Implicit Reasoning Steering via Concept Chaining

2026-07-15

AI Summary

Large language models exhibit decision fragility during reasoning, often yielding inconsistent answers to the same question across multiple samplings. This work proposes Concept Chaining, a method that implicitly constructs conceptual chains between problem entities and target answer options through natural language text. By leveraging continued pretraining, the approach systematically steers model preferences without requiring explicit instructions or answer hints. Experimental results demonstrate that Concept Chaining effectively biases model predictions in a manner more covert than direct prompt rewriting. The findings reveal that such reasoning vulnerabilities can be exploited using ordinary textual inputs to achieve stealthy manipulation, thereby highlighting significant practical safety risks inherent in current large language models.

Abstract

Large language models often appear to reason reliably, yet on many questions repeated sampling yields both correct and incorrect answers, revealing an underlying fragility in how final decisions are formed. We study whether this fragility can be exploited through implicit reasoning steering: using natural-language text to bias a model toward a designated answer without explicit instructions, triggers, or direct answer cues. Our approach, Concept Chaining, generates a short connection paragraph that links question entities to a target option through one or two intermediate concepts. We then continue pretraining a victim model on these connection paragraphs and evaluate whether its answer preference shifts on the original multiple-choice questions. Our results show that indirect, natural-looking text can systematically steer model predictions while remaining substantially less inferable than direct paraphrases, which shows that reasoning brittleness is not merely an evaluation artifact: it creates a practical channel through which latent biases can be amplified by ordinary-looking text to covertly redirect model decisions.

Problem

• implicit reasoning steering
• reasoning brittleness
• large language models
• concept chaining
• answer bias

Innovation

• Implicit Reasoning Steering
• Concept Chaining
• Reasoning Brittleness
• Natural-Language Biasing
• Covert Model Manipulation