Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Why Thinking Hurts? Diagnosing and Rectifying the Reasoning Shift in Foundation Recommender Models

Luankang Zhang, Yonghao Huang, Hang Lv, Mingjia Yin, Liangyue Li, Zulong Chen, Hao Wang, Enhong Chen

TL;DR

By compressing reasoning chains and applying bias-subtracted contrastive decoding, this approach mitigates ungrounded textual drift and effectively calibrates inference, allowing foundation models to leverage reasoning without sacrificing ID-grounded accuracy.

Abstract

Integrating Chain-of-Thought (CoT) reasoning into Semantic ID-based recommendation foundation models (such as OpenOneRec) often paradoxically degrades recommendation performance. We identify the root cause as textual inertia from the General Subspace, where verbose reasoning dominates inference and causes the model to neglect critical Semantic ID. To address this, we propose a training-free Inference-Time Subspace Alignment framework. By compressing reasoning chains and applying bias-subtracted contrastive decoding, our approach mitigates ungrounded textual drift. Experiments show this effectively calibrates inference, allowing foundation models to leverage reasoning without sacrificing ID-grounded accuracy.

Why Thinking Hurts? Diagnosing and Rectifying the Reasoning Shift in Foundation Recommender Models

TL;DR

By compressing reasoning chains and applying bias-subtracted contrastive decoding, this approach mitigates ungrounded textual drift and effectively calibrates inference, allowing foundation models to leverage reasoning without sacrificing ID-grounded accuracy.

Abstract

Integrating Chain-of-Thought (CoT) reasoning into Semantic ID-based recommendation foundation models (such as OpenOneRec) often paradoxically degrades recommendation performance. We identify the root cause as textual inertia from the General Subspace, where verbose reasoning dominates inference and causes the model to neglect critical Semantic ID. To address this, we propose a training-free Inference-Time Subspace Alignment framework. By compressing reasoning chains and applying bias-subtracted contrastive decoding, our approach mitigates ungrounded textual drift. Experiments show this effectively calibrates inference, allowing foundation models to leverage reasoning without sacrificing ID-grounded accuracy.
Paper Structure (22 sections, 5 equations, 3 figures, 2 tables)

This paper contains 22 sections, 5 equations, 3 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. LLM-Based Recommendation
  4. Reasoning-Enhanced Recommendation
  5. Empirical Evaluation
  6. Bias Analysis in the Thinking Process
  7. Case Study of Bias Impact
  8. Method: Inference-Time Subspace Alignment
  9. Reasoning-Chain Compression
  10. Compression operator.
  11. Instantiation via constrained summarization.
  12. Bias-Subtracted Contrastive Inference
  13. Contextual scoring.
  14. Score normalization.
  15. Bias-subtracted contrastive scoring.
  16. ...and 7 more sections

Figures (3)

  • Figure 1: Impact of Thinking Mode on OpenOneRec.
  • Figure 2: Visiualization of Two Token Subspaces.
  • Figure 3: Prompt Structure for Section \ref{['sec:cot-compression']}.