gedi (generative discriminator)
**GeDi (Generative Discriminator)** is the **controllable generation technique that uses class-conditional language models as discriminators to guide text generation toward or away from specified attributes** — developed by Salesforce Research as a method to steer any language model's output in real-time by using smaller "guide" models that score candidate tokens for their alignment with desired properties like topic relevance, safety, or sentiment.
**What Is GeDi?**
- **Definition**: A generation-time control method that uses class-conditional language models (trained on attribute-labeled text) to compute per-token guidance signals that steer a base model's generation.
- **Core Innovation**: Treats small fine-tuned language models as Bayesian classifiers that score each candidate next token for its alignment with desired attributes.
- **Key Advantage**: Works with any frozen base model — no base model modification needed, attribute control is applied purely at decoding time.
- **Publication**: Krause et al. (2021), Salesforce Research.
**Why GeDi Matters**
- **Plug-and-Play Control**: Add attribute control to any base model without retraining or fine-tuning it.
- **Real-Time Steering**: Guidance is computed per-token during generation, enabling dynamic control.
- **Multi-Attribute**: Multiple GeDi guides can be combined for simultaneous control over multiple attributes.
- **Detoxification**: Particularly effective at steering generation away from toxic content while maintaining fluency.
- **Efficiency**: Guide models are small (124M parameters), adding minimal computational overhead.
**How GeDi Works**
**Training**: Train small class-conditional LMs on text labeled by attribute (e.g., "toxic" vs. "non-toxic"). Each class-conditional model learns language patterns specific to that attribute.
**Inference**: At each generation step:
1. Compute next-token probabilities from the base model.
2. Compute next-token probabilities from the desired-class guide model.
3. Compute next-token probabilities from the anti-class guide model.
4. Use Bayes' rule to weight base model probabilities toward desired class.
**Guidance Strength**: A control parameter adjusts how strongly the guide influences base model generation — from subtle bias to strong enforcement.
**Applications**
| Application | Desired Class | Anti-Class | Effect |
|-------------|--------------|------------|--------|
| **Detoxification** | Non-toxic | Toxic | Safe generation |
| **Topic Control** | On-topic | Off-topic | Relevant content |
| **Sentiment** | Positive | Negative | Upbeat text |
| **Formality** | Formal | Informal | Professional tone |
**Comparison with Alternatives**
| Method | Base Model Change | Control Granularity | Overhead |
|--------|-------------------|-------------------|----------|
| **GeDi** | None (frozen) | Per-token | Small guide model |
| **PPLM** | Gradient updates during generation | Per-step | Backpropagation per step |
| **RLHF** | Full fine-tuning | Global behavior | Training cost |
| **Prompting** | None | Instructions only | No overhead |
GeDi is **an elegant solution for real-time attribute control in text generation** — proving that small, specialized guide models can effectively steer any base model's output through Bayesian per-token weighting without requiring base model modification.