Research

Paper

AI LLM February 20, 2026

Thinking by Subtraction: Confidence-Driven Contrastive Decoding for LLM Reasoning

Authors

Lexiang Tang, Weihao Gao, Bingchen Zhao, Lu Ma, Qiao jin, Bang Yang, Yuexian Zou

Abstract

Recent work on test-time scaling for large language model (LLM) reasoning typically assumes that allocating more inference-time computation uniformly improves correctness. However, prior studies show that reasoning uncertainty is highly localized: a small subset of low-confidence tokens disproportionately contributes to reasoning errors and unnecessary output expansion. Motivated by this observation, we propose Thinking by Subtraction, a confidence-driven contrastive decoding approach that improves reasoning reliability through targeted token-level intervention. Our method, Confidence-Driven Contrastive Decoding, detects low-confidence tokens during decoding and intervenes selectively at these positions. It constructs a contrastive reference by replacing high-confidence tokens with minimal placeholders, and refines predictions by subtracting this reference distribution at low-confidence locations. Experiments show that CCD significantly improves accuracy across mathematical reasoning benchmarks while substantially reducing output length, with minimal KV-cache overhead. As a training-free method, CCD enhances reasoning reliability through targeted low-confidence intervention without computational redundancy. Our code will be made available at: https://github.com/bolo-web/CCD.

Metadata

arXiv ID: 2602.18232
Provider: ARXIV
Primary Category: cs.CL
Published: 2026-02-20
Fetched: 2026-02-23 05:33

Related papers

Raw Data (Debug)
{
  "raw_xml": "<entry>\n    <id>http://arxiv.org/abs/2602.18232v1</id>\n    <title>Thinking by Subtraction: Confidence-Driven Contrastive Decoding for LLM Reasoning</title>\n    <updated>2026-02-20T14:13:22Z</updated>\n    <link href='https://arxiv.org/abs/2602.18232v1' rel='alternate' type='text/html'/>\n    <link href='https://arxiv.org/pdf/2602.18232v1' rel='related' title='pdf' type='application/pdf'/>\n    <summary>Recent work on test-time scaling for large language model (LLM) reasoning typically assumes that allocating more inference-time computation uniformly improves correctness. However, prior studies show that reasoning uncertainty is highly localized: a small subset of low-confidence tokens disproportionately contributes to reasoning errors and unnecessary output expansion. Motivated by this observation, we propose Thinking by Subtraction, a confidence-driven contrastive decoding approach that improves reasoning reliability through targeted token-level intervention. Our method, Confidence-Driven Contrastive Decoding, detects low-confidence tokens during decoding and intervenes selectively at these positions. It constructs a contrastive reference by replacing high-confidence tokens with minimal placeholders, and refines predictions by subtracting this reference distribution at low-confidence locations. Experiments show that CCD significantly improves accuracy across mathematical reasoning benchmarks while substantially reducing output length, with minimal KV-cache overhead. As a training-free method, CCD enhances reasoning reliability through targeted low-confidence intervention without computational redundancy. Our code will be made available at: https://github.com/bolo-web/CCD.</summary>\n    <category scheme='http://arxiv.org/schemas/atom' term='cs.CL'/>\n    <category scheme='http://arxiv.org/schemas/atom' term='cs.AI'/>\n    <published>2026-02-20T14:13:22Z</published>\n    <arxiv:primary_category term='cs.CL'/>\n    <author>\n      <name>Lexiang Tang</name>\n    </author>\n    <author>\n      <name>Weihao Gao</name>\n    </author>\n    <author>\n      <name>Bingchen Zhao</name>\n    </author>\n    <author>\n      <name>Lu Ma</name>\n    </author>\n    <author>\n      <name>Qiao jin</name>\n    </author>\n    <author>\n      <name>Bang Yang</name>\n    </author>\n    <author>\n      <name>Yuexian Zou</name>\n    </author>\n  </entry>"
}