BrainTRACE
Harbor adapterA report-grounded benchmark with 7,273 VQA instances from 1,778 longitudinal patients evaluating vision-language models on brain MRI evidence tracing.
- Samples
- 7,273
- Released
- 2026
- Scores
- 13
InstanceBench
InstanceBench: Diagnosing Referential Reasoning and Target Identity in Referring Expression Segmentation
An instance-centered diagnostic benchmark for Referring Expression Segmentation with 6,194 images, 9,264 targets, and 25,077 expressions evaluating referential logic consistency.
How to run
ltb · instancebench
$ npm i -g lettokenburn-cli$ ltb bench info instancebench
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 | |||
|---|---|---|---|---|---|---|
| Qwen3.8 27B | 63.5 | N/A | Independent | medium | Oct 7, 2026 | arxiv.org |
| Qwen3-VL-4B | 54.0 | N/A | Independent | medium | Oct 7, 2026 | arxiv.org |
| GLM-4.1V-9B-Thinking | 52.8 | N/A | Independent | medium | Oct 7, 2026 | arxiv.org |
| Qwen3-VL-8B-Instruct | 52.0 | N/A | Independent | medium | Oct 7, 2026 | arxiv.org |
| Qwen2.5-VL-72B-AWQ | 51.5 | N/A | Independent | medium | Oct 7, 2026 | arxiv.org |
| Qwen2.5-VL-32B-AWQ | 49.0 | N/A | Independent | medium | Oct 7, 2026 | arxiv.org |
| CogVLM-Grounding | 46.7 | N/A | Independent | medium | Oct 7, 2026 | arxiv.org |
| Qwen2.5-VL-7B | 42.2 | N/A | Independent | medium | Oct 7, 2026 | arxiv.org |
| InternVL3.5-8B-Instruct | 34.9 | N/A | Independent | medium | Oct 7, 2026 | arxiv.org |
| MiniCPM-V 4.6 | 34.8 | N/A | Independent | medium | Oct 7, 2026 | arxiv.org |
| LLaVA-OneVision-7B | 0.7 | N/A | Independent | medium | Oct 7, 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
InstanceBench comprises 6,194 images and 25,077 human-verified expressions organized into target-centric expression sets. It evaluates referential reasoning across nine logic categories using COCO and OpenImages data.
It employs identity-aware metrics to separate target selection from mask generation errors. Evaluations of 22 RES checkpoints reveal target selection as the primary bottleneck, with top models achieving 59.6% [email protected].