CapRL: Stimulating Dense Image Caption Capabilities via Reinforcement Learning

Kari Jaaskelainen

29 Sep 2025 — 1 min read

Objective rewards for a subjective task: captions judged by how well they enable QA.

Supervised captioning often memorizes ground-truth phrasing and struggles to generalize. CapRL brings Reinforcement Learning with Verifiable Rewards (RLVR) to image captioning by redefining “good” captions in terms of utility: can a non-visual LLM answer questions about the image using the caption alone?

The two-stage pipeline first has an LVLM generate a caption, then computes an objective reward from a separate, vision-free LLM answering multiple-choice questions based solely on that caption. This decoupled setup turns a subjective task into one with verifiable signals, encouraging captions that are informative, precise, and semantically rich—without overfitting to a single phrasing.

Pretraining on CapRL-5M (captions produced by a 3B model) yields substantial gains across 12 benchmarks. Within the Prism framework for caption quality evaluation, CapRL matches the performance of a much larger Qwen2.5-VL-72B and exceeds the baseline by an average of 8.4%. The approach improves open-ended descriptiveness while maintaining factual grounding.

Why it matters: stronger captions amplify downstream utility—from retrieval and VQA to accessibility and content moderation. By rewarding captions that enable reliable QA, CapRL aligns training with practical goals rather than surface form.

Future work may refine reward design (e.g., counterfactual QA, hallucination penalties) and extend to dense region-level descriptions and video.

Paper: http://arxiv.org/abs/2509.22647v1

Register: https://www.AiFeta.com

#AI #ComputerVision #Captioning #RL #LVLM #Evaluation #VQA

From f(x) and g(x) to f(g(x)): LLMs Learn New Skills in RL by Composing Old Ones

Evidence that RL teaches genuinely new abilities: compositional skills emerge and transfer across tasks Does RL merely reweight what an LLM already knows—or can it teach genuinely new skills? This paper offers concrete evidence for the latter. Using a controlled, synthetic framework, the authors define “skills” as string transformation

OpenGPT-4o-Image: A Comprehensive Dataset for Advanced Image Generation and Editing

A structured 80k instruction–image corpus spanning 11 domains and 51 subtasks to train unified visual editors Unified models for image generation and editing hit a data ceiling: existing corpora emphasize basic manipulations but miss real‑world complexity. OpenGPT‑4o‑Image tackles this with a hierarchical task taxonomy and automated

Random Policy Valuation is Enough for LLM Reasoning with Verifiable Rewards

ROVER replaces PPO loops with uniform‑policy Q‑values—boosting quality and diversity in math reasoning Popular RLVR methods for LLM reasoning lean on generalized policy iteration (e.g., PPO/GRPO), but suffer instability and diversity collapse. This paper reframes math RLVR as a specialized finite‑horizon MDP with deterministic

CLPO: Curriculum Learning meets Policy Optimization for LLM Reasoning

A dynamic, self‑paced curriculum that restructures problems to match model ability in RLVR Online RL with Verifiable Rewards (RLVR) has boosted LLM reasoning—but most methods treat all problems equally, wasting effort on solved items and flailing on those beyond current capability. CLPO fixes that with a dynamic pedagogy:

Read more

From f(x) and g(x) to f(g(x)): LLMs Learn New Skills in RL by Composing Old Ones

OpenGPT-4o-Image: A Comprehensive Dataset for Advanced Image Generation and Editing

Random Policy Valuation is Enough for LLM Reasoning with Verifiable Rewards

CLPO: Curriculum Learning meets Policy Optimization for LLM Reasoning