Reproduce “Out-of-Distribution Detection and Selective Generation for Conditional Language Models” on OpenResearch

About this paper

Jie Ren, Jiaming Luo, Yao Zhao, Kundan Krishna, Mohammad Saleh, Balaji Lakshminarayanan, Peter J. Liu

The authors introduce selective generation for encoder–decoder Transformers, using Relative Mahalanobis Distance (RMD) on input or generated-output representations alongside perplexity to rank examples by quality. In pooled in-domain and shifted tests, combined scores correlate more strongly with human summary ratings (Kendall’s τ 0.415 versus 0.300) and translation BLEURT (0.361 versus 0.286) than perplexity alone; this ranks quality, not guarantees it.

Key insights

• Perplexity, a common metric for language models, is an unreliable indicator for out-of-distribution (OOD) detection and quality prediction under distribution shifts in conditional language models (CLMs).
• OOD scores derived from the CLM's internal encoder and decoder embeddings (e.g., Relative Mahalanobis Distance or Binary Logits) are highly effective and computationally lightweight for detecting OOD inputs.
• OOD scores and perplexity capture complementary aspects of uncertainty, and their combination significantly improves the prediction of generated text quality across in-domain and OOD examples.

Summary by alphaXiv.