Research map / Vision-language and multimodal / Multimodal LLMs

BalCapRL : A Balanced Framework for RL-Based MLLM Image Captioning

NeurIPS 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Multimodal LLM benchmarks · 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 BalCapRL, an RL framework for MLLM image captioning that jointly rewards utility-aware precision, reference recall, and linguistic quality, using c-GDPO (per-reward normalization) and length-conditional reward masking.

Problem

Captioning-RL methods optimize a narrow notion of caption quality (utility, correctness/completeness, or arena preference), causing biased trade-offs and regressions on other evaluation views.

Approach

Combine three continuous rewards (pointability-aware precision via atomic assertion verification, recall against reference units, and a linguistic score), normalize each reward separately before weighted aggregation (c-GDPO) to avoid summed-reward collapse, and gate the linguistic reward by a two-sided length ratio mask.

Key results

• On LLaVA-1.5-7B, BalCapRL raises DCScore from 23.0 to 36.6 (+13.6) and CapArena from -94.0 to -65.0, outperforming FEEDQUILL's DCScore of 34.5.
• On QwenVL2.5-3B, BalCapRL reaches DCScore 50.8 and CapArena -3.8 versus CapRL-3B's 48.6 and -50.6, while CapRL-3B reaches higher CaptionQA (82.6 vs 75.0).
• On QwenVL2.5-7B, BalCapRL achieves DCScore 53.4, CaptionQA 79.1, CapArena 28.5, versus RubiCap-7B at 50.5, 76.0, 22.7.

Surprising findings

• Removing the linguistic reward raises CaptionQA and DCScore above the full method but drops CapArena sharply and roughly triples caption length.
• Setting only an upper length-ratio bound is insufficient; lower-bound gating (tau_l=0.5) is needed to prevent overly short captions.

Limitations

• Pointability rubric may under-reward reasonable world-knowledge inferences not directly pointable.
• Reliance on MLLM/LLM-as-judge introduces latency and cost, and results depend on reference caption quality.
• Some caption-quality aspects are not explicitly modeled in the reward design.

Experimental setup

Datasets: ShareGPT4V DCScore reference captions CaptionQA CapArena

Benchmarks: DCScore CaptionQA CapArena b-CapScore BLINK ChartQA DocVQA InfoVQA MMBench MMStar OCRBench ScienceQA

Metrics: DCScore CaptionQA accuracy CapArena win-rate score average caption length b-CapScore (harmonic mean) Spearman correlation with human ranking

Baselines: FEEDQUILL CapRL RubiCap vanilla GRPO SFT on ShareGPT4V

Builds on: GRPO GDPO FEEDQUILL CapRL DAPO

Closest papers by meaning

• CapRL: Stimulating Dense Image Caption Capabilities via Reinforcement Learning (ICLR 2026 · similarity 0.95)
• RePIC: Reinforced Post-Training for Personalizing Multi-Modal Language Models (NeurIPS 2025 · similarity 0.90)
• Dual-Stage Value-Guided Inference with Margin-Based Reward Adjustment for Fast and Faithful VLM Captioning (NeurIPS 2025 · similarity 0.88)
• CapsFusion: Rethinking Image-Text Data at Scale (CVPR 2024 · similarity 0.87)
• Top-Down Semantic Refinement for Image Captioning (AAAI 2026 · similarity 0.87)
• Painting with Words: Elevating Detailed Image Captioning with Benchmark and Alignment Learning (ICLR 2025 · similarity 0.87)