FABLE: Lightweight Policy Layer for Personalizing LLM Agents
A team of researchers has unveiled FABLE (Factorized Adaptive Bandit Layer for Execution), a streamlined policy layer aimed at customizing large language model (LLM) agents without the need for fine-tuning the core model. This method views personalization as an online learning process that develops a unique execution policy for each user based on scalar feedback from actions taken. FABLE functions independently of a potentially opaque host agent, breaking down memory, information gathering, and response choices so that feedback can influence these decisions. It screens actions through a pre-defined feasible set prior to exploration and learns user-specific residual preferences against a stable baseline. This strategy tackles the difficulty of tailoring execution choices—like memory retrieval, tool usage, clarifying inquiries, and response styles—to individual users. Fine-tuning distinct LLMs can be expensive or unfeasible for proprietary systems, and prompts and memory mainly reveal user data rather than modify execution choices based on feedback. The research paper is accessible on arXiv with the identifier 2608.00215.
Key facts
- FABLE is a lightweight policy layer for personalizing LLM agents.
- It formulates personalization as online learning of a per-user execution policy.
- FABLE works with frozen, potentially black-box host agents.
- It factorizes memory, information-acquisition, and response decisions.
- Feedback updates related choices through factorization.
- Actions are filtered through an externally specified feasible set before exploration.
- FABLE learns user-specific residual preferences relative to a fixed baseline.
- The paper is available on arXiv with ID 2608.00215.
Entities
Institutions
- arXiv