Reproduce “Tool-Adaptive LLM Reranker” on OpenResearch

About this paper

Zichuan Liu, Ruijin Hua

TALRanker is an LLM reranking framework that dynamically decides whether to use internal knowledge or invoke external search tools for relevance estimation. The system achieves state-of-the-art accuracy on reasoning-intensive benchmarks, such as BRIGHT, while maintaining high inference throughput comparable to non-reasoning models.

Key insights

• Adaptive tool invocation can effectively balance high relevance accuracy with efficient inference speed, overcoming the traditional trade-off in LLM-based reranking.
• Preserving the LLM's native generative and reasoning capabilities during discriminative fine-tuning is crucial for enabling effective tool-adaptive behavior in subsequent stages.
• An asymmetric, cost-aware reward function in reinforcement learning is key to teaching LLMs to dynamically manage their epistemic boundaries, using tools only when internal knowledge is insufficient or uncertain.

Summary by alphaXiv.