AI

RE-SciVQA: Improving Scientific VQA With Context-Relevant Retrieval and Lightweight MLLM Adaptation

Read full article at source ↗

Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 03/11/2026, 09:03:25 UTC.

Briefings

RE-SciVQA improves Scientific Visual Question Answering by enhancing retrieved context quality and using lightweight LoRA training. On the SciVQA benchmark, it achieved a ROUGE-1 F1 score of 73.9%, significantly outperforming chart-focused models like UniChart (14.7%) and ChartGemma (6.9%). The method also improved exact match by 11.9% and fuzzy match by 29.5% compared to a standard RAG baseline. Additionally, plot digitization instead of synthetic captions yielded an 18.5% improvement for closed-ended infinite visual questions.

Key points

• RE-SciVQA achieves a ROUGE-1 F1 score of 73.9% on the SciVQA benchmark, outperforming chart-focused models such as UniChart (14.7%) and ChartGemma (6.9%).
• The method improves exact match by 11.9% and fuzzy match by 29.5% compared to a standard RAG baseline, resulting in an overall gain of 14.3%.

Context

This research addresses the challenge of interpreting scientific figures beyond main text by leveraging Large Language Models for knowledge discovery. The study focuses on improving context retrieval quality and fine-tuning small LLMs with parameter-efficient LoRA training, using a simple query expansion step and similarity score adjustments.

Limitations

The study is based on preprint arXiv:2611.00242v1, which may not be fully peer-reviewed. Results are limited to the SciVQA benchmark and closed-ended infinite visual questions; performance on other benchmarks or question types remains untested.

Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 04/11/2026, 02:22:24 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.

View archived briefing on IPFS
bafybeigzxgtguotuqvjeywyuocnlznqi5q2lho4busib7rhj2nhrchibni