Abstract
This study tests how correct logical judgments respond to learned contextual influences. We prepend a soft prefix to an exactly labeled syllogistic reasoning benchmark while keeping the model fixed. Soft prefixes are opaque continuous vectors; thus, we characterize them through their induced behavior across controlled variations in logical form and interface.
By examining which prefixes succeed and how their effects generalize, we illustrate how learned contextual pressure can override correct judgments and expose limits in a model's logical stability.
In experiments with Qwen3.6-35B-A3B MoE, Qwen3-8B, and Gemma 4 31B, learned prefixes redirect many correct answers and remain effective across unseen forms and interface changes. In repeated tests with Qwen3.6 MoE and Gemma, they outperform paired random controls in all 16 model-direction-split comparisons by 37 to 99 percentage points.
The flip rates for Qwen3.6 MoE remain between 72% and 90% across wording and prompt changes, while Gemma's validity prefixes retain a 54% to 56% flip compared to less than 1% for matched random prefixes.
Diagnostic tests show that the dominant effect is a broad preference for one answer meaning rather than fixed-symbol forcing or a logical operation that transfers reliably between tasks. This bias manifests differently across models.
In both Qwen models, simple score models often predict which judgments will flip but not how far their margins will move, whereas Gemma's overall response is more closely approximated by the same models.
These results indicate that the dominant behavioral effect of successful soft prefixes is a broad answer preference, while the remaining response reveals substantial model-specific differences in logical stability.
Blogger's Review: This study highlights the profound impact of learned context on logical judgments, particularly under pressure. The introduction of soft prefixes reveals the fragility of model logical stability, exposing limitations in understanding and reasoning, warranting further exploration of its implications in practical applications.