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

• URL: http://arxiv.org/abs/2509.21487v2
• Date: Mon, 29 Sep 2025 02:58:38 GMT
• Title: Dual-Head Reasoning Distillation: Improving Classifier Accuracy with Train-Time-Only Reasoning
• Abstract summary: We introduce Dual-Head Reasoning Distillation (DHRD), a simple training method for decoder-only language models (LMs)<n>On seven SuperGLUE tasks, DHRD yields relative gains of 0.65-5.47% over pooled baselines, with notably larger gains on entailment/causal tasks.<n>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.
• Score: 4.1343183246697475
• License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
• 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.

Related papers

• Better Starts, Better Ends: Bootstrapped Iterative Self-Reasoning Distillation for Compressed Reasoning [27.08087977200613]
  BIRD(Bootstrapped Iterative Self-Reasoning Distillation) is a two-stage self-reasoning distillation method.<n>It first samples concise solutions from a base model under a brevity instruction, keeps only answer-correct traces, and performs a lightweight prompt-switch SFT step.<n>It achieves a stronger accuracy-efficiency trade-off than prompting and cold-start on-policy distillation on MATH-500 and AIME benchmarks.
  arXiv  Detail & Related papers  (2026-07-17T08:15:35Z)
• Rethinking Reward Supervision: Rubric-Conditioned Self-Distillation [60.55792673956761]
  We propose textbfRubric-Conditioned Self-Distillation, a framework that incorporates rubrics as structured, fine-grained feedback for on-policy self-distillation.<n>We evaluate on a diverse suite of science reasoning benchmarks and results show that rubric-conditioned self-distillation effectively converts rubric-level criteria into token-level guidance.
  arXiv  Detail & Related papers  (2026-06-17T17:54:04Z)
• Dynamic Rollout Editing for Reducing Overthinking in RL-Trained Reasoning Models [102.76983747945836]
  Long-form chain-of-thought reasoning can improve performance on complex tasks.<n>But models often continue generating unnecessary reasoning after a correct answer has emerged.<n>We study this phenomenon from the perspective of GRPO-style reinforcement learning.
  arXiv  Detail & Related papers  (2026-06-16T13:10:30Z)
• Unified Data Selection for LLM Reasoning [66.0118347918729]
  Training Large Language Models (LLMs) for complex, long-CoT reasoning is often bottlenecked by the need for massive high-quality reasoning data.<n>We propose High-Entropy Sum (HES), a training-free metric that quantifies reasoning quality by summing only the entropy of the top (e.g., 0.5%) highest-entropy tokens in each reasoning sample.
  arXiv  Detail & Related papers  (2026-05-21T12:21:41Z)
• Constraint-Rectified Training for Efficient Chain-of-Thought [60.52883907721588]
  Chain-of-Thought (CoT) has significantly enhanced the reasoning capabilities of Large Language Models (LLMs)<n>While longer reasoning traces can improve answer quality and unlock abilities such as self-correction, they also incur high inference costs and often introduce redundant steps, known as overthinking.<n>Recent research seeks to develop efficient reasoning strategies that balance reasoning length and accuracy.
  arXiv  Detail & Related papers  (2026-02-13T02:13:45Z)
• Scaling Reasoning Hop Exposes Weaknesses: Demystifying and Improving Hop Generalization in Large Language Models [66.36240676392502]
  Chain-of-thought (CoT) reasoning has become the standard paradigm for enabling Large Language Models (LLMs) to solve complex problems.<n>Recent studies reveal a sharp performance drop in reasoning hop generalization scenarios.<n>We propose test-time correction of reasoning, a lightweight intervention method that dynamically identifies and deactivates ep heads in the reasoning process.
  arXiv  Detail & Related papers  (2026-01-29T03:24:32Z)
• Correct, Concise and Complete: Multi-stage Training For Adaptive Reasoning [11.179446105672461]
  We propose a multi-stage efficient reasoning method that combines supervised fine-tuning and reinforcement learning.<n>Our approach reduces response length by an average of 28% for 8B models and 40% for 32B models.<n>It achieves a superior trade-off compared to more complex state-of-the-art efficient reasoning methods.
  arXiv  Detail & Related papers  (2026-01-06T12:31:51Z)
• Can Large Reasoning Models Improve Accuracy on Mathematical Tasks Using Flawed Thinking? [0.0]
  Chain-of-thought (CoT) prompting has become central to mathematical reasoning in large language models.<n>We investigate whether training on intentionally flawed reasoning traces can teach models to detect and recover from such errors.
  arXiv  Detail & Related papers  (2025-12-18T21:20:21Z)
• Generative Adversarial Reasoner: Enhancing LLM Reasoning with Adversarial Reinforcement Learning [19.473649388687484]
  Large language models (LLMs) with explicit reasoning capabilities excel at mathematical reasoning yet still commit process errors.<n>We introduce Geneversarative Adrial Reasoner, an on-policy joint training framework designed to enhance reasoning.<n>A compute-efficient review schedule partitions each reasoning chain into logically complete slices of comparable length, and the discriminator evaluates each slice's soundness with structured justifications.
  arXiv  Detail & Related papers  (2025-12-18T18:59:54Z)
• DLER: Doing Length pEnalty Right - Incentivizing More Intelligence per Token via Reinforcement Learning [134.03095505580276]
  Doing Length pEnalty Right (DLER) is a training recipe combining batch-wise reward normalization, higher clipping, dynamic sampling, and a simple truncation length penalty.<n>DLER achieves state-of-the-art accuracy--efficiency trade-offs, cutting output length by over 70 percent while surpassing all previous baseline accuracy.
  arXiv  Detail & Related papers  (2025-10-16T20:05:57Z)
• Inducing Faithfulness in Structured Reasoning via Counterfactual Sensitivity [6.908972852063454]
  Large language models often generate a correct answer while relying on a flawed or irrelevant reasoning trace.<n>This paper introduces textbfCounterfactual Sensitivity Regularization (CSR), a novel training objective.<n>CSR improves faithfulness over standard fine-tuning and process supervision by up to 70 percentage points.
  arXiv  Detail & Related papers  (2025-09-01T15:18:46Z)
• Skip-Thinking: Chunk-wise Chain-of-Thought Distillation Enable Smaller Language Models to Reason Better and Faster [51.89995713333108]
  Chain-of-thought (CoT) distillation allows a large language model (LLM) to guide a small language model (SLM) in reasoning tasks.<n>Existing methods train the SLM to learn the long rationale in one iteration.<n>We propose chunk-wise training (CWT), which uses a search to divide the rationale into internal semantically coherent chunks.
  arXiv  Detail & Related papers  (2025-05-24T11:04:52Z)
• Interpretable Traces, Unexpected Outcomes: Investigating the Disconnect in Trace-Based Knowledge Distillation [14.489157453882767]
  This work aims to address the challenge of evaluating reasoning traces and their correlation with the final performance.<n>We employ a KD method leveraging rule-based problem decomposition to generate interpretable traces.<n>Specifically, we demonstrate this approach on Open Book QA, decomposing the problem into a Classification step and an Information Retrieval step.
  arXiv  Detail & Related papers  (2025-05-20T00:49:19Z)
• Rethinking Classifier Re-Training in Long-Tailed Recognition: A Simple Logits Retargeting Approach [102.0769560460338]
  We develop a simple logits approach (LORT) without the requirement of prior knowledge of the number of samples per class. Our method achieves state-of-the-art performance on various imbalanced datasets, including CIFAR100-LT, ImageNet-LT, and iNaturalist 2018.
  arXiv  Detail & Related papers  (2024-03-01T03:27:08Z)
• Dual Compensation Residual Networks for Class Imbalanced Learning [98.35401757647749]
  We propose Dual Compensation Residual Networks to better fit both tail and head classes. An important factor causing overfitting is that there is severe feature drift between training and test data on tail classes. We also propose a Residual Balanced Multi-Proxies classifier to alleviate the under-fitting issue.
  arXiv  Detail & Related papers  (2023-08-25T04:06:30Z)
• You Only Need End-to-End Training for Long-Tailed Recognition [8.789819609485225]
  Cross-entropy loss tends to produce highly correlated features on imbalanced data. We propose two novel modules, Block-based Relatively Balanced Batch Sampler (B3RS) and Batch Embedded Training (BET) Experimental results on the long-tailed classification benchmarks, CIFAR-LT and ImageNet-LT, demonstrate the effectiveness of our method.
  arXiv  Detail & Related papers  (2021-12-11T11:44:09Z)
• Pre-training Is (Almost) All You Need: An Application to Commonsense Reasoning [61.32992639292889]
  Fine-tuning of pre-trained transformer models has become the standard approach for solving common NLP tasks. We introduce a new scoring method that casts a plausibility ranking task in a full-text format. We show that our method provides a much more stable training phase across random restarts.
  arXiv  Detail & Related papers  (2020-04-29T10:54:40Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.