ReaCon and SG-LoRI: Reducing Content Effect in Large Language Models via Controlled Reasoning Interventions

This article introduces an innovative solution to the content effect problem in large language models: the ReaCon benchmark dataset and the SG-LoRI method. ReaCon separates logical validity from semantic plausibility through fine-grained control, while SG-LoRI corrects model representations via pattern-guided low-rank interventions, making model reasoning rely more on formal logic rather than superficial semantic credibility.

Large language models have a 'content effect' bias: they tend to prefer semantically plausible conclusions even if they are logically invalid. For example, models are more likely to accept logically invalid but commonsense-consistent reasoning chains than logically valid but counterintuitive ones. This stems from exposure to large-scale text during pre-training, which makes models learn to 'sound right' rather than be logically valid, leading to systematic reasoning errors.

ReaCon is a controlled reasoning benchmark dataset for studying content effects, with the core goal of separating key variables:

• Controllable dimensions: logical validity, semantic plausibility, numerical correctness, counterfactual perturbation, reasoning difficulty, out-of-distribution generalization
• Annotation structure: JSONL format, including fields such as input text, logical validity label, numerical correctness label, reasoning difficulty, logical pattern, counterfactual flag, etc., supporting precise measurement of model reasoning behavior.

SG-LoRI is a parameter-efficient training-time intervention method with core designs:

• Freeze the pre-trained model backbone and only train lightweight components, with advantages of high parameter efficiency, modularity, interpretability, and reversibility
• Architectural components: pattern gating (identifies reasoning patterns), pattern-specific low-rank matrices, validity classifier, content effect metrics
• Workflow: pattern gating identifies patterns → activates corresponding low-rank adapters → intervenes on hidden representations → outputs logical validity predictions.

The experimental design includes ablation experiments and multi-dimensional evaluation:

• Ablation settings: no pattern supervision, full linear adapters, shared adapters, no pattern gating, to verify the value of each component
• Dataset splits: dev (tuning), test_iid (standard generalization), test_ood_vocab (vocabulary OOD), test_ood_structure (structural OOD), to comprehensively test generalization ability.

Theoretical contributions: ReaCon provides a diagnostic tool, SG-LoRI demonstrates a bias correction method, and low-rank interventions offer insights into representation structures Practical value: Can be applied in fields such as legal analysis (based on provisions rather than plausibility), medical diagnosis (avoiding misguidance from symptom descriptions), financial risk control (based on real risk signals), scientific reasoning (ensuring logical rigor), etc.

Current limitations: Training data needs to be prepared by users themselves, supported models are limited, scale expansion remains to be verified, only training-time intervention Future directions: Combining activation interventions, expanding model architectures, multi-modal reasoning, unsupervised pattern discovery, etc.