Exploring the “logit_bias” Feature in the API: Eliminating Em Dashes and Suppressing 106 Tokens—Insights and Sample Code for a “Dash-Free” Output Test

Understanding and Suppressing Em Dash Usage in AI Prompts: A Technical Approach

In the realm of AI prompt engineering, controlling the stylistic output of language models can be a challenging endeavor—especially when dealing with specific characters like the em dash (—). Recently, I explored an innovative method to significantly reduce or eliminate the presence of em dashes in responses generated by GPT-4 models by leveraging the logit_bias parameter in the OpenAI API.

The Challenge

Em dashes are often used for emphasis, interruption, or stylistic separation—yet, for certain applications, their usage may be undesirable. Classic strategies such as tweaking instructions or utilizing memory features often fall short due to the model’s ingrained tendencies. To address this, I turned to the logit_bias parameter, which allows direct influence over token probabilities during generation.

Implementation Details

The core idea was to identify the token IDs corresponding to em dashes and related characters, then assign them a bias of -100—effectively making their selection impossible during response generation. Initial attempts targeting only the exact em dash token were insufficient because the model tends to generate similar symbols like en dashes (–) or hyphens (-) when prompted.

To comprehensively suppress these variants, I adopted an iterative approach:

• First, I identified tokens that include the em dash as part of their representation, such as tokens with “—” attached to other characters.
• Next, I expanded the biasing to encompass tokens representing en dashes and hyphens, recognizing that the model may substitute these for em dashes.
• Finally, after applying biases to over 100 tokens—specifically, 106 tokens to cover all relevant variants—the model’s output consistently avoided any dash-like characters.

This aggressive token biasing did not notably impair response relevance or grammaticality. Responses remained coherent, with the primary difference being the absence of em dashes, thus achieving the stylistic goal.

Sample Evaluation

To test this approach, I used two prompts:

1. Asking for a ‘hot take’ on productivity culture.
2. Requesting pragmatic solutions for political division.

Across multiple model versions, responses biased against dash tokens predominantly aligned with the non-dash versions. For example, the “anti-dash” responses provided clearer, more straightforward language without stylistic dashes, while preserving overall coherence and informativeness.

Practical Guide

I have compiled a Python script that automates this process. The script:

• Retrieves token IDs for