Skip to content
OpenSmartRoute

LLMs1 min read

CLLPU Benchmark Reveals Propagation Challenges in Multilingual LLM Unlearning

A new benchmark, CLLPU, assesses multilingual LLM unlearning by evaluating the extent of knowledge suppression across languages. Experiments with Llama-3.1-8B-Instruct show that universal suppression often fails to fully remove target knowledge, while language-conditioned confinement leads to uncontrolled propagation.

By OpenSmartRoute editorial · written through the router by writer-small

From arXiv cs.CL - “Beyond Cross-Lingual Transfer: Benchmarking Propagation Boundaries in Multilingual LLM Unlearning”

The research introduces CLLPU (Cross-Lingual and Language-Bound Protocol for LLM Unlearning), a benchmark designed to evaluate multilingual LLM unlearning beyond cross-lingual transfer. The benchmark utilizes two distinct settings: common-goal forgetting, aiming for suppression across all languages, and language-conditioned forgetting, restricting suppression to a specific language. CLLPU constructs 800 matched knowledge-unit pairs and 72,000 QA instances across ten languages, employing goal-guided topic pairing, schema-aware relation matching, and dual-anchor multilingual translation. Six representative methods were evaluated using Llama-3.1-8B-Instruct, revealing divergent failure modes. Specifically, universal suppression resulted in incomplete forgetting, while language-conditioned confinement caused propagation beyond the intended boundary. The research indicates that general multilingual utility can mask damage to neighboring knowledge. These findings highlight propagation control as a key challenge in multilingual LLM unlearning.

Source: https://arxiv.org/abs/2609.05976

Published Sep 9, 2026 · updated Sep 9, 2026 · 124 words

Keep reading

Related posts

More in LLMs

LLMs1 min read

AlignDiff: Filtering Preference Data with Model Signals

AlignDiff is a new framework that improves LLM alignment by prioritizing challenging preference samples based on model-intrinsic signals like negative log-likelihood gaps. Evaluations on LLaMA, Qwen, and common benchmarks demonstrate consistent performance improvements over established baselines.

LLMs1 min read

CONDUIT: KV Cache Reuse Policy for VLMs

CONDUIT is a training-free refresh policy for KV caches in vision-language models, unifying single and multi-image reuse. It achieves high accuracy and speed improvements with a small refresh budget, outperforming budgeted methods on latency.