MoDA Lifts LLM Output Diversity 265% Without Hurting Capability

A training recipe called MoDA lifts language-model output diversity by 265% on the Infinite-Chat benchmark while also improving general-capability pass@1 by 10.3% over a Qwen3-8B baseline, according to a paper posted to arxiv on 14 September 2026.

'A notable byproduct of LLM alignment training is mode collapse: the progressive loss of output diversity that narrows a model's expressivity at inference time,' the authors write. The team, made up of Jiayi Yuan, Hangoo Kang, James Jihao Liu, Yejin Choi, Vikram Iyer, Liwei Jiang and Natasha Jaques, trains a single policy conditioned on numbered abstract roles that each act as an agent competing to generate distinct outputs. They describe the approach as one that 'jointly optimizes generation quality and diversity, inspired by the coordination perspective in multi-agent reinforcement learning.' A prompt-adaptive quality gating mechanism calibrates a reference quality threshold so that widening the output distribution does not collapse into low-quality generations.

Against DivPO, a prior diversity-oriented baseline, MoDA raises SBERT diversity from 0.274 to 0.482 and E-Vendi from 2.86 to 4.4, with a 7.0% pass@1 gain. Two researchers on our radar posted the arxiv link the same week it appeared. The evaluation spans seven general-capability tasks and four domain-specific ones in scientific ideation and creative writing.

The abstract publishes no compute-cost figures against the baselines it beats, and stays silent on whether the gains carry to models larger than the 8B tested.