Asking a Language Model for Diverse Responses

2025-09-22

AI Summary

This paper addresses the challenge of generating diverse yet plausible responses from large language models (LLMs) given a fixed context. To this end, it systematically compares and enhances sampling strategies, proposing two non-i.i.d. sampling methods: *enumerative sampling*—which explicitly covers logical and lexical paths—and *iterative sampling*—which dynamically adjusts generation based on prior outputs—against standard parallel (i.i.d.) sampling. All methods integrate chain-of-thought (CoT) prompting to jointly regulate logical coherence and representational diversity. Experiments under identical computational budgets show that enumerative and iterative sampling significantly improve response diversity (+23.6%–38.1% in n-gram and semantic diversity) while maintaining or slightly exceeding parallel sampling in generation quality, as measured by BLEU, BERTScore, and human evaluation. The core contribution is the empirical validation and formal characterization of non-i.i.d. sampling as an effective mechanism for balancing diversity and quality, yielding a reproducible methodological framework for controllable diverse generation.

Technology Category

• Natural Language Processing: Generation
• Search and Optimization: Sampling/Simulation-based Search
• Machine Learning: Large Multimodal Models (LMMs)

Application Category

• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• Economics, Online Markets and Human Computation: LLM based quality controls for crowd work

Abstract

Large language models increasingly rely on explicit reasoning chains and can produce multiple plausible responses for a given context. We study the candidate sampler that produces the set of plausible responses contrasting the ancestral (parallel) sampling against two alternatives: enumeration, which asks the model to produce $n$ candidates in one pass, and iterative sampling, which proposes candidates sequentially while conditioning on the currently generated response set. Under matched budgets, we compare these samplers on quality, lexical and computation flow diversity, and efficiency. Our empirical results demonstrate that enumeration and iterative strategies result in higher diversity at comparable quality. Our findings highlight the potential of simple non-independent sampling strategies to improve response diversity without sacrificing generation quality.

Problem

• Studying methods to generate diverse responses from language models
• Comparing ancestral sampling with enumeration and iterative strategies
• Evaluating diversity and quality under matched computational budgets

Innovation

• Enumeration generates multiple candidates simultaneously
• Iterative sampling conditions on existing response set
• Non-independent strategies enhance diversity without quality loss