RGBT Index

COMPOSITE PERFORMANCE · 0–100

RGB-T Overall Index

Four benchmarks. Complete results only.

GroundBench 50%RGB-Th 15%Degraded-1K 20%DV-1K 15%

RelMLLM included · click a model for details

Released and archived scores are descriptive. Cross-version ordering is not a controlled comparison; use the version selector for within-cohort views. No statistical tie groups are inferred.

Model configurations11Versions tracked separately

RGB-T benchmarks4Three capability dimensions

Tasks per full evaluation13,980Across four benchmarks

General benchmarks5Scores awaiting verification

Model Results

Percentage scores · higher is better

Complete only

		Default weights	Acc@0.5	Macro QAcc	Acc@0.5	Accuracy	Result version	Status

Overall requires four complete scores within a disclosed result cohort.Snapshot · 2026-10-16

Three Dimensions. Four Benchmarks.

Localization

GroundBench · Acc@0.5

50% of the default index

Understanding

RGB-Th + DV-1K · mean score

30% of the default index

Degraded Localization

Degraded-1K · Acc@0.5

20% of the default index

General evaluations are awaiting model-specific verification.

No scores from unrelated research adapters are assigned to released models.

01

MMStar

Multimodal understanding

02

OCRBench

Text recognition & understanding

03

MMMU

Multidisciplinary reasoning

04

MMMU-Pro

Advanced visual reasoning

05

IFEval

Text instruction following

General Capability Results

Absolute scores · no base-model disclosure required

Model	MMStar	OCRBench	MMMU	MMMU-Pro	IFEval	Status

A dash means not evaluated. Benchmark split, variant, scoring scale and generation settings must be disclosed before publication. No combined general-capability index is defined.

SCORING

A weighted benchmark index

Scoring version 1.1, effective 19 October 2026: GroundBench 50%, RGB-Th 15%, Degraded-1K 20%, DV-1K 15%. This replaces the previous 25%-each convention; benchmark scores themselves are unchanged.

Overall = 0.50 × GroundBench + 0.15 × RGB-Th + 0.20 × Degraded-1K + 0.15 × DV-1K

Missing scores are never zero-filled. Partial results receive no Overall Index. Custom weighting is a personal browser-local view, not an official ranking. Index values are only interpretable within their disclosed cohort; archived and paper averages are descriptive.

Benchmark definitions

Degraded-scene accuracy is not itself a clean-to-degraded robustness drop. That requires matched clean/degraded samples. GroundBench uses IoU ≥ 0.5; parser validity and token-limit rates are separate diagnostics.

Keep provenance visible

The current grounding runner imports the recovered RelMLLM build_grounding_prompt: RGB image first, TIR second; explicit evidence and relation assessment; single-image coordinates normalized to 0–1000; a final numeric bbox in JSON. No pair-concatenated coordinates. Greedy decoding, BF16, native processor, SDPA, no thinking, batch 1 and max 1,024 new tokens for the Qwen3.5 runs. RGB-Th uses 8 tokens and DV-1K 32 tokens, with a 10-second generation time limit. Historical settings differ and remain separate. Prompt alignment alone does not establish full runtime equivalence.

Scores are an author-provided snapshot, not automatically synchronized jobs or an independent certification. No confidence intervals are fabricated from summary scores.

General capability coverage

MMStar, OCRBench, MMMU, MMMU-Pro and IFEval are separate metrics. Dataset versions, split, MMMU-Pro variant, prompt, scorer and generation budget must be recorded. OCRBench scoring scales must be identified rather than silently treated as percentages. IFEval measures text instruction following, not multimodal instruction following.

Base-model identity is optional. A base-versus-finetuned delta is an additional analysis only when the mapping and evaluation protocol are verified. Absolute capability scores alone cannot establish forgetting.

Submit a model or result

Include a model URL, checkpoint version, benchmark split, evaluation configuration, scoring rules and full result artifacts where permitted. Submissions are reviewed manually; models are not executed by this Space.

RGB-T Model Leaderboard