Scaling Multimodal Search and Recommendation with Small Language Models via Upside-Down Reinforcement Learning
The research investigates scaling small language models (SLMs) to handle multimodal search and recommendation tasks efficiently. By combining synthetic data distillation from Llama-3 with upside-down reinforcement learning, researchers trained a 100M-parameter GPT-2 model. This lightweight SLM achieved relevance and diversity scores comparable to much larger models, proving that small models can effectively drive multimodal discovery while significantly reducing latency and memory needs.