Why multilingual LLMs encode translations identically: new theory explains the platonic representation hypothesis
Hidenori8Tanaka · x · 2026-10-12
- The phenomenon: multilingual LLMs encode a sentence and its translation almost identically — the "platonic representation hypothesis," observed but never explained.
- Accumulated facts: cross-lingual similarity peaks in middle layers (not the ends), coexists with language-specific structure, and grows with language proximity, model quality, and data volume.
- New theory and prediction: Matthieu Wyart and team explain all of these facts and predict something new — removing the part of each layer linearly predictable from the previous one should raise cross-lingual similarity.
- Validation: the prediction is confirmed in real LLMs, with plots matching trained transformers against theoretical (BP) networks on II and mKNN metrics, same-model and cross-model.
More from Research
- Meituan LongCat's QuantMLA: joint INT4 MLA KV caching with 3.59x compression at 128K — meituan-longcat · 2026-10-12
- MOTIVE teaches VLMs multi-perspective self-verification and when to rethink answers — Ziquan Zhu · 2026-10-12
- Free-threaded Python 3.14t plus fastsimd parses JSON at 65.7 GB/s on 64 cores — lemire · 2026-10-12
- NVIDIA's open-source SDM library unifies tabular foundation models — pandeyparul · 2026-10-12
- Qwen releases QwenGyre: elastic GPU scheduling speeds long-horizon agent RL training by up to 1.85x — Alibaba_Qwen · 2026-10-12
- Crowdsourced push on integer multiplication bounds stalls: κ improves only 0.7%, geometry experts sought — iskander · 2026-10-12