The Override Gap: A Magnitude Account of Knowledge Conflict Failure in Hypernetwork-Based Instant LLM Adaptation
Researchers from arXiv have published a study addressing systematic failures in hypernetwork-based methods, such as Doc-to-LoRA, when adapting Large Language Models (LLMs) to documents that contradict pretraining knowledge. The paper identifies the root cause as a magnitude problem rather than a representational one, where the adapter margin remains constant while pretrained margins grow with training frequency. This leads to significant accuracy drops in deep conflicts. To resolve this, the authors propose two training-free solutions: Selective Layer Boosting and Conflict-Aware Internalization. These methods scale adapters at top-norm layers and trigger boosting only when the base model is confident. Experimental results show substantial accuracy improvements on Gemma-2B and Mistral-7B models, outperforming vanilla retrieval-augmented generation. Additionally, the team released KID-Bench, a new benchmark designed to evaluate novel recall, cross-knowledge combination, and prior-graded conflicts in LLMs.
Wire timeline
The Override Gap: A Magnitude Account of Knowledge Conflict Failure in Hypernetwork-Based Instant LLM Adaptation
Researchers from arXiv have published a study addressing systematic failures in hypernetwork-based methods, such as Doc-to-LoRA, when adapting Large Language Models (LLMs) to documents that contradict pretraining knowledge. The paper identifies the root cause as a magnitude problem rather than a representational one, where the adapter margin remains constant while pretrained margins grow with training frequency. This leads to significant accuracy drops in deep conflicts. To resolve this, the authors propose two training-free solutions: Selective Layer Boosting and Conflict-Aware Internalization. These methods scale adapters at top-norm layers and trigger boosting only when the base model is confident. Experimental results show substantial accuracy improvements on Gemma-2B and Mistral-7B models, outperforming vanilla retrieval-augmented generation. Additionally, the team released KID-Bench, a new benchmark designed to evaluate novel recall, cross-knowledge combination, and prior-graded conflicts in LLMs.
cs.AI updates on arXiv.org