Test-Time Scaling with Repeated Sampling Improves Multilingual Text Generation

2025-05-28

AI Summary

This paper systematically investigates the effectiveness of test-time repetition sampling in multilingual text generation. To address inconsistent performance across open-ended generation and reasoning-intensive tasks (e.g., mathematical reasoning, code generation), we propose a dual-verification mechanism: perplexity-based verification for open-ended generation and reward-driven verification—grounded in human preferences or task-specific objectives—for reasoning tasks, significantly improving reasoning performance. We conduct the first comprehensive evaluation on multilingual benchmarks Aya and m-ArenaHard, demonstrating consistent quality gains: average generation quality improves substantially, with some metrics increasing by over 35%. Our core contributions are threefold: (1) establishing the general efficacy of repetition sampling for multilingual generation; (2) identifying the critical design principle that verifier type must align with task paradigm; and (3) introducing a lightweight, scalable test-time scaling framework.

Technology Category

• Natural Language Processing: Generation
• Machine Learning: Large Multimodal Models (LMMs)
• Knowledge Representation and Reasoning: Computational Complexity of Reasoning

Application Category

• Search and Retrieval-Augmented AI: Multilingual and cross-lingual Web search
• Economics, Online Markets and Human Computation: LLM based quality controls for crowd work
• Web Mining and Content Analysis: Mining multimedia, multimodal, multilingual, cross-lingual Web data

Abstract

Inference-time scaling via repeated sampling has shown promise in reasoning tasks, but its effectiveness in multilingual generation remains underexplored. We evaluate this approach using perplexity- and reward-based verifiers on two multilingual benchmarks: the Aya Evaluation Suite and m-ArenaHard. Our results show consistent quality improvements, with gains exceeding 35% in some cases. While perplexity-based scoring is effective for open-ended prompts, only reward-based verifiers improve performance on tasks requiring reasoning (e.g., math, code). Our results demonstrate the broader utility of repeated sampling for multilingual text generation and underscore the importance of selecting right verifiers for the task.

Problem

• Evaluates repeated sampling for multilingual text generation
• Compares perplexity- and reward-based verifiers on benchmarks
• Demonstrates performance gains exceeding 35% in some cases

Innovation

• Repeated sampling enhances multilingual text generation
• Perplexity and reward verifiers improve generation quality
• Task-specific verifiers optimize reasoning performance