HOMURA: Reinforcement Learning for Time-Constrained LLM Translation (EMNLP 2026 Oral)

Overview

Large language models (LLMs) have advanced multilingual translation but suffer from a systematic "cross-lingual verbosity bias," producing translations that are too long for time-budgeted scenarios such as subtitling and video dubbing. Prompt engineering alone cannot balance semantic fidelity with strict length constraints. This work introduces the Sand-Glass benchmark for syllable-level length-constrained evaluation and proposes HOMURA , a reinforcement learning framework that combines a KL-regularized objective with a novel "dynamic syllable-ratio reward" to explicitly optimize the trade-off between semantic preservation and time compliance. Experiments show HOMURA precisely controls output length while respecting cross-lingual density differences, significantly outperforming mainstream LLM baselines. HOMURA is now deployed at scale in Bilibili's video voice-translation pipeline.

Paper: https://arxiv.org/pdf/2601.10187 Online Demo: https://index-translate.bilibili.com/ GitHub: https://github.com/bilibili/Index-Translate Hugging Face: https://huggingface.co/collections/IndexTeam/index-translate Free API:

https://github.com/bilibili/Index-Translate#option-1-free-online-api-zero-gpu-setup

Index-Translate technical report: https://arxiv.org/abs/2609.40181 HOMURA accepted as EMNLP 2026 Oral!

Background and Challenges

LLMs exhibit exceptional semantic accuracy in neural machine translation, yet they display a systematic cross-lingual verbosity bias . As illustrated, the proposed Roundtrip Expansion Ratio reveals that LLM-generated translations are significantly longer than source utterances even when semantics are preserved. Traditional metrics overlook this, but in time-constrained scenarios (subtitles, dubbing, simultaneous interpretation) translation length forms a rigid "time budget" determining synchronization with video or audio.

From an information-theoretic view, the time budget acts as a strict rate constraint. The model must compress maximal information density into minimal length via lexical consolidation and syntactic restructuring. However, LLMs struggle to achieve this balance spontaneously due to language density differences (e.g., German, Spanish naturally occupy more space than Chinese) and, more fundamentally, because alignment reward models exhibit a "length bias" favoring longer, more detailed responses. Taming model length without losing core semantics is a key modeling challenge for multilingual localization.

Core contributions:

Diagnostic analysis: Systematically quantifies LLM translation verbosity bias and proposes cross-lingual diagnostic metrics including roundtrip expansion ratio.

Benchmark construction: Introduces the Sand-Glass dataset, using syllable-level budgets to simulate real-world dubbing and subtitle constraints.

Method innovation: Proposes the HOMURA RL framework that effectively balances strict length constraints with semantic preservation.

Sand-Glass Benchmark and Verbosity Analysis

3.1 Sand-Glass Benchmark and Syllable Metrics

Sand-Glass is a translation benchmark sourced from real video corpora. Unlike general translation datasets, it adopts syllables as a proxy for time.

Iso-Information Principle: Research shows that while syllable rates vary widely across languages (e.g., Spanish very fast, Chinese slower), human information transmission rate is constant at ~39 bits/s.

Language density differences: Chinese information density (0.94) far exceeds Spanish (0.63), meaning Chinese-to-Spanish translation naturally requires ~1.49× length expansion to preserve semantics.

3.2 Diagnostic Metric: Roundtrip Expansion Ratio

Traditional "forward expansion ratio" cannot distinguish whether length increase stems from inherent language density differences or model verbosity. The paper proposes Roundtrip Expansion Ratio (ρ_rtp) :

Translate source to target.

Back-translate target to source language.

Compute expansion ratio.

Conclusion: Ideally, same-language roundtrip expansion ratio should approach 1.0. Tests on Sand-Glass show mainstream models' ρ_rtp universally >1, proving LLMs have a systematic, language-agnostic verbosity tendency .

3.3 Calibration Target Intervals

Comparing "theoretical expansion ratio" with "actual LLM expansion ratio" reveals LLMs universally exhibit 30%-40% extra redundancy (bias). HOMURA sets strict target syllable intervals:

English (En): [0.8, 0.9] (forces summarization-style compression).

German (De): [0.9, 1.0].

Spanish (Es): Reduces original 1.84× syllable ratio to [1.0, 1.1].

This forces models to achieve maximal information alignment within extremely short time budgets via rewriting rather than simple truncation.

HOMURA Framework Design

The task is modeled as controllable text generation: maximize translation quality under strict syllable-ratio constraints.

4.1 KL-Regularized Objective with GRPO Optimization

HOMURA uses a KL-regularized RL objective. A reward function R(x,c,y) encourages length compliance while maximizing generation quality.

Optimization algorithm: GRPO (Group Relative Policy Optimization) . No extra critic network; computes advantage via in-group normalization, greatly reducing compute overhead, suitable for sequence-level reward optimization.

4.2 Reward Mechanism Design

The reward combines length compliance reward and quality reward .

1. Dynamic Syllable-Ratio Reward:

Dynamic interval setting: Base target intervals [L₀, R₀] set per language density tiers.

Short-sentence smoothing: For short sentences (few syllables), small length changes cause large ratio fluctuations (discrete noise). A dynamic scaling factor moderately relaxes the lower bound for short sentences.

Smooth reward function: Uses squared exponential decay instead of binary 0/1 judgment, providing continuous gradient guidance for gradual convergence to target interval.

2. Quality Reward: Ensures compression does not cause semantic loss or broken sentences. Two evaluation paths:

Rubric-based evaluation:

Semantic fidelity: Uses back-translation mechanism; re-translates generated translation back to source language, measures embedding similarity with original sentence to gauge semantic retention.

Linguistic fluency: Uses LLM-as-judge to score grammar, readability, and language consistency.

Reason-based evaluation (GenRM): Trains a generative reward model (GenRM) that produces chain-of-thought reasoning. It internally performs back-translation and semantic comparison, then judges fluency, finally outputting an interpretable binary reward signal. This "reason-then-score" approach significantly improves reward robustness.

Experimental Results and Analysis

Evaluated on Sand-Glass benchmark against GPT-5, Claude-4.1, Gemini-pro-2.5 and other mainstream models.

5.1 Length Compliance and Quality Trade-off

Precise length control: Compared to prompt-only models, HOMURA achieves a much higher proportion of translations within target intervals. Across English, German, Spanish, HOMURA's over-length rate is far lower than baselines.

Semantic density improvement: Introduces semantic density metric measuring semantic value per syllable. HOMURA maintains extremely high semantic retention while compressing length, proving it uses semantic consolidation rather than simple deletion to achieve higher information density.

5.2 Behavioral Analysis During RL Training

Monitoring training reveals significant policy evolution:

Breaking verbosity prior (beta parameter effect): When KL penalty coefficient β > 0, models fall into a "conservative trap," reluctant to aggressively restructure sentences. With β = 0, models fully escape the base model's verbosity tendency, learning more aggressive and efficient expression.

Gradient norm and optimization momentum: Curves show stable gradient updates, effectively moving toward the rate-distortion frontier.

5.3 Discovery of the Compression Wall

Probed the physical limits of semantic compression via automated compression search.

Experiments find an "effective compression threshold" for each language pair (e.g., ~0.49 for Chinese-English). Below this threshold, semantic accuracy collapses non-linearly. HOMURA's advantage lies in approaching this limit while maintaining fluency.

5.4 Quality-Compression Trade-off Analysis

Analyzes translation quality vs output length trade-off from a rate-distortion perspective.

Baseline limitations: Unconstrained baselines cluster in "high quality, long text" region. Prompting LLMs to shorten translations causes quality to plummet with increasing compression, showing ordinary LLMs are inefficient at lowering syllable rate and struggle to identify core semantic information under extreme budgets.

HOMURA's efficient trade-off: HOMURA demonstrates significantly higher rate-distortion efficiency. At same syllable budget, HOMURA's semantic fidelity significantly outperforms baselines; equivalently, at same quality level, HOMURA achieves shorter translations.

Conclusion

This work addresses the systematic cross-lingual verbosity bias in LLM translation, the main barrier for time-constrained tasks like subtitling and dubbing.

Benchmark construction: Introduces Sand-Glass benchmark based on Iso-Information Principle, using syllable budgets to simulate real-world time constraints, providing a scientific tool for quantifying translation "time feasibility."

Method breakthrough: Proposes HOMURA RL framework. Combining KL-regularized objective with dynamic syllable-ratio reward, it achieves strict length compliance without reference translations and significantly improves information density per unit length.

The study not only achieves precise length control in engineering but also theoretically delineates semantic compression boundaries (e.g., 0.49 feasibility threshold for Chinese-English). Future work will verify compression boundary universality across more language pairs and extend HOMURA to end-to-end speech translation, exploring joint semantic-duration optimization.

Bilibili Deployment and EMNLP 2026 Oral

Leveraging HOMURA's precise duration control, the system generates translations strictly aligned with original video timelines, fundamentally solving the industry pain point of audio-visual desynchronization or abnormal speech rates in machine dubbing, delivering a native-like immersive experience. Demo:

https://www.bilibili.com/video/BV111BeBPEaD/?spm_id_from=333.337.search-card.all.click&vd_source=e946351224816c78650f22c3d26579ad

.

HOMURA accepted as EMNLP 2026 Oral. Oral presentation scheduled for October 27, 2026, 09:00–10:30 in Machine Translation and Semantics session, room F9-10 . Also presenting at Poster Session D (Oct 26, 11:00–12:30, Hall H) and WMT poster session. Team swag available at venue.

EMNLP program: https://2026.emnlp.org/program/ WMT program:

https://www2.statmt.org/wmt26/program.html