NeurIPS Paper Explains Why Knowledge Editing in English Fails to Propagate Across Languages in Multilingual LLMs
Tanmoy_Chak · x · 2026-11-16
A NeurIPS paper offers the first theoretical explanation for why knowledge editing in multilingual LLMs fails to propagate across languages.
- Using associative-memory locate-then-edit methods (ROME, MEMIT, AlphaEdit) as analytical frameworks, the authors test the assumption that facts live in a language-agnostic interlingua.
- They derive a closed-form decomposition of the cross-lingual gap into three terms: key-space propagation, parallel residual mismatch, and orthogonal residual mismatch — with the orthogonal residual component dominating, meaning source-language edits never span the update directions target languages need.
- Fixes relying only on source-language residual statistics hit a provable lower bound; contrastive realignment improves key-space propagation but value-space mismatch remains.
- Experiments span Llama-3.1-8B and Gemma-3-4B across EN, DE, HI, UR, and ZH.
The takeaway: this is a geometric bottleneck in the model's representations, not a flaw in the editing algorithms.
Related event: Study reveals cross-lingual limits of knowledge editing in LLMs(2 posts)→
More from Research
- Translated open-sources Lara Mini Base: 0.6B-active MoE beats models trained on 4-6x more tokens — NielsRogge · 2026-11-16
- Researchers formally specify ops interfaces for CockroachDB, Kafka and RabbitMQ to make AI operators verifiable — tianyin_xu · 2026-11-16
- Astera Residencies open: up to $2M in funding for off-network researchers — dr_alphalyrae · 2026-11-16
- 'It Knows Google Is Watching': Viral Chat Surfaces J-Lens Evaluation-Awareness Research — repligate · 2026-11-16
- MDPI review surveys reinforcement learning approaches in quantitative trading — PtrPomorski · 2026-11-16
- JEM trained in 20 H100-hours lands 1 point behind fine-tuned DINOv3 — gabriberton · 2026-11-16