Efficient and Trainable Language Model Test-Time Scaling via Local Branch Routing

A new token-level framework called Local Branch Routing (LBR) aims to improve language model reasoning at test time while lowering computational cost, according to a paper submitted on 24 Jun 2026 [1]. The method uses a small local lookahead tree and a lightweight router to select the most promising token path [1]. The framework, detailed on the preprint server arXiv, addresses a known trade-off in test-time scaling for large language models [1]. Long chain-of-thought sampling remains single-threaded, while broader search strategies at the sentence or solution level are computationally expensive and difficult to train end-to-end [2]. LBR operates at the token level, expanding a small local lookahead tree and forwarding all sampled branches through the model [1]. A lightweight router then examines the hidden states of these candidate local futures to select a depth-1 subtree to commit to, a process the authors describe as prune-shift-grow decoding [2]. This design preserves discrete branch identities and defines a tractable tree-trajectory likelihood, where newly grown nodes are counted when first sampled and router decisions receive explicit probabilities [2]. The structure enables end-to-end reinforcement learning with verifiable rewards, jointly optimizing the base model and the router under the same likelihood-ratio principle used in discrete-token reinforcement learning from verifiable rewards (RLVR) [1]. Neural networks, the underlying architecture for such models, consist of connected layers of artificial neurons that process signals through weighted connections, with training typically accelerated by graphics processing units and large datasets [6]. On synthetic hierarchical-planning tasks, the researchers found that post-candidate hidden states provided useful routing evidence [2]. On mathematical reasoning benchmarks, LBR improved both Pass@1 and Pass@32 metrics over discrete chain-of-thought, vanilla discrete-token RLVR, and RL-compatible soft-token branching baselines [1]. The paper, posted on arXiv—an open-access repository that hosts preprints across physics, computer science, and related fields and receives about 24,000 submissions per month as of late 2024—has not yet been peer reviewed [7]. The authors suggest that lightweight local branching offers an efficient, trainable, and discrete form of language-model test-time scaling [2].