Logical Judgments Under Pressure: Diagnosing Syllogistic Stability with Learned Soft Prefixes

This paper investigates the robustness of logical reasoning in large language models (LLMs) by testing whether learned contextual pressure—in the form of "soft prefixes"—can force models to abandon correct logical judgments. The author asks whether these interventions act as logical edits, fixed-symbol forcing, or broad answer biases, and whether these effects generalize to unseen logical forms and different prompt interfaces.

The researcher prepends trainable, continuous vector sequences (soft prefixes) to syllogistic reasoning prompts while keeping the underlying model weights frozen. By using exactly labeled syllogistic benchmarks, the study ensures that the correct logical answer remains fixed. The author evaluates three models—Qwen3.6-35B-A3B MoE, Qwen3-8B, and Gemma 4 31B—across two types of logical reversals: unsatisfiable-to-satisfiable and valid-to-invalid. To interpret the opaque prefixes, the study employs a diagnostic suite including matched random controls, rephrased prompts, and score-based modeling to determine if the model's behavior reflects a general preference for a specific answer or a more complex logical operation.

The study finds that learned soft prefixes are highly effective at redirecting correct judgments, with flip rates ranging from 40% to 90% depending on the model and condition. These interventions significantly outperform random controls, often by 37 to 99 percentage points. The diagnostic analysis reveals that the dominant mechanism is a broad preference for a specific answer meaning rather than fixed-symbol forcing or a transferable logical operation. While aggregate flip rates are high across all models, the underlying response patterns differ; score models based on answer class and baseline margins approximate the behavior of Gemma more accurately than that of the Qwen models, suggesting that aggregate metrics can mask distinct internal stability profiles.