Paper Overview
Field: NLP Authors: Bangji Yang, Hongbo Ma, Jiajun Fan Published: 2025-04-01 arXiv: 2504.01258
Summary
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.
The authors introduce Batched Contextual Reinforcement (BCR), 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. A novel task-scaling law: as the number of concurrent problems N increases, per-problem token usage monotonically decreases, while accuracy declines far more gently than baselines — establishing N as a controllable throughput dimension. 2. A "free lunch" phenomenon: BCR challenges the traditional accuracy-efficiency trade-off. Across 1.5B and 4B model families, BCR reduces token usage by 15.8% to 62.6% while consistently maintaining or improving accuracy on five major math benchmarks. 3. Emergent self-regulating efficiency: qualitative analysis reveals that models autonomously eliminate redundant metacognitive loops without any explicit length supervision. 4. Stable length control: the implicit budget constraint empirically avoids the adversarial gradients and catastrophic optimization collapse inherent in explicit length penalties, providing a highly stable constraint-based alternative for length control.
These results demonstrate the practicality of BCR, showing that simple structural incentives can unlock latent high-density reasoning in LLMs.
---
*Auto-collected on 2026-04-04*