Anti-Length Shift: Dynamic Outlier Truncation for Training Efficient Reasoning Models

TL;DR

The paper tackles the inefficiency from lengthy reasoning in RL-based training of large language models by identifying a length shift phenomenon and proposing Dynamic Outlier Truncation (DOT), a post-hoc tail-truncation technique applied only to all-correct rollout groups. DOT is paired with KL-Cov regularization and Predictive Dynamic Sampling to stabilize training and preserve exploration, enabling substantial token reductions without sacrificing accuracy. Across multiple model scales and reasoning benchmarks, DOT expands the efficiency–performance Pareto frontier, achieving major token savings on AIME-24 and robust generalization to out-of-distribution tasks. The work offers a simple, robust, and scalable approach for training efficient reasoning models with potential applicability to future post-training or agentic settings.

Abstract

Large reasoning models enhanced by reinforcement learning with verifiable rewards have achieved significant performance gains by extending their chain-of-thought. However, this paradigm incurs substantial deployment costs as models often exhibit excessive verbosity on simple queries. Existing efficient reasoning methods relying on explicit length penalties often introduce optimization conflicts and leave the generative mechanisms driving overthinking largely unexamined. In this paper, we identify a phenomenon termed length shift where models increasingly generate unnecessary reasoning on trivial inputs during training. To address this, we introduce Dynamic Outlier Truncation (DOT), a training-time intervention that selectively suppresses redundant tokens. This method targets only the extreme tail of response lengths within fully correct rollout groups while preserving long-horizon reasoning capabilities for complex problems. To complement this intervention and ensure stable convergence, we further incorporate auxiliary KL regularization and predictive dynamic sampling. Experimental results across multiple model scales demonstrate that our approach significantly pushes the efficiency-performance Pareto frontier outward. Notably, on the AIME-24, our method reduces inference token usage by 78% while simultaneously increasing accuracy compared to the initial policy and surpassing state-of-the-art efficient reasoning methods.

Figures (23)

• Figure 1: Performance-efficiency comparison on AIME-24 across two model scales.
• Figure 2: Evolution of average response length on all-correct queries during RL and SFT training.
• Figure 3: Co-evolution of reasoning word count and response length on test problems of varying difficulty.
• Figure 4: Evolution of policy entropy during training.
• Figure 5: Evolution of group ratios during training.