Academic

AdaptFuse: Training-Free Sequential Preference Learning via Externalized Bayesian Inference

arXiv:2604.03925v1 Announce Type: new Abstract: Large language models struggle to accumulate evidence across multiple rounds of user interaction, failing to update their beliefs in a manner consistent with Bayesian inference. Existing solutions require fine-tuning on sensitive user interaction data, limiting their applicability in privacy-conscious settings. We propose AdaptFuse, a training-free framework that externalizes probabilistic computation entirely from the LLM: a symbolic module maintains a Bayesian posterior over a discrete hypothesis set, while a frozen LLM contributes semantic reasoning via multi-sample Dirichlet aggregation. The two signals are combined through entropy-adaptive fusion, which automatically weights each source by its predictive confidence, shifting reliance from the LLM to the symbolic posterior as evidence accumulates. We evaluate across three domains: flight recommendation, hotel recommendation, and web shopping; on Gemma 2 9B, Llama 3 8B, and Qwen 2.5 7

F
Fangzhou Lin, Peiran Li, Shuo Xing, Siyuan Yang, Qianwen Ge, Kazunori Yamada, Ziming Zhang, Haichong Zhang, Zhengzhong Tu · · 1 min read · 5 views
#cs.CL #cs.AI

arXiv:2604.03925v1 Announce Type: new Abstract: Large language models struggle to accumulate evidence across multiple rounds of user interaction, failing to update their beliefs in a manner consistent with Bayesian inference. Existing solutions require fine-tuning on sensitive user interaction data, limiting their applicability in privacy-conscious settings. We propose AdaptFuse, a training-free framework that externalizes probabilistic computation entirely from the LLM: a symbolic module maintains a Bayesian posterior over a discrete hypothesis set, while a frozen LLM contributes semantic reasoning via multi-sample Dirichlet aggregation. The two signals are combined through entropy-adaptive fusion, which automatically weights each source by its predictive confidence, shifting reliance from the LLM to the symbolic posterior as evidence accumulates. We evaluate across three domains: flight recommendation, hotel recommendation, and web shopping; on Gemma 2 9B, Llama 3 8B, and Qwen 2.5 7B. AdaptFuse consistently outperforms both prompting baselines and fine-tuned Bayesian Teaching models on all tasks, with accuracy improving monotonically over interaction rounds. These results demonstrate that principled inference-time algorithms can substitute for fine-tuning in personalized recommendation, without storing or training on sensitive user data. All the code and materials will be open-sourced.

Executive Summary

AdaptFuse, a training-free sequential preference learning framework, has been proposed to address the limitations of large language models (LLMs) in accumulating evidence across multiple rounds of user interaction. By externalizing probabilistic computation from the LLM and combining it with a symbolic module, AdaptFuse achieves principled inference-time algorithms that outperform fine-tuned Bayesian Teaching models in personalized recommendation tasks without storing or training on sensitive user data. The framework's entropy-adaptive fusion mechanism automatically weighs each source by its predictive confidence, shifting reliance from the LLM to the symbolic posterior as evidence accumulates. Evaluations across three domains demonstrate AdaptFuse's consistency in outperforming prompting baselines and fine-tuned models, with accuracy improving monotonically over interaction rounds.

Key Points

  • ▸ AdaptFuse proposes a training-free framework for sequential preference learning.
  • ▸ The framework externalizes probabilistic computation from the LLM and combines it with a symbolic module.
  • ▸ AdaptFuse achieves principled inference-time algorithms that outperform fine-tuned Bayesian Teaching models in personalized recommendation tasks.

Merits

Strength in Privacy-Conscious Settings

AdaptFuse addresses the limitation of LLMs in accumulating evidence across multiple rounds of user interaction without storing or training on sensitive user data, making it an attractive solution for privacy-conscious settings.

Demerits

Potential Overreliance on Symbolic Module

The quality and effectiveness of AdaptFuse's symbolic module may impact the overall performance of the framework, and potential overreliance on this module could limit the framework's adaptability in certain scenarios.

Expert Commentary

AdaptFuse's proposed framework marks an important step towards developing principled inference-time algorithms for LLMs in personalized recommendation tasks. By externalizing probabilistic computation and combining it with a symbolic module, AdaptFuse achieves a level of performance that outperforms fine-tuned Bayesian Teaching models. While potential limitations, such as overreliance on the symbolic module, need to be addressed, the framework's implications for real-world applications and policy development are significant. As the field continues to evolve, it will be essential to monitor AdaptFuse's performance in various scenarios and explore its potential applications and limitations.

Recommendations

  • ✓ Further research is needed to explore the potential limitations of AdaptFuse, particularly in relation to the symbolic module's quality and effectiveness.
  • ✓ The development and deployment of AdaptFuse should be closely monitored to ensure that its benefits are realized while also addressing potential concerns around user data privacy and protection.

Sources

Original: arXiv - cs.CL

Related Articles