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

TL;DR

This paper investigates why integrating Chain-of-Thought reasoning into recommendation models can harm performance and proposes a training-free inference-time method to align reasoning with semantic IDs, improving accuracy.

Contribution

It introduces a novel inference-time subspace alignment technique that mitigates textual inertia and enhances reasoning calibration without additional training.

Findings

Effective in reducing textual drift during inference

Improves recommendation accuracy with reasoning integration

Calibrates foundation models to leverage reasoning without losing ID-grounded performance

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.