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

2026-02-18

AI Summary

This work identifies and addresses a critical issue in semantic ID-based recommendation: the integration of Chain-of-Thought (CoT) reasoning induces “reasoning drift,” which impairs the model’s sensitivity to ID semantics and consequently degrades recommendation performance. To mitigate this, the authors propose a training-free, inference-time subspace alignment framework that effectively integrates CoT reasoning while preserving the model’s original ID dependency. The approach combines reasoning chain compression, debiased contrastive decoding, and subspace alignment to harmonize semantic reasoning with ID-aware representations. This method significantly enhances recommendation accuracy without incurring additional training costs, offering a novel paradigm for controllable reasoning with large language models in recommender systems.

Technology Category

• Cognitive Modeling & Cognitive Systems: Conceptual Inference and Reasoning
• Knowledge Representation and Reasoning: Argumentation
• Data Mining & Knowledge Management: Recommender Systems

Application Category

• User Modeling, Personalization and Recommendation: Attacks and countermeasures in recommendation systems
• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs

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.

Problem

• Chain-of-Thought reasoning
• recommendation foundation models
• Semantic ID
• reasoning shift
• textual inertia

Innovation

• Chain-of-Thought
• Semantic ID
• Subspace Alignment
• Contrastive Decoding
• Foundation Recommender Models