Compress once, personalize forever - ComMer's approach to efficient LLM adaptation.
ComMer introduces a framework that compresses and merges user data into compact representations for efficient LLM personalization, reducing computational costs while maintaining performance.
-----
https://arxiv.org/abs/2501.03276
🤔 Original Problem:
Personalizing LLMs faces two major challenges: prompt engineering hits context window limits and is computationally expensive, while fine-tuning requires substantial resources for training individual user models.
-----
🔧 Solution in this Paper:
→ ComMer compresses each user document independently into a soft prompt using a frozen LLM with trainable compression embeddings
→ These compressed representations are merged through mean pooling into a single compact form
→ The merged representation is fed into a frozen LLM for generating personalized responses
→ The entire process is trained end-to-end using cross-entropy loss
-----
💡 Key Insights:
→ ComMer excels in personalized skill learning tasks but shows limitations in knowledge-intensive scenarios
→ Performance improves with more documents in skill learning tasks, following a power-law relationship
→ Mean pooling proves more effective than concatenation for merging document representations
→ The choice of pretraining dataset has minimal impact on final performance
-----
📊 Results:
→ Achieves superior quality with fewer resources compared to prompt-tuning within 128-token budget
→ Shows improved perplexity metrics when exposed to more documents than training
→ Demonstrates degraded performance in knowledge-intensive tasks with multiple documents
Share this post