Cross-Lingual Transfer in LLMs Falls Short of Expectations, Sparking Debate
Technicians debate claims that pretraining transfer learning barely exists, agreeing cross-lingual transfer is real but weaker than expected five years ago. Jack Rae notes low-resource languages like Welsh consume tiny compute fractions despite lower capability, while others argue miscalibrated benchmarks like MMMLU may understate true transfer.
2026-08-19 ~ 2026-08-20 · 4 related posts
- Tech opinion: Pre-training transfer learning weaker than expected 5 years ago — jack_w_rae · 2026-08-19
- Cross-lingual transfer in LLMs is weaker than expected — willdepue · 2026-08-20
- Jack Rae on low-resource language eval and transfer efficiency — jack_w_rae · 2026-08-20
- Experts debate low-resource language transfer: calibration issues vs true capability — willdepue · 2026-08-20