Research map / Vision-language and multimodal / Human and embodied modeling
Fusing Pre-Trained Language Models With Multimodal Prompts Through Reinforcement Learning
Youngjae Yu, Jiwan Chung, Heeseung Yun, Jack Hessel, Jae Sung Park, Ximing Lu, Rowan Zellers, Prithviraj Ammanabrolu, Ronan Le Bras, Gunhee Kim, Yejin Choi
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use CLIP-reward RL alignment with an unpaired caption-domain language model | ESPER-CaptionLM · COCO test split | CIDEr: 63.6 -> 78.2 | Improved |
| Use CLIP-reward RL alignment with an unpaired caption-domain language model | ESPER-CaptionLM · COCO test split | inference time: 65s -> 0.65s | Improved |
| Initialize supervised fine-tuning with an encoder trained by ESPER RL (ablation) | ESPERInit-Full · COCO Captions test split | CIDEr: 108.4 -> 111.1 | Improved |
| Initialize encoder-only supervised fine-tuning with an encoder trained by ESPER RL (ablation) | ESPERInit-MLP · COCO Captions test split | CIDEr: 94.4 -> 103.1 | Improved |
| Apply ESPER RL multimodal alignment to a text-only commonsense model | ESPER-VCGLM · Visual Commonsense Graph validation split | BLEU-4: 9.9 -> 13.0 | Improved |
| Apply ESPER RL multimodal alignment to COMET | ESPER-COMET · COCO Captions + COMET validation split | BLEU-4: 17.8 -> 28.3 | Improved |
| Apply ESPER RL alignment to a news-domain language model | ESPER-NewsLM · VisualNews test split | CIDEr: 1.3 -> 4.6 | Improved |
| Apply ESPER RL alignment to a dialogue-domain language model | ESPER-DialogLM · VisDial validation split | R@1: 5.7 -> 14.6 | Improved |
| Use an audio reward model and unpaired audio data for multimodal alignment | ESPER-CaptionLM-Audio · AudioCaps test split | CIDEr: 6.21 -> 7.92 | Improved |
| Use an unpaired domain-specific caption language model rather than general GPT-2 | ESPER · COCO test split | CIDEr: 29.1 -> 78.2 | Improved |
| Use an unpaired domain-specific language model rather than general GPT-2 | ESPER · AudioCaps test split | CIDEr: 2.84 -> 7.92 | Improved |
| Use zero-shot ESPER-NewsLM rather than ESPERInit-MLP | ESPER · VisualNews test split | CIDEr: 15.7 -> 4.6 | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- ZeroCap — ZeroCap: Zero-Shot Image-to-Text Generation for Visual-Semantic Arithmetic (CVPR 2022) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Activating Visual Context and Commonsense Reasoning Through Masked Prediction in VLMs (AAAI 2026 · similarity 0.90)
- Perception-Aware Policy Optimization for Multimodal Reasoning (ICLR 2026 · similarity 0.89)
- Perception-R1: Advancing Multimodal Reasoning Capabilities of MLLMs via Visual Perception Reward (ICLR 2026 · similarity 0.89)
- Spotlight on Token Perception for Multimodal Reinforcement Learning (ICLR 2026 · similarity 0.89)
- Perceptual-Evidence Anchored Reinforced Learning for Multimodal Reasoning (CVPR 2026 · similarity 0.88)
- TokenSwap: Benchmarking and Reducing the Modality Gap in Multimodal LLMs (NeurIPS 2026 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
