ArborKV: Structure-Aware KV Cache Management for Scaling Tree-based LLM Reasoning

2026-05-21

AI Summary

This work addresses the severe memory bottleneck in tree-based reasoning frameworks—such as Tree-of-Thoughts (ToT)—which arises from retaining extensive intermediate key-value (KV) caches when scaling search depth and breadth. To mitigate this, the authors propose a structure-aware KV cache management mechanism that employs a lightweight value estimator to guide cache allocation. This approach integrates token-level extractive eviction with a lazy rehydration strategy, substantially reducing memory overhead while preserving the ability to backtrack during reasoning. Evaluated on ToT reasoning benchmarks, the method achieves up to a 4× reduction in peak KV cache memory usage compared to full retention, with minimal degradation in reasoning accuracy. Consequently, it enables significantly larger-scale tree search configurations previously hindered by memory constraints.

Abstract

Recent progress in LLM reasoning has increasingly shifted from single-pass generation to explicit search over intermediate reasoning states. Tree-of-Thoughts (ToT) organizes inference to tree-structured search with branching and backtracking, but it substantially amplifies the Key--Value (KV) cache: retaining KV states for a frontier of partial trajectories quickly becomes a memory bottleneck that limits throughput and constrains search depth and width under fixed hardware budgets. We address this challenge by observing that KV reuse in ToT-style inference is governed by search dynamics: near-term decoding depends primarily on the active branch and its ancestors, whereas inactive subtrees have low short-term reuse probability yet must remain recoverable for backtracking. Motivated by this, we propose ArborKV, a structure-aware eviction framework that couples a lightweight value estimator with a tree-aware allocation policy, and performs purely token-extractive eviction with lazy rehydration to support revisits. Experiments on ToT-style reasoning benchmarks show that ArborKV achieves up to ~4x peak KV-memory reduction while preserving near-full-retention accuracy, enabling larger search configurations under fixed device budgets that would otherwise run out of memory.

Problem

• KV cache
• Tree-of-Thoughts
• memory bottleneck
• LLM reasoning
• tree-based search

Innovation

• structure-aware KV cache
• Tree-of-Thoughts
• memory-efficient LLM inference
• token-extractive eviction
• lazy rehydration