BioReason: Retrieve-Then-Reason MLLMs for Interpretable Fine-Grained Species Classification
Year: 2026
Type: article
Abstract: Fine-grained species classification involves distinguishing visually similar taxa and remains challenging for general-purpose visual recognition models. We propose BioReason, a retrieve-then-reasonframework that uses BioCLIP candidates to formulate a multiple-choice task for Qwen2.5-VL-7B, followed by two-source supervised fine-tuning (SFT) and group relative policy optimization (GRPO). We clarify that the implemented evaluation retrieves the Top-5 labels from each benchmark-specific class vocabulary, rather than directly from the complete 454K taxonomy. When the ground-truth label is absent, it is inserted into the candidate set; consequently, the reported MLLM accuracy measures conditional candidate reranking and is not an end-to-end open-world zero-shot result. The executed SFT uses 31,023 caption samples and 4929 reasoning samples, while GRPO combines binary format and answer-accuracy rewards. Across nine benchmarks, SFT(A+B) obtains 40.81% macro accuracy and the subsequent GRPO checkpoint obtains 40.33%; GRPO improves the five-subset Camera Trap mean from 46.06% to 46.42% but does not yield a consistent overall gain. Generated responses follow the requested – structure in 99.95% of cases, although this syntactic measure does not establish biological correctness or faithfulness of the explanations.
Source: Electronics
Institutions: Northeast Forestry University, Lanzhou University
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: gold
Estimated APC paid: $2,896
OpenAlex ID: W7221663328
FWCI:
Cites:
Cited by:
Related to: