aiqbench analysis
This research addresses the inference cost of Chain-of-Thought reasoning by introducing SCA, a method that selectively compresses thinking segments while strictly preserving answer alignment to prevent degradation. By routing compression rewards only to successful thought tokens and aligning output distributions with a frozen base model, it achieves state-of-the-art efficiency without sacrificing accuracy across multiple domains. The work offers significant practical value for deploying structured reasoning models in cost-sensitive environments where maintaining precise outputs is critical.
SCA: Segment-Wise CoT Compression with Answer Alignment
arXiv:2603.07598v2 Announce Type: replace Abstract: Chain-of-thought (CoT) reasoning improves problem solving, but long think traces increase inference cost. Existing CoT compression methods usually optimize completion-level length. For structured thinking models, however, a completion contains both a think segment and an answer segment, so completion-level compression can save tokens by compressing not only the CoT but also the answer. We call this failure mode answer drift. We propose Segment-wise CoT Compression with Answer Alignment (SCA), an answer-preserving think-compression method. SCA parses completions into functional segments, routes compression rewards only to successful think tokens, and protects answer tokens through length and distribution alignment to a frozen base model. Experiments show that, across datasets from multiple domains, SCA achieves state-of-the-art-level chain-of-thought compression while preserving the base model's performance and answer alignment. Training data and code are included in the supplementary code and data package.
Abstract:Chain-of-thought (CoT) reasoning improves problem solving, but long think traces increase inference cost. Existing CoT compression methods usually optimize completion-level length. For structured thinking models, however, a completion contains both a think segment and an answer segment, so completion-level compression can save tokens by compressing not only the CoT but also the answer. We call this failure mode answer drift. We propose Segment-wise CoT Compression with Answer Alignment (SCA), an answer-preserving think-compression method. SCA parses completions into functional segments, routes compression rewards only to successful think tokens, and protects answer tokens through length and distribution alignment to a frozen base model. Experiments show that, across datasets from multiple domains, SCA achieves state-of-the-art-level chain-of-thought compression while preserving the base model's performance and answer alignment. Training data and code are included in the supplementary code and data package.
9 pages, 5 figures. This version substantially revises the previous preprint with a new method, updated experiments, and rewritten analysis. Code available at the GitHub project repository this https URL
Artificial Intelligence (cs.AI); Machine Learning (cs.LG)
arXiv:2603.07598 [cs.AI]
(or arXiv:2603.07598v2 [cs.AI] for this version)
Full text under the publisher's open licence · 180 words
Related coverage
ViSAGE: Constructing Self-Correcting Memories for Long-Form Video Understanding
arXiv · Aug 3, 2026
D3S2: Diffusion-Guided Dataset Distillation for Semantic Segmentation
arXiv · Oct 7, 2026
Geometric Alignment without Functional Equivalence: A Layer-wise Analysis of the Speech-Text Modality Gap
arXiv · Oct 15, 2026
[AINews] not much happened today
Latent Space · Oct 3, 2026