English

Dual-Head Reasoning Distillation: Improving Classifier Accuracy with Train-Time-Only Reasoning

Computation and Language 2025-09-30 v2 Artificial Intelligence

Abstract

Chain-of-Thought (CoT) prompting often improves classification accuracy, but it introduces a significant throughput penalty with rationale generation (Wei et al., 2022; Cheng and Van Durme, 2024). To resolve this trade-off, we introduce Dual-Head Reasoning Distillation (DHRD), a simple training method for decoder-only language models (LMs) that adds (i) a pooled classification head used during training and inference and (ii) a reasoning head supervised by teacher rationales used only in training. We train with a loss function that is a weighted sum of label cross-entropy and token-level LM loss over input-plus-rationale sequences. On seven SuperGLUE tasks, DHRD yields relative gains of 0.65-5.47% over pooled baselines, with notably larger gains on entailment/causal tasks. Since we disable the reasoning head at test time, inference throughput matches pooled classifiers and exceeds CoT decoding on the same backbones by 96-142 times in QPS.

Keywords

Cite

@article{arxiv.2509.21487,
  title  = {Dual-Head Reasoning Distillation: Improving Classifier Accuracy with Train-Time-Only Reasoning},
  author = {Jillian Xu and Dylan Zhou and Vinay Shukla and Yang Yang and Junrui Ruan and Shuhuai Lin and Wenfei Zou and Yinxiao Liu and Karthik Lakshmanan},
  journal= {arXiv preprint arXiv:2509.21487},
  year   = {2025}
}

Comments

39th Conference on Neural Information Processing Systems (NeurIPS 2025) Efficient Reasoning Workshop

R2 v1 2026-07-01T05:56:56.763Z