LetTokenBurn
Sign in
Verified86%Harbor: pendingMultimodalAgentic / Tool Use

EviLens

EviLens: A Human-Verified Benchmark for Perception Under Insufficient Evidence

A human-verified benchmark of 688 instances across five perception categories (segmentation, counting, localization, recognition, spot-the-difference) designed to evaluate agents' ability to seek evidence beyond a single glance.

PaperRepoDataset

How to run

ltb · evilens
$ npm i -g lettokenburn-cli
$ ltb bench info evilens

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

13 sourced entries
HarnessSourceConfidenceLink
Gemini-3.5-flash61.8N/AIndependentmediumSep 30, 2026arxiv.org
Seed-2.1-Turbo60.5N/AIndependentmediumSep 30, 2026arxiv.org
GPT-5.6 Sol51.3N/AIndependentmediumSep 30, 2026arxiv.org
EviRover50.0N/AIndependentmediumSep 30, 2026arxiv.org
Seed-1.848.0N/AIndependentmediumSep 30, 2026arxiv.org
Qwen3.7 Plus46.7N/AIndependentmediumSep 30, 2026arxiv.org
GPT-5.6-terra43.4N/AIndependentmediumSep 30, 2026arxiv.org
MiniMax-M337.8N/AIndependentmediumSep 30, 2026arxiv.org
GPT-5.6 Luna35.5N/AIndependentmediumSep 30, 2026arxiv.org
OneThinker-8B25.6N/AIndependentmediumSep 30, 2026arxiv.org
InternVL-3.5-8B20.5N/AIndependentmediumSep 30, 2026arxiv.org
Qwen3VL-4B-Inst19.2N/AIndependentmediumSep 30, 2026arxiv.org
Perception-R117.3N/AIndependentmediumSep 30, 2026arxiv.org

Every number links to where it was published. Low-confidence rows are muted and need per-cell verification.

Reproduced by the community

All runs
Loading community runs…

About

Content & Construction: EviLens addresses 'perception under insufficient evidence,' where fine-grained details or external knowledge are required. It comprises 688 instances spanning segmentation, counting, and grounding tasks (localization, recognition, spot-the-difference). The dataset was constructed using two pipelines: one collecting high-resolution images with small targets (<0.1% area) requiring multi-step visual exploration, and another synthesizing group photographs of public figures/anime characters to test external knowledge retrieval.

Scoring & Findings: Models are evaluated on their ability to resolve perceptual queries through interaction (searching within or beyond the image). The benchmark shows that EviRover outperforms its backbone by 30 points on average, reaching performance comparable to advanced proprietary models like Seed-2.1-turbo and Gemini-3.5-Flash.