LatentMT: Machine Translation with Latent Reasoning

Traditional machine translation (MT) models typically scale by increasing parameter count, which necessitates massive, resource-intensive deployments. This paper investigates an alternative scaling path: LatentMT. The authors ask whether a compact, 2.6B-parameter 'looped' language model (LoopLM) can achieve high-quality translation by performing recurrent computation within its hidden states, rather than relying on explicit chain-of-thought tokens or massive model scale.

The authors utilize a frozen 2.6B-parameter LoopLM backbone, which applies shared Transformer weights repeatedly to update hidden states before each token prediction. To adapt this backbone for specific translation tasks, they employ lightweight LoRA adapters for each language pair. This architecture allows the model to increase its 'recurrent depth'—the number of internal computation steps—without increasing the number of parameters. The study systematically varies this recurrent depth (u=1 to 6) across 32 translation directions, covering high-, mid-, and low-resource languages, and compares the results against larger, non-recurrent models trained under identical conditions.

LatentMT demonstrates that recurrent computation is a powerful tool for MT. Across all tested resource regimes, LatentMT achieves performance comparable to models three to five times its size. In high-resource settings, it closes the gap with domain-specialized models, while in mid- and low-resource settings, it frequently outperforms larger baselines like the 8B-parameter Qwen3-8B. The authors observe that translation quality improves significantly with the first few recurrent steps but saturates quickly thereafter. Mechanistic analysis confirms this, showing that hidden-representation differences between steps diminish as the model reaches a stable output distribution.