Logical Judgments Under Pressure: Diagnosing Syllog... | AI Research

Logical Judgments Under Pressure: Diagnosing Syllogistic Stability with Learned Soft Prefixes
This research investigates the robustness of logical reasoning in large language models by testing how they respond to "learned context." The author explores whether models that correctly solve syllogistic logic problems can be forced to abandon those correct answers when a specific, optimized sequence of vectors—called a "soft prefix"—is added to the prompt. By keeping the model’s internal weights frozen and only training these prefixes, the study exposes the limits of logical stability and identifies how contextual pressure can override accurate reasoning.

Testing Logical Stability

The study uses syllogistic logic because it provides a clear, objective way to measure accuracy. In these problems, the correct answer is mathematically determined by the premises. The researchers prepend a "soft prefix"—a series of continuous, non-readable embedding vectors—to the input. Because these vectors are not natural language, they cannot be interpreted as instructions. Instead, the researchers treat them as a "stress test" to see if they can force the model to flip its judgment from a correct answer to an incorrect one, even when the underlying logic of the problem remains unchanged.

How the Prefixes Work

To understand what these prefixes are actually doing, the researchers compared them against several controls, including random prefixes and readable text. They tested three models: Qwen3.6-35B-A3B MoE, Qwen3-8B, and Gemma 4 31B. The results showed that these learned prefixes are highly effective at redirecting answers. In all 16 comparisons across different models and logical splits, the learned prefixes significantly outperformed random controls, with success rates (flip rates) ranging from 40% to 90%. This suggests that the prefixes are not just adding noise, but are specifically tapping into the model's internal biases.

Key Findings and Model Differences

The diagnostic tests reveal that the primary effect of these prefixes is a "broad answer preference." Rather than forcing the model to use a specific symbol or performing a complex logical operation, the prefixes seem to nudge the model toward a preferred answer category. However, the study also highlights significant differences between models. For example, simple score models were able to predict the behavior of the Gemma model quite well, suggesting its response to pressure is relatively consistent. In contrast, while these models could predict when the Qwen models would flip their answers, they struggled to predict the intensity of those changes, indicating that different models have unique ways of maintaining—or losing—their logical stability.

Limitations and Interpretations

A central challenge in this research is the "interpretive ambiguity" of the results. Because the prefixes move a minority answer toward a majority answer, it is difficult to determine if the model is performing a sophisticated logical edit or simply defaulting to a more frequent response. The researchers emphasize that these prefixes do not represent a new way to "program" a model, but rather a diagnostic tool. By showing that similar aggregate steering rates can hide very different internal response patterns, the study demonstrates that even when models appear to behave similarly on the surface, their underlying logical stability is fundamentally different.