llava-le-eval
LLaVA-LE: Large Language-and-Vision Assistant for Lunar Exploration — Inal et al. (2026) (arXiv:2603.24696, 2026)
What this evaluates
Evaluates multimodal instruction-following and complex geological reasoning on lunar surface imagery. It probes the model's ability to interpret crater morphology, degradation states, and inferred geological processes beyond simple visual description.
Datasets
- LUCID (held-out eval set) — total 190; splits: test (190)
Metrics
average_overall_score(primary) — range: [1, 10]- Each of two independent LLM judges (ChatGPT and Gemini) assigns a score from 1 to 10 based on relevance, clarity, and accuracy. The final metric is the arithmetic mean of the two judges' scores.
Input / output format
Input: Raw lunar image and a text question.
Output: Free-form textual response.
Scoring recipe
def compute_metric(predictions, references, captions):
scores = []
for pred, ref, cap in zip(predictions, references, captions):
s1 = judge_chatgpt(cap, ref, pred) # returns 1-10
s2 = judge_gemini(cap, ref, pred) # returns 1-10
scores.append((s1 + s2) / 2)
return sum(scores) / len(scores)
Common pitfalls
- Judges do not access the raw image during scoring; evaluation is grounded solely in the scientific caption.
- Reference answers are synthetically generated by GPT-5.1 and Gemini-2.5 rather than human-annotated ground truth.
- The 1-10 scale is averaged across two different LLM judges, which may introduce model-specific bias or calibration differences.
Evidence (verbatim from paper)
Each judge assigns a score from 1 to 10 based on relevance, clarity, and accuracy. ... LLaVA-LE Stage 2 achieves an average overall score of 0.921 (averaged across GPT and Gemini judges), representing a 3.3× improvement over Base LLaVA (0.278) and a 2.1× improvement over LLaVA-LE Stage 1 (0.443).
Citation
@misc{inal2026llavale,
title={LLaVA-LE: Large Language-and-Vision Assistant for Lunar Exploration},
author={Inal et al. (2026)},
year={2026},
note={arXiv:2603.24696}
}
- arXiv: 2603.24696