AgentHorizon
Harbor: pendingA benchmark of 1,373 computer-use tasks with paired instructions to evaluate agentic judges on long-horizon trajectory success verification across three OSs.
- Samples
- 1,373
- Released
- 2026
- Scores
- 10
VisualTrust
VisualTrust: Video Reasoning under On-Screen Claim Framing
A counterfactual benchmark with 5,063 cases measuring video MLLMs' tendency to trust misleading on-screen text claims over visual evidence.
How to run
ltb · visualtrust
$ npm i -g lettokenburn-cli$ ltb bench info visualtrust
External harness. No Harbor adapter has been generated for this benchmark yet. `ltb bench info` prints the known links.
Runs happen on your machine with your own subscription; only the score summary is pushed. CLI docs
Scores
11 sourced entries| Harness | Source | Confidence | Link | |||
|---|---|---|---|---|---|---|
| Seed 2.1 Pro | 93.4 | N/A | Independent | medium | Oct 15, 2026 | arxiv.org |
| Qwen3.5 397B A17B | 93.1 | N/A | Independent | medium | Oct 15, 2026 | arxiv.org |
| Qwen3.5 122B A10B | 92.8 | N/A | Independent | medium | Oct 15, 2026 | arxiv.org |
| Qwen3.5 Plus | 92.7 | N/A | Independent | medium | Oct 15, 2026 | arxiv.org |
| Gemini-3.5-flash | 92.5 | N/A | Independent | medium | Oct 15, 2026 | arxiv.org |
| qwen3.5-27B | 92.5 | N/A | Independent | medium | Oct 15, 2026 | arxiv.org |
| Qwen3.5 35B A3B | 91.3 | N/A | Independent | medium | Oct 15, 2026 | arxiv.org |
| Gemini 3.1 Pro | 90.9 | N/A | Independent | medium | Oct 15, 2026 | arxiv.org |
| qwen3.5-9B | 89.1 | N/A | Independent | medium | Oct 15, 2026 | arxiv.org |
| google/gemma-4-31B-it | 83.9 | N/A | Independent | medium | Oct 15, 2026 | arxiv.org |
| google/gemma-4-26B-A4B-it | 78.6 | N/A | Independent | medium | Oct 15, 2026 | arxiv.org |
Every number links to where it was published. Low-confidence rows are muted and need per-cell verification.
Reproduced by the community
All runsLoading community runs…
About
VisualTrust pairs questions with three video versions (no claim, accurate claim, misleading claim) across 14 claim-fact classes. It isolates 'claim deference' by varying only the rendered text.
Scoring uses accuracy and trap-answer selection rates. Models' accuracy drops from 90.1% to 51.9% with misleading claims, showing strong susceptibility to textual framing despite high baseline performance.