Research map / Vision-language and multimodal / Multimodal LLMs

Specificity-aware reinforcement learning for fine-grained open-world classification

CVPR 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Reinforcement learning for visual reasoning · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes SpeciaRL, a GRPO-based RL method with a dynamic, judge-verified reward anchored to the best rollout's specificity level, to make reasoning LMMs more specific in open-world fine-grained classification without losing correctness.

Problem

Reasoning large multimodal models produce overly generic labels in open-world fine-grained image classification, and naively promoting specificity (prompting, SFT, or static-reward RL) reduces correctness.

Approach

A sample-wise dynamic reward: each rollout receives reward 1 if its LLM-judged category is at least as specific as the best category achieved among the online GRPO rollouts for that sample (anchored to model capability), with categories ordered W < A < G < S- < S < S+.

Key results

• On the fine-grained set, SpeciaRL-7B reaches HM 0.883 (spec 0.920, corr 0.848), above Qwen2.5VL-7B base (HM 0.790) and rft (HM 0.825).
• On the very fine-grained set, SpeciaRL-7B achieves HM 0.830, second to rft's 0.821 only in ordering; it is best overall HM among methods listed in the main table (BoN-64 is 0.868 as reference).
• Following the external evaluation protocol, SpeciaRL-7B achieves best TI, SS and CS on fine-grained and best TI, SS, CS on very fine-grained, state-of-the-art on three of four metrics per group.

Surprising findings

• Binary static reward (rewarding only Specific/More Specific) performs worst among the static reward variants, indicating value in rewarding less-informative correct predictions.
• Increasing online rollouts to N=15 lowers specificity and correctness relative to N=5 and N=10.

Limitations

• Training uses only a single domain (CUB birds), so generalization is evaluated out-of-domain on a limited set of datasets.
• The reward depends on an LLM judge, and occasional wrong predictions or unusual scientific names for generic concepts appear as failure cases.
• On the very fine-grained set SpeciaRL is not best on every metric (e.g., LI is lower than SFT).

Experimental setup

Datasets: CUB

Benchmarks: Flowers102 Food101 OxfordPets FGVCAircraft StanfordCars CUB (in-domain)

Metrics: specificity correctness harmonic mean TI LI SS CS

Baselines: CaSED Qwen2.5VL-7B (Be specific prompt) Qwen2.5VL-7B (sft) Qwen2.5VL-7B (rft) InternVL2.5-8B Qwen2.5VL-3B InternVL2.5-4B

Builds on: GRPO Visual-RFT RLVR CaSED

Closest papers by meaning

• Reason-RFT: Reinforcement Fine-Tuning for Visual Reasoning of Vision Language Models (NeurIPS 2025 · similarity 0.88)
• DiVE-k: DIFFERENTIAL VISUAL REASONING FOR FINE-GRAINED IMAGE RECOGNITION (ICLR 2026 · similarity 0.88)
• SAM-R1: Leveraging SAM for Reward Feedback in Multimodal Segmentation via Reinforcement Learning (NeurIPS 2025 · similarity 0.88)
• RewardMap: Tackling Sparse Rewards in Fine-grained Visual Reasoning via Multi-Stage Reinforcement Learning (ICLR 2026 · similarity 0.87)
• PeRL: Permutation-Enhanced Reinforcement Learning for Interleaved Vision-Language Reasoning (NeurIPS 2025 · similarity 0.87)
• Activating Visual Context and Commonsense Reasoning Through Masked Prediction in VLMs (AAAI 2026 · similarity 0.87)