Multi-Novelty: Improve the Diversity and Novelty of Contents Generated by Large Language Models via inference-time Multi-Views Brainstorming

Summary

Model-agnostic inference-time brainstorming: GPT-4o text views and crawled-image descriptions (Qwen-2VL then GPT-4o-mini rewrite) prepend extra perspectives before generation. DNC eval (diversity, novelty, correctness) on 10 prompts × 100–2000 samples (909,500 answers). Text views lift GPT-4o novelty ~5–9× (SBERT) and DeepSeek-R1 ~2×; correctness often drops (GPT-4o 99.6%→92.6% text / 94.6% image). No official code on abs.

Keywords

multi-novelty · diversity · brainstorming · decoding · nus · gpt-4o · deepseek-r1

Topics

decoding, diversity, novelty, inference-time methods

Research notes

• Primary: arxiv abs (cs.CL). License CC BY 4.0 on HTML at check. NUS; correspondence alagzian@visitor.nus.edu.sg and {srinu_pd, dianbo}@nus.edu.sg. No official code on abs. HF has no paper page (API 404). Discord posted AlphaXiv 2502.12700. 909kPR is an internal generated-answer pool, not a hosted corpus, so no datasets_local row.