RubiCap: Rubric-Guided Reinforcement Learning for Dense Image Captioning

Dimensions overall score 8.0

GitHub Code Pulse

No public code linked for this paper yet.

• RubiCap achieves the highest win rates on CapArena benchmark compared to supervised distillation, prior RL methods, human-expert annotations, and GPT-4V-augmented outputs.result 90%
  Evidencepartial

  RubiCap achieves the highest win rates on CapArena, outperforming supervised distillation, prior RL methods, human-expert annotations, and GPT-4V-augmented outputs.

  Implicationpartial

  Directly stated in abstract with clear comparative results

  Verificationpartial

  partial

• RubiCap demonstrates superior word efficiency on CaptionQA benchmark, with the 7B model matching Qwen2.5-VL-32B-Instruct performance.result 90%
  Evidencepartial

  On CaptionQA, it demonstrates superior word efficiency: our 7B model matches Qwen2.5-VL-32B-Instruct

  Implicationpartial

  Directly stated in abstract with specific model comparisons

  Verificationpartial

  partial

• RubiCap-3B model surpasses the 7B counterpart of Qwen2.5-VL on CaptionQA benchmark.result 90%
  Evidencepartial

  our 3B model surpasses its 7B counterpart

  Implicationpartial

  Directly stated in abstract with clear model size comparison

  Verificationpartial

  partial

• Using RubiCap-3B as a captioner produces stronger pretrained vision-language models than those trained on captions from proprietary models.result 85%
  Evidencepartial

  using the compact RubiCap-3B as a captioner produces stronger pretrained VLMs than those trained on captions from proprietary models

  Implicationpartial

  Directly stated in abstract with comparative claim about model performance

  Verificationpartial

  partial

• RubiCap uses LLM-written rubrics to derive fine-grained, sample-specific reward signals for reinforcement learning in dense image captioning.method 90%
  Evidencepartial

  RubiCap, a novel RL framework that derives fine-grained, sample-specific reward signals from LLM-written rubrics

  Implicationpartial

  Directly stated in abstract describing the core method

  Verificationpartial

  partial

• RubiCap replaces coarse scalar rewards with structured, multi-faceted evaluations by decomposing holistic quality assessment.technical 85%
  Evidencepartial

  enabling an LLM judge to decompose holistic quality assessment and replace coarse scalar rewards with structured, multi-faceted evaluations

  Implicationpartial

  Directly stated in abstract describing technical approach

  Verificationpartial

  partial

• Supervised distillation for synthetic captioning via vision-language models yields limited output diversity and weak generalization.limitation 80%
  Evidencepartial

  supervised distillation often yields limited output diversity and weak generalization

  Implicationpartial

  Directly stated in abstract as motivation for the work

  Verificationpartial

  partial

• Reinforcement learning successes in captioning have been limited by the lack of deterministic checkers available in verifiable domains.limitation 80%
  Evidencepartial

  reinforcement learning (RL) could overcome these limitations, but its successes have so far been concentrated in verifiable domains that rely on deterministic checkers -- a luxury not available in open-ended captioning

  Implicationpartial

  Directly stated in abstract as problem statement

  Verificationpartial

  partial

Startup potential card