HOMURA: Taming the Sand-Glass for Time-Constrained LLM Translation via Reinforcement Learning

TL;DR

The paper tackles the verbosity bias of LLMs in time-constrained translation tasks such as subtitling and dubbing. It introduces Sand-Glass, a syllable-budget benchmark, and HOMURA, a KL-regularized RL framework with a dynamic syllable-ratio reward to tightly control output length while preserving meaning. The method avoids supervised compression data and demonstrates strong improvements over strong baselines in both length compliance and semantic adequacy, across several language pairs. The contributions enable more reliable, real-time multilingual localization and offer insights into the rate-distortion frontier for neural translation under strict temporal budgets.

Abstract

Large Language Models (LLMs) have achieved remarkable strides in multilingual translation but are hindered by a systemic cross-lingual verbosity bias, rendering them unsuitable for strict time-constrained tasks like subtitling and dubbing. Current prompt-engineering approaches struggle to resolve this conflict between semantic fidelity and rigid temporal feasibility. To bridge this gap, we first introduce Sand-Glass, a benchmark specifically designed to evaluate translation under syllable-level duration constraints. Furthermore, we propose HOMURA, a reinforcement learning framework that explicitly optimizes the trade-off between semantic preservation and temporal compliance. By employing a KL-regularized objective with a novel dynamic syllable-ratio reward, HOMURA effectively "tames" the output length. Experimental results demonstrate that our method significantly outperforms strong LLM baselines, achieving precise length control that respects linguistic density hierarchies without compromising semantic adequacy.

Figures (7)

• Figure 1: Illustration of cross-lingual verbosity bias. Unlike forward metrics, the Roundtrip Expansion Ratio$\rho_{\text{rtp}}$ isolates model-induced redundancy from linguistic density shifts. A value of $\rho_{\text{rtp}} > 1.0$ indicates systemic inflation despite constant semantic content. Formal definition and diagnostic study are provided in Section \ref{['sec:pilot']}.
• Figure 2: Forward vs. backward expansion relationship (colored by round-trip ratio $\rho_{\text{rtp}}$). The red curve represents the unity baseline $\rho_{\text{rtp}}=1.0$ ($y=1/x$). Points above this curve indicate systemic model-induced length inflation. Unconstrained models show a median $\rho_{\text{rtp}} > 1.10$.
• Figure 3: Overview of our HOMURA framework. Given a source utterance $x$ and a syllable-ratio budget $c$, we optimize a KL-regularized policy $\pi_\theta(y\!\mid\!x,c)$ with GRPO. The reward combines a length-compliance term $R_{\mathrm{len}}$ and a quality term $R_{\mathrm{qual}}$ (rubric-based or reason-based), encouraging concise yet faithful translations.
• Figure 4: Quality--compression trade-off on Zh$\rightarrow$En. We plot BLEU (left) and BT-CERR (right) versus the average syllable ratio $\rho$. Each point is a model setting (marker: backbone; color: variant: Unconstrained, Rubric, GenRM, Multi-length Prompting). The red curve shows the empirical trend: quality degrades as compression strengthens.
• Figure 5: Training dynamics of syllable ratios during policy optimization (Zh$\to$En). The blue curve represents our dynamic bound relaxation, while the red curve represents static fixed bounds. Values indicate the mean syllable ratio $\rho$ across training steps.