| name | tower-plus-multilingual |
| title | Tower+: Bridging Generality and Translation Specialization in Multilingual LLMs |
| version | 0.0.2 |
| engine | skillxiv-v0.0.2-claude-opus-4.6 |
| license | MIT |
| url | https://arxiv.org/abs/2506.17080 |
| keywords | ["Multilingual","Machine Translation","Post-training Pipeline","Preference Optimization","Reinforcement Learning"] |
| description | Build multilingual LLMs that excel at machine translation while preserving general-purpose capabilities through a staged training pipeline combining continued pretraining, SFT, preference optimization, and RL with verifiable rewards. |
Tower+: Balanced Translation and General Capabilities in Multilingual Models
The core challenge in multilingual model development is that specialist translation models often sacrifice general-purpose reasoning and chat abilities, while general-purpose models underperform on translation tasks. This creates a false choice: either have a translation powerhouse that struggles with general tasks, or a generalist model that performs mediocrely at translation. Tower+ demonstrates this choice is unnecessary by systematically balancing both capabilities through an integrated four-stage post-training pipeline.
Current open-weight multilingual models typically excel in one domain at the expense of others. Tower+ achieves a Pareto frontier between specialization and generality by carefully composing multilingual signals during training while preserving chat abilities through reward-based filtering and RL optimization. The result is a model family (2B, 9B, 72B parameters) that matches frontier proprietary models in translation while exceeding open-weight competitors on general benchmarks.
Core Concept
Tower+ uses a staged post-training approach that treats translation capability and general understanding as complementary goals rather than competing objectives. The key insight is that you can simultaneously improve specialized performance and maintain general capabilities by:
- Loading the model with diverse multilingual signals during continued pretraining
- Selecting training examples through reward-based filtering rather than naive mixing
- Using preference optimization to balance multiple objectives
- Applying reinforcement learning with specialized rewards that preserve instruction-following accuracy
This approach avoids the common pitfall where translation-focused training causes "catastrophic forgetting" of general capabilities—a problem documented in math reasoning transfer studies where SFT-based approaches often degrade on non-specialized tasks.
Architecture Overview
The Tower+ post-training pipeline consists of four sequential stages:
- Continued Pretraining (CPT): 66% monolingual data, 33% parallel translation pairs, plus 1% instruction-following data across 27 languages. This stage grounds the model in diverse linguistic patterns before specialization.
- Supervised Fine-tuning (SFT): 1.3M curated samples where only 22% are translation tasks; the remainder are general instruction-following examples. Responses are selected from multiple teacher models using reward-based filtering to ensure high quality.
- Preference Optimization: Weighted Preference Optimization (WPO) combining SFT prompt data with UltraFeedback data, using trained reward models to guide preference selection.
- Reinforcement Learning: GRPO (Group Relative Policy Optimization) with dual verifiable rewards—one for translation quality and one for instruction-following precision—to maintain both capabilities.
