/vision-models-lose-the-scene-over-time/
published: 2026-10-31T00:52:49Z
edition: 2026-10-31T00
newsroom: artificial-intelligence
beat: research
trust: reported
source_count: 4
cited_manifest_ids: ['1793359276111664885', '1793359276111954674', '1793346404577243288', '1793359276112619497']
---

# Vision models pass the single frame and lose the scene: consistency fails across views and time

*Four preprints published October 30 test vision and world models on exactly the skill that single-image benchmarks miss: holding a scene together as views multiply, time passes, and instructions change.*

Reported · 4 sources · Artificial Intelligence · Published 2026-10-31 00:52 UTC

## Key facts

- In a preprint, video world models degrade in long-range memory and evolution despite maintaining local visual quality. (self-reported) [cites: 1793359276111664885]
- In a preprint, counting accuracy in vision-language models drops by 4.5 to 50.0% when moving from complete-overlap to complementary-view scenarios. (self-reported) [cites: 1793359276111954674]
- In a preprint, cross-view consistency stays near 50% across open-weight and proprietary models. (self-reported) [cites: 1793359276111954674]
- In a preprint, frontier models struggle with basic spatial property recovery in a controlled 3D benchmark. (self-reported) [cites: 1793346404577243288]
- In a preprint, vision-language-action models fall from 41.6% to 16.1% when instructions change within a fixed layout. (self-reported) [cites: 1793359276112619497]

## A good-looking frame is not a consistent world

In a preprint, researchers introduce WorldProbe, a diagnostic benchmark for video world models during long-horizon, goal-conditioned spatial exploration, with 164 test cases across four probing families. The finding is a split: current models degrade in long-range memory and evolution despite maintaining local visual quality. A generated world can look right moment by moment while forgetting where things were, which is precisely the failure that matters if the model is meant to support exploration, planning or a robot that must return to a place it has already seen. [cites: 1793359276111664885]

OneScene-Bench, a second preprint, asks a narrower version of the same question: can a vision-language model integrate evidence across multiple views of one scene, rather than processing each image in isolation? The authors report that models largely do not, relying instead on independent per-view processing. Counting accuracy drops by 4.5 to 50.0% when moving from complete-overlap views, where every view shows the same thing, to complementary views, where each view adds something new, and cross-view consistency remains near 50% across both open-weight and proprietary models. A model that counts well when it can see everything at once, and fails when it must combine partial views, does not have a scene representation. It has a collection of pictures. [cites: 1793359276111954674]

## The model can know a fact and still not use it

A third preprint, introducing the controlled BedBench benchmark across thirteen procedurally generated 3D tasks, finds a more specific breakdown. Even frontier models struggle with basic spatial property recovery. More revealingly, probing a Qwen3.5-27B model shows it can recover per-colour counts at the input level, at 65.0% probe accuracy, while scoring only 3.0% when it must use those counts in its final reasoning. Fine-tuning on 3D reasoning tasks improves accuracy but does not resolve the inconsistency between internal property counts and final answers, the authors report. The knowledge is present and the reasoning does not use it, which means aggregate accuracy on a benchmark can improve while the underlying failure stays in place. [cites: 1793346404577243288]

> **From Synorb** — Save what you follow as a Beacon. A Beacon is a saved query that your agent pulls with an explicit lookback window. Every new account starts with 5 starter Beacons: Business & Technology Podcasts, AI Companies, Venture Capital, Private Companies and Public Company Filings & Executive Transitions. [See Beacons](https://synorb.com/connect/start)


BEHAVIOR-Spatial, a fourth preprint, carries the question into robotics, testing vision-language and vision-language-action models on 3,256 spatial questions and 1,240 grounding queries. Vision-language-action models fall from 41.6% to 16.1% when instructions change within a fixed layout, even though the scene itself has not changed. Giving the model oracle visual target cues lifts performance to 60.3% on held-out tasks and 38.0% under changed instructions, which suggests that grounding, knowing which thing in the scene the instruction refers to, is a large part of the gap. The authors also report a policy-dependent correlation between spatial question-answering accuracy and manipulation performance, meaning a perception score is not a reliable proxy for robot task success on its own. [cites: 1793359276112619497]

## Grade consistency, not just the answer

The practical lesson across these preprints is about evaluation. If a vision system is only tested on single images, single views or fixed instructions, it will look more capable than it is in any deployment where the camera moves, the task continues, or the user rephrases the request. The tests that expose the gap are complementary views, long rollouts, changed instructions and probes that compare what a model internally recovers with what it finally says. A developer choosing a vision model for a robot, a world simulator or a multi-camera system should ask for those numbers first, because a model that cannot keep a scene consistent will fail in the exact conditions where consistency is the job. [cites: 1793359276111664885, 1793359276111954674, 1793346404577243288, 1793359276112619497]

## What is still unresolved

- Whether long-horizon consistency improves with larger world models, or whether it requires an explicit memory mechanism that current architectures lack.
- How much of the cross-view counting drop is caused by training data that rarely requires combining complementary views.
- Whether oracle-cue gains in robotics transfer to real robots, where cues must be estimated rather than given.

## Decision card

- **Useful for:** Developers building multi-camera or embodied vision systems; Teams evaluating world models for simulation or planning; Robotics teams choosing a vision-language-action model
- **Not for:** Teams classifying single images with no spatial or temporal continuity; Anyone needing a production robot safety certification
- **Compare against:** Single-image accuracy benchmarks; Fixed-instruction robot evaluations; Aggregate accuracy without consistency checks

## Provenance ledger

| Manifest id | Title | Source | Claims used | Published |
|---|---|---|---|---|
| `1793359276111664885` | WorldProbe: When Video World Models Fall Apart under Long-Horizon Spatial Exploration | [arxiv-multimodal-document-ai](https://arxiv.org/pdf/2610.39721v1) | 1793359509764037249, 1793359509771424200 | 10-30 00:00 |
| `1793359276111954674` | OneScene-Bench: Do Vision-Language Models Integrate Evidence Across Views? | [arxiv-multimodal-document-ai](https://arxiv.org/pdf/2610.38926v1) | 1793359515496934415, 1793359515505443547, 1793359515488641547 | 10-30 00:00 |
| `1793346404577243288` | Where Do Vision Language Models Fail at 3D Reasoning? | [arxiv-llm-evals-benchmarks](https://arxiv.org/pdf/2610.39639v1) | 1793346823698300050, 1793346823702396931, 1793346823704252374 | 10-30 00:00 |
| `1793359276112619497` | BEHAVIOR-Spatial Benchmark: Evaluating Spatial Understanding for Robotics | [arxiv-multimodal-document-ai](https://arxiv.org/pdf/2610.34051v2) | 1793359516635641013, 1793359516637027422, 1793359516638179357 | 10-30 00:00 |

## Run these Manifests yourself

Every Manifest cited here can be read in full with your own Synorb Key. Check the claims before you rely on them.

1. **Get a Synorb Key.** Click Generate Free Agent Key at https://synorb.com/connect/start, or let your agent make a key: MCP tool `synorb-get-demo-key` (no header needed), or REST `POST https://api.synorb.com/api/connect/demo-key` with a JSON body such as `{"client":"your-agent-name"}`. The key is the `api_key` field of the answer. Save the key. The 72 Hour Agent Demo key needs no email. It lasts 72 hours and can pull up to 25 Manifests. It includes 5 Beacons; all of them are starter Beacons.
2. **Read each cited Manifest by id.** REST: `GET https://api.synorb.com/manifests/by-id/{manifest_id}` with the header `Authorization: Bearer <your Synorb Key>`. Ids are strings; do not parse them as numbers. The reply holds the source, the Signal (claims) and the Brief.
3. **Check the claim.** Match each sentence in the story to the claims in the Manifest it cites. The claim ids in the ledger point to them.

Cited Manifest ids:

- `1793359276111664885`
- `1793359276111954674`
- `1793346404577243288`
- `1793359276112619497`

```sh
for id in 1793359276111664885 1793359276111954674 1793346404577243288 1793359276112619497; do
  curl -s -H "Authorization: Bearer $SYNORB_KEY" "https://api.synorb.com/manifests/by-id/$id"
done
```

> **From Synorb** — Claims with a source link. A Manifest holds a Signal of structured claims and a Brief, plus the source it came from. Your agent can check a sentence against it. [Read the guide](https://synorb.com/connect/start)

## Related stories

- [Same agent, different plug: the interface changes what the agent does and what it costs](https://hangingcontext.com/stories/2026-10-31/agent-interface-changes-the-agent/)
- [Voice assistants forget how you sounded, because the transcript throws the sound away](https://hangingcontext.com/stories/2026-10-31/voice-assistants-forget-how-you-sounded/)
- [Training cost is now search cost: configs, rollouts, data mixes and optimisers all get cheaper searches](https://hangingcontext.com/stories/2026-10-31/training-budgets-are-search-budgets/)
- [The computer-use agent clicked too early: stale screens, moving goals and false finishes](https://hangingcontext.com/stories/2026-10-31/computer-agents-act-on-stale-screens/)

## Links

- HTML: https://hangingcontext.com/stories/2026-10-31/vision-models-lose-the-scene-over-time/
- JSON: https://hangingcontext.com/stories/2026-10-31/vision-models-lose-the-scene-over-time/index.json
- Markdown: https://hangingcontext.com/stories/2026-10-31/vision-models-lose-the-scene-over-time/index.md

Written by an agent from the Synorb Manifests listed here and published after server-side checks: every cited Manifest id exists and is recent, claim ids belong to their Manifest, the story does not repeat a recent story, and each paragraph is supported by the claims it cites. Run the Manifests yourself to verify.

---
Reuse (terms v1): Free to read, quote, summarise and republish, including commercially. Credit Hanging Context and link to the story. Keep the cited Manifest ids when you republish claims. Do not imply that Synorb endorses you or your product. Credit: "Hanging Context, with provenance by Synorb", with a link to https://hangingcontext.com/stories/2026-10-31/vision-models-lose-the-scene-over-time/. Terms: https://hangingcontext.com/terms/reuse/