FUDGE (Future Discriminators for Generation) is a controllable text generation method that uses a learned discriminator to predict whether a particular continuation of text will satisfy a desired constraint or attribute in the future. Unlike PPLM which uses gradients to modify hidden states, FUDGE directly adjusts token probabilities at each generation step.

How FUDGE Works

• Base Language Model: A pretrained LM generates candidate next tokens as usual.
• Future Discriminator: A separately trained classifier takes a partial sequence and predicts the probability that the completed sequence will have the desired attribute (e.g., ending with a certain word, being about a specific topic, having a particular format).
• Probability Adjustment: At each step, token probabilities from the base LM are multiplied by the discriminator's predictions, boosting tokens that are likely to lead toward compliant completions.
• Decoding: Standard sampling or beam search is applied to the adjusted distribution.

Key Advantages

• Forward-Looking: Unlike methods that only condition on past context, FUDGE's discriminator is trained to predict whether future text will satisfy constraints — enabling better planning.
• Lightweight: The discriminator is small and fast, adding minimal overhead to generation.
• Flexible Constraints: Can enforce hard constraints like "must end with word X" or soft attributes like "should be formal."
• No LM Modification: The base language model remains unchanged.

Comparison with Other Methods

• PPLM: Uses gradients on hidden states — slower and less stable.
• FUDGE: Uses a learned discriminator on surface text — faster and more targeted.
• GeDi: Similar discriminator-based approach but guides generation using contrastive class probabilities.

Limitations

• Requires training a separate discriminator for each desired attribute.
• The discriminator must generalize to unseen partial sequences, which can be challenging.

FUDGE demonstrated that future-aware discriminators provide an effective and efficient mechanism for constrained text generation.