Research map / Label-efficient and robust learning / Zero-/few-shot and meta-learning
Incentivizing Generative Zero-Shot Learning via Outcome-Reward Reinforcement Learning with Visual Cues
Wenjin Hou, Xiaoxiao Sun, Hehe Fan
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 |
|---|---|---|---|
| add outcome-reward RL (ablation) | RLVC · AWA2 | CZSL Acc and GZSL H (%): 79.4/73.9 -> 84.0/80.4 | Improved |
| add visual cues with prototype-distillation loss (ablation) | RLVC · AWA2 | CZSL Acc (%): 74.9 -> 84.0 | Improved |
| fine-tune the visual encoder with the reward model (ablation) | RLVC · CUB, SUN, AWA2 | GZSL H (%): H +3.7%, +1.2%, +4.1% | Improved |
| use EMA-baseline advantage instead of raw reward (ablation) | RLVC · AWA2 | GZSL H (%): 78.2 -> 80.4 | Improved |
| add RL and visual cues to vanilla diffusion-GAN generator (attribute vectors) (ablation) | Vanilla model · CUB | GZSL H (%): 75.1 -> 81.2 | Improved |
| add RL and visual cues to vanilla generator (class-name word embeddings) (ablation) | Vanilla model · CUB, SUN, AWA2 | CZSL Acc (%): +1.0, +0.1, +1.4 | Improved |
| use prototype-distillation loss instead of KL (ablation) | RLVC · CUB, SUN, AWA2 | CZSL Acc (%): CUB 88.9 -> 90.1; AWA2 76.4 -> 84.0 | Improved |
| use prototype-distillation loss instead of l1 (ablation) | RLVC · AWA2 | CZSL Acc (%): 75.9 -> 84.0 | Improved |
| use prototype-distillation loss instead of KL (ablation) | RLVC · SUN | GZSL H (%): 58.2 -> 57.6 | Hurt |
| start RL after 30 cold-start epochs (ablation) | RLVC · CUB | accuracy | Improved |
| set visual loss coefficient to 20 (ablation) | RLVC · CUB | accuracy | Improved |
| synthesize 400 unseen samples per class (ablation) | RLVC · CUB | GZSL H (%) | Improved |
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.
- PSVMA+ — PSVMA+: Exploring Multi-Granularity Semantic-Visual Adaption for Generalized Zero-Shot Learning (TPAMI 2024) · 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
- Zero-Shot Reinforcement Learning from Low Quality Data (NeurIPS 2024 · similarity 0.87)
- Vision-Language Models are Zero-Shot Reward Models for Reinforcement Learning (ICLR 2024 · similarity 0.87)
- Code as Reward: Empowering Reinforcement Learning with VLMs (ICML 2024 · similarity 0.87)
- Vision-Zero: Scalable VLM Self-Improvement via Strategic Gamified Self-Play (ICLR 2026 · similarity 0.86)
- GenZSL: Generative Zero-Shot Learning Via Inductive Variational Autoencoder (ICML 2025 · similarity 0.86)
- MSRL: Scaling Generative Multimodal Reward Modeling via Multi-Stage Reinforcement Learning (CVPR 2026 · similarity 0.86)
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.
