Batched Contextual Reinforcement: A Task-Scaling Law for Efficient Reasoning

2026-04-02

AI Summary

This work addresses the high token consumption and associated costs of chain-of-thought reasoning in large language models, where existing optimizations often compromise output quality or require complex training procedures. The authors propose Batched Contextual Reinforcement (BCR), a single-stage reinforcement learning approach that concurrently solves N problems within a shared context window, using individual problem accuracy as the reward signal to implicitly enforce a token budget constraint. BCR reveals a task-scaling law: as N increases, tokens per problem decrease monotonically while accuracy degrades significantly less than baselines. Remarkably, under standard single-problem inference, BCR yields a “free lunch” effect—improving accuracy without additional cost—and autonomously eliminates redundant metacognitive loops, avoiding optimization collapse caused by explicit length penalties. Experiments on 1.5B and 4B models show token reductions of 15.8%–62.6% while maintaining or enhancing performance across five major mathematical benchmarks.

Abstract

Large Language Models employing Chain-of-Thought reasoning achieve strong performance but suffer from excessive token consumption that inflates inference costs. Existing efficiency methods such as explicit length penalties, difficulty estimators, or multi-stage curricula either degrade reasoning quality or require complex training pipelines. We introduce Batched Contextual Reinforcement, a minimalist, single-stage training paradigm that unlocks efficient reasoning through a simple structural modification: training the model to solve N problems simultaneously within a shared context window, rewarded purely by per-instance accuracy. This formulation creates an implicit token budget that yields several key findings: (1) We identify a novel task-scaling law: as the number of concurrent problems N increases during inference, per-problem token usage decreases monotonically while accuracy degrades far more gracefully than baselines, establishing N as a controllable throughput dimension. (2) BCR challenges the traditional accuracy-efficiency trade-off by demonstrating a "free lunch" phenomenon at standard single-problem inference. Across both 1.5B and 4B model families, BCR reduces token usage by 15.8% to 62.6% while consistently maintaining or improving accuracy across five major mathematical benchmarks. (3) Qualitative analyses reveal emergent self-regulated efficiency, where models autonomously eliminate redundant metacognitive loops without explicit length supervision. (4) Crucially, we empirically demonstrate that implicit budget constraints successfully circumvent the adversarial gradients and catastrophic optimization collapse inherent to explicit length penalties, offering a highly stable, constraint-based alternative for length control. These results prove BCR practical, showing simple structural incentives unlock latent high-density reasoning in LLMs.

Problem

• token efficiency
• Chain-of-Thought reasoning
• inference cost
• reasoning quality
• length penalty

Innovation

• Batched Contextual Reinforcement
• task-scaling law
• implicit token budget
• efficient reasoning
• length control