Explain the arrangement-dependent sign reversal in the post-trained Qwen3.5-9B model

Explain why the post-trained Qwen3.5-9B model produces arrangement-dependent sign reversals in the influence of a risk disclosure across document layouts, with negative effects when filler precedes the focal filing and positive effects when filler follows it.

Background

The paper finds that the post-trained Qwen3.5-9B model behaves differently from the base model: at context lengths of 8,000 tokens and beyond, the disclosure has significantly negative influence when filler precedes the focal filing, but significantly positive influence when filler follows it. This instability is not attributable to weak risk sensitivity, because the post-trained model has the strongest severity discrimination in the model ladder.

The authors explicitly identify the mechanism of this layout-dependent sign structure as unresolved. Determining its cause would clarify how post-training affects the transmission and interpretation of risk information across different document arrangements, and would be relevant to deployment-oriented evaluation of AI financial analysts.