Learning Adaptive LLM Decoding

TL;DR

This paper introduces a method for learning adaptive decoding policies for large language models, dynamically selecting sampling strategies during inference to improve accuracy and efficiency across tasks like math and coding.

Contribution

It proposes lightweight reinforcement learning-based decoding adapters that adapt sampling strategies at both sequence and token levels without fine-tuning the language model.

Findings

Token-level adapter improves Pass@1 accuracy by up to 10.2% on MATH.

Sequence-level adapter yields 2-3% accuracy gains under fixed sampling budgets.

Both sequence- and token-level adaptations contribute to improved performance.

Abstract

Decoding from large language models (LLMs) typically relies on fixed sampling hyperparameters (e.g., temperature, top-p), despite substantial variation in task difficulty and uncertainty across prompts and individual decoding steps. We propose to learn adaptive decoding policies that dynamically select sampling strategies at inference time, conditioned on available compute resources. Rather than fine-tuning the language model itself, we introduce lightweight decoding adapters trained with reinforcement learning and verifiable terminal rewards (e.g. correctness on math and coding tasks). At the sequence level, we frame decoding as a contextual bandit problem: a policy selects a decoding strategy (e.g. greedy, top-k, min-p) for each prompt, conditioned on the prompt embedding and a parallel sampling budget. At the token level, we model decoding as a partially observable Markov decision process (POMDP), where a policy selects sampling actions at each token step based on internal model features and the remaining token budget. Experiments on the MATH and CodeContests benchmarks show that the learned adapters improve the accuracy-budget tradeoff: on MATH, the token-level adapter improves Pass@1 accuracy by up to 10.2% over the best static baseline under a fixed token budget, while the sequence-level adapter yields 2-3% gains under fixed parallel sampling. Ablation analyses support the contribution of both sequence- and token-level adaptation.