Knowledge Interaction in Hypernetwork-based Continual Learning: Conflict and Composition
Continual learning aims to adapt pretrained large language models (LLMs) to evolving data while preserving their pretrained capabilities.
ProofPaper ↗
Key points
- A central challenge is catastrophic forgetting: adaptation to new data can impair performance on previously learned knowledge.
- We study these interactions in a hypernetwork-based adaptation framework, where a shared network generates adapters from context, allowing a frozen base model to adapt without task-specific updates.
- We find that, under knowledge conflict, models preferentially follow the context, and their choices depend more on where knowledge is presented than on its content.
- In multi-hop reasoning, we find that models fail to reliably compose knowledge across the three sources, with the key breakdown arising from the interaction of the hypernetwork with both the base model and the context.
Sources (1)
- [1]Knowledge Interaction in Hypernetwork-based Continual Learning: Conflict and CompositionarXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 11, 02:57 PM
Continual learning aims to adapt pretrained large language models (LLMs) to evolving data while preserving their pretrained capabilities.
A central challenge is catastrophic forgetting: adaptation to new data can impair performance on previously learned knowledge.
Extractive summary: sentences quoted from the sources.
Before this
- Oct 10, 2026Exploring Jailbreak Vulnerability Across Depth-Based Dynamic Inference in Large Language Models
- Oct 10, 2026Beyond Adjacent Layers: Graph-Guided Layer Fusion for Compressing Large Language Models
- Oct 10, 2026SAGE: Pushing 2-Bit Vector-Quantized LLMs toward Real-World Usability
- Oct 10, 2026Uniform Information Density-Based Preference Optimization
- Oct 10, 2026LongHisDoc: A Comprehensive Benchmark for Chinese Long Historical Document Understanding
- Oct 9, 2026SAT-RL: Training Language Models for Logical Reasoning with Symbolic Solvers