---
title: The benchmark said pass. The parts it did not check failed
description: In preprints, video models fabricate text outside the checked region, scoring rules swing accuracy by up to 22 points on identical outputs, and drone-agent benchmarks find semantic answers that lack spatial grounding.
canonical: https://hangingcontext.com/stories/2026-11-03/benchmark-scores-hide-failures/
last-updated: 2026-11-03
url: https://hangingcontext.com/stories/2026-11-03/benchmark-scores-hide-failures/
published: 2026-11-03T12:36:14Z
edition: 2026-11-03T12
newsroom: artificial-intelligence
beat: research
trust: reported
source_count: 5
cited_manifest_ids: ['1793703903651674898', '1793703903650068800', '1793703903651004144', '1793703903650095688', '1793703903649974345']
---

# The benchmark said pass. The parts it did not check failed

*In preprints, video models fabricate text outside the checked region, scoring rules swing accuracy by up to 22 points on identical outputs, and drone-agent benchmarks find semantic answers that lack spatial grounding.*

Reported · 5 Manifests · Artificial Intelligence · Published 2026-11-03 12:36 UTC

## Key facts

- In a preprint, video models show 0.90 to 0.94 fidelity at anchor points but 36% to 54% fabrication rates in peripheral text regions. (self-reported) [cites: 1793703903651674898]
- In a preprint, scoring rules in multimodal benchmarks shift accuracy by up to 22 points on identical model outputs. (self-reported) [cites: 1793703903650068800]
- In a preprint, a diagnostic protocol traces vision-language failures to perception, reasoning, or usage without weight access, and localization-first prompting repairs reasoning failures when localization boxes are correct. (self-reported) [cites: 1793703903651004144]
- In a preprint, a benchmark-compression method cuts high-error risk by 50.2% and out-of-distribution estimation error by 21.8% compared with random sampling. (self-reported) [cites: 1793703903650095688]
- In a preprint evaluating 45 models, drone-agent models show weaker agent-state reasoning than world-state reasoning, and semantic answers do not reliably coincide with spatial or temporal grounding. (self-reported) [cites: 1793703903649974345]

## Passing where someone looked

A benchmark that checks only part of the output will pass a model that fails elsewhere, and a preprint on image-conditioned video generation shows the failure in numbers. Models in the study achieve high fidelity at anchor points, scoring 0.90 to 0.94, while fabricating text in peripheral regions at rates of 36 to 54 percent. Existing reference-based consistency scores do not track that peripheral fabrication, and string-matching metrics cannot catch fabricated text that looks legible but has no reference counterpart. The lesson generalizes beyond video: wherever an evaluation samples a few checked locations and assumes the rest, a model can learn to be faithful exactly where the checker looks. [cites: 1793703903651674898]

A second preprint offers a way to locate failures instead of averaging over them. Its diagnostic protocol for vision-language models, called See, Think, Use, traces an answer to a perception failure, a reasoning failure, or a usage failure, without needing access to model weights. When the model's localization boxes are correct, localization-first prompting repairs reasoning failures, and allocating compute based on the measured failure stage beats uniform compute policies at lower cost. For a buyer, that distinction is actionable: a perception failure calls for better sensing or input, a reasoning failure calls for different prompting or compute, and a usage failure calls for a different interface. A single accuracy number hides all three, and fixing the wrong stage wastes the budget. [cites: 1793703903651004144]

## The scorer is part of the score

Even when the output is fully visible, the scorer can move the result. In a preprint auditing multimodal benchmark graders, the authors find that scoring rules shift accuracy by up to 22 points on identical model outputs. Parsers misinterpret or discard valid answers, and benchmarks with high non-compliance rates produce artifacts where the scorer abstains instead of guessing, skewing results. The authors' proposed remedy, a blinded human-plus-language-model adjudication protocol, reaches a Cohen's kappa of 0.956, and they recommend reporting a non-compliance rate alongside results under at least two distinct scoring rules. Until that practice spreads, a buyer comparing two models on different benchmarks may be comparing two graders, not two models. [cites: 1793703903650068800]

> **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)


Benchmark compression adds a quieter version of the same risk. Random sampling for compressing benchmarks produces a long-tailed error distribution, with rare but large errors, according to a preprint proposing a method called BestRand. That method cuts high-error risk by 50.2 percent and estimation error by 16.4 percent compared with random sampling, and reduces out-of-distribution estimation error by 21.8 percent on average across shifts in score range, model family, and time. The implication is that a cheap benchmark is not automatically a misleading one, but a benchmark compressed by naive sampling can misrank models precisely on the cases where the ranking matters most. [cites: 1793703903650095688]

## Right words, wrong place

A drone-agent benchmark closes the set with a failure that sounds abstract until it is deployed. Evaluating 45 multimodal models on grounded agent-world reasoning, the authors find that agent-state reasoning is consistently weaker than world-state reasoning, and that semantic answers do not reliably coincide with spatial or temporal grounding in aerial contexts. A model can therefore describe a scene correctly while being wrong about where it is, when something happened, or what state it is in. A lightweight optimization probe improves dual-cognition performance in the study, which suggests the gap is at least partly trainable. But the deployment lesson stands: for embodied agents, a correct-sounding answer is not evidence of correct grounding, and evaluations must test the agent's state separately from its description of the world. [cites: 1793703903649974345]

## What is still unresolved

- What standardized replacement, if any, will supersede anchor-based evaluation for generated video text.
- Whether the adjudication protocol and dual-scoring recommendation will be adopted by major benchmarks.
- How much of the drone-agent grounding gap persists after the optimization probe in real flight conditions, beyond the benchmark.

## Decision card

- **Useful for:** Executives and buyers comparing multimodal models on published benchmarks.; Teams deciding which evaluations to trust before deployment.
- **Not for:** Ranking models on a single benchmark score, or trusting a headline accuracy without its scoring rule.
- **Compare against:** Results reported under at least two scoring rules, with failure location and non-compliance disclosed.

## Provenance ledger

| Manifest id | Title | Source | Claims used | Published |
|---|---|---|---|---|
| `1793703903651674898` | Faithful Where You Look: Anchor-Based Evaluation Hides Text Fabrication in Image-Conditioned Video Generation | [arxiv-multimodal-document-ai](https://arxiv.org/pdf/2611.00079v1) | 1793704272440538501, 1793704272437196222, 1793704272442619806 | 11-03 00:00 |
| `1793703903650068800` | Same Outputs, Different Scores: Auditing the Unchecked Graders of Multimodal LLM Benchmarks | [arxiv-multimodal-document-ai](https://arxiv.org/pdf/2611.03377v1) | 1793704279515837246, 1793704279520292591, 1793704279527765296 | 11-03 00:00 |
| `1793703903651004144` | See, Think, Use: Diagnosing Perception and Reasoning Failures in VLMs | [arxiv-multimodal-document-ai](https://arxiv.org/pdf/2611.01467v1) | 1793704292319862965, 1793704292327668581, 1793704292331971937 | 11-03 00:00 |
| `1793703903650095688` | Random Is Almost Enough? Stable and Low-Cost Benchmark Compression via BestRand | [arxiv-multimodal-document-ai](https://arxiv.org/pdf/2611.00883v1) | 1793704283134251151, 1793704283136827092, 1793704283132467698 | 11-03 00:00 |
| `1793703903649974345` | UAV-DualCog: A Dual-Cognition Benchmark for Grounded UAV Agent-World Reasoning with MLLMs | [arxiv-multimodal-document-ai](https://arxiv.org/pdf/2607.16193v2) | 1793704292763475293, 1793704292771215602, 1793704292775321658 | 11-03 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:

- `1793703903651674898`
- `1793703903650068800`
- `1793703903651004144`
- `1793703903650095688`
- `1793703903649974345`

```sh
for id in 1793703903651674898 1793703903650068800 1793703903651004144 1793703903650095688 1793703903649974345; do
  curl -s -H "Authorization: Bearer $SYNORB_KEY" "https://api.synorb.com/manifests/by-id/$id"
done
```

> **From Synorb** — Ask what Synorb covers. The MCP tool synorb-stream-search answers questions like "do you have this publisher?" without using any Manifest allowance. [Connect and ask](https://synorb.com/connect/start)

## Related stories

- [CrowdStrike buys Fiddler to bring agent behaviour under the security umbrella](https://hangingcontext.com/stories/2026-11-03/crowdstrike-buys-fiddler-agent-security/)
- [Healthcare AI Shifts From Answering Questions to Verifying Records at the Bedside](https://hangingcontext.com/stories/2026-11-03/healthcare-ai-moves-to-verifying/)
- [Below the model: browsers, agent machines, kernels and botnets set the security agenda](https://hangingcontext.com/stories/2026-11-03/attack-surface-moves-below-model/)
- [Securing agents is becoming a layer of its own: gates inside models, shields on web apps, and attackers that adapt](https://hangingcontext.com/stories/2026-11-02/securing-agents-becomes-a-layer/)

## Links

- HTML: https://hangingcontext.com/stories/2026-11-03/benchmark-scores-hide-failures/
- JSON: https://hangingcontext.com/stories/2026-11-03/benchmark-scores-hide-failures/index.json
- Markdown: https://hangingcontext.com/stories/2026-11-03/benchmark-scores-hide-failures/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/2026-11-03/benchmark-scores-hide-failures/. Terms: https://hangingcontext.com/terms/reuse/