# Compatibility Matrix

> **Auto-generated** by `docs/scripts/generate_compat_matrix.py` — do not edit manually.
> Since it is based on rough dynamic evaluation **it may include some errors---but is useful for quick reference**.

Each cell lists the concrete loss functions supported for the given optimizer-model pair,
or shows **Unsupported** if the model does not satisfy the optimizer's requirements.
Optimizers with both token and text flows appear as two rows, one per flow.

| Optimizer | CLIPTextEncoderHFModel | ClassifierHFModel | EncoderGeminiModel | EncoderHFModel | EncoderOpenAIModel | EncoderVoyageModel | LMHFModel | LiteLLMModel |
|---|---|---|---|---|---|---|---|---|
| **ARCAOptimizer** | `ExternalTriggerPerplexityLoss`, `SimilarityLoss` | `ExternalTriggerPerplexityLoss`, `MisclassCELoss` | **Unsupported** | `ExternalTriggerPerplexityLoss`, `SimilarityLoss` | **Unsupported** | **Unsupported** | `AttentionEnhLoss`, `ExternalTriggerPerplexityLoss`, `FirstTokenNLLLoss`, `PrefillCELoss`, `PrefillCWLoss`, `PrefillDistillationLoss`, `PrefillMellowMaxLoss`, `ResponseHarmfulnessLoss`, `SteeringActivationLoss`, `TriggerPerplexityLoss` | **Unsupported** |
| **AutoPromptOptimizer** | `ExternalTriggerPerplexityLoss`, `SimilarityLoss` | `ExternalTriggerPerplexityLoss`, `MisclassCELoss` | **Unsupported** | `ExternalTriggerPerplexityLoss`, `SimilarityLoss` | **Unsupported** | **Unsupported** | `AttentionEnhLoss`, `ExternalTriggerPerplexityLoss`, `FirstTokenNLLLoss`, `PrefillCELoss`, `PrefillCWLoss`, `PrefillDistillationLoss`, `PrefillMellowMaxLoss`, `ResponseHarmfulnessLoss`, `SteeringActivationLoss`, `TriggerPerplexityLoss` | **Unsupported** |
| **BeamSearchOptimizer** | `SimilarityLoss` | `MisclassCELoss` | `SimilarityLoss` | `SimilarityLoss` | `SimilarityLoss` | `SimilarityLoss` | `FirstTokenNLLLoss`, `PrefillCELoss`, `PrefillCWLoss`, `PrefillDistillationLoss`, `PrefillMellowMaxLoss`, `ResponseHarmfulnessLoss` | `FirstTokenNLLLoss`, `ResponseHarmfulnessLoss` |
| **GASLITEOptimizer** | `ExternalTriggerPerplexityLoss`, `SimilarityLoss` | `ExternalTriggerPerplexityLoss`, `MisclassCELoss` | **Unsupported** | `ExternalTriggerPerplexityLoss`, `SimilarityLoss` | **Unsupported** | **Unsupported** | `AttentionEnhLoss`, `ExternalTriggerPerplexityLoss`, `FirstTokenNLLLoss`, `PrefillCELoss`, `PrefillCWLoss`, `PrefillDistillationLoss`, `PrefillMellowMaxLoss`, `ResponseHarmfulnessLoss`, `SteeringActivationLoss`, `TriggerPerplexityLoss` | **Unsupported** |
| **GASLITEPlusOptimizer** | `ExternalTriggerPerplexityLoss`, `SimilarityLoss` | `ExternalTriggerPerplexityLoss`, `MisclassCELoss` | **Unsupported** | `ExternalTriggerPerplexityLoss`, `SimilarityLoss` | **Unsupported** | **Unsupported** | `AttentionEnhLoss`, `ExternalTriggerPerplexityLoss`, `FirstTokenNLLLoss`, `PrefillCELoss`, `PrefillCWLoss`, `PrefillDistillationLoss`, `PrefillMellowMaxLoss`, `ResponseHarmfulnessLoss`, `SteeringActivationLoss`, `TriggerPerplexityLoss` | **Unsupported** |
| **GBDAOptimizer** | `ExternalTriggerPerplexityLoss`, `SimilarityLoss` | `ExternalTriggerPerplexityLoss`, `MisclassCELoss` | **Unsupported** | `ExternalTriggerPerplexityLoss`, `SimilarityLoss` | **Unsupported** | **Unsupported** | `AttentionEnhLoss`, `ExternalTriggerPerplexityLoss`, `FirstTokenNLLLoss`, `PrefillCELoss`, `PrefillCWLoss`, `PrefillDistillationLoss`, `PrefillMellowMaxLoss`, `ResponseHarmfulnessLoss`, `SteeringActivationLoss`, `TriggerPerplexityLoss` | **Unsupported** |
| **GCGOptimizer** | `ExternalTriggerPerplexityLoss`, `SimilarityLoss` | `ExternalTriggerPerplexityLoss`, `MisclassCELoss` | **Unsupported** | `ExternalTriggerPerplexityLoss`, `SimilarityLoss` | **Unsupported** | **Unsupported** | `AttentionEnhLoss`, `ExternalTriggerPerplexityLoss`, `FirstTokenNLLLoss`, `PrefillCELoss`, `PrefillCWLoss`, `PrefillDistillationLoss`, `PrefillMellowMaxLoss`, `ResponseHarmfulnessLoss`, `SteeringActivationLoss`, `TriggerPerplexityLoss` | **Unsupported** |
| **GCGPlusOptimizer** | `SimilarityLoss` | `MisclassCELoss` | `SimilarityLoss` | `SimilarityLoss` | `SimilarityLoss` | `SimilarityLoss` | `FirstTokenNLLLoss`, `PrefillCELoss`, `PrefillCWLoss`, `PrefillDistillationLoss`, `PrefillMellowMaxLoss`, `ResponseHarmfulnessLoss` | `FirstTokenNLLLoss`, `ResponseHarmfulnessLoss` |
| **HotFlipOptimizer** | `ExternalTriggerPerplexityLoss`, `SimilarityLoss` | `ExternalTriggerPerplexityLoss`, `MisclassCELoss` | **Unsupported** | `ExternalTriggerPerplexityLoss`, `SimilarityLoss` | **Unsupported** | **Unsupported** | `AttentionEnhLoss`, `ExternalTriggerPerplexityLoss`, `FirstTokenNLLLoss`, `PrefillCELoss`, `PrefillCWLoss`, `PrefillDistillationLoss`, `PrefillMellowMaxLoss`, `ResponseHarmfulnessLoss`, `SteeringActivationLoss`, `TriggerPerplexityLoss` | **Unsupported** |
| **PALOptimizer** | `SimilarityLoss` | `MisclassCELoss` | `SimilarityLoss` | `SimilarityLoss` | `SimilarityLoss` | `SimilarityLoss` | `FirstTokenNLLLoss`, `PrefillCELoss`, `PrefillCWLoss`, `PrefillDistillationLoss`, `PrefillMellowMaxLoss`, `ResponseHarmfulnessLoss` | `FirstTokenNLLLoss`, `ResponseHarmfulnessLoss` |
| **PEZOptimizer** | `ExternalTriggerPerplexityLoss`, `SimilarityLoss` | `ExternalTriggerPerplexityLoss`, `MisclassCELoss` | **Unsupported** | `ExternalTriggerPerplexityLoss`, `SimilarityLoss` | **Unsupported** | **Unsupported** | `AttentionEnhLoss`, `ExternalTriggerPerplexityLoss`, `FirstTokenNLLLoss`, `PrefillCELoss`, `PrefillCWLoss`, `PrefillDistillationLoss`, `PrefillMellowMaxLoss`, `ResponseHarmfulnessLoss`, `SteeringActivationLoss`, `TriggerPerplexityLoss` | **Unsupported** |
| **QCGOptimizer** | `SimilarityLoss` | `MisclassCELoss` | `SimilarityLoss` | `SimilarityLoss` | `SimilarityLoss` | `SimilarityLoss` | `FirstTokenNLLLoss`, `PrefillCELoss`, `PrefillCWLoss`, `PrefillDistillationLoss`, `PrefillMellowMaxLoss`, `ResponseHarmfulnessLoss` | `FirstTokenNLLLoss`, `ResponseHarmfulnessLoss` |
| **RASLITEPlusOptimizer** | `SimilarityLoss` | `MisclassCELoss` | `SimilarityLoss` | `SimilarityLoss` | `SimilarityLoss` | `SimilarityLoss` | `FirstTokenNLLLoss`, `PrefillCELoss`, `PrefillCWLoss`, `PrefillDistillationLoss`, `PrefillMellowMaxLoss`, `ResponseHarmfulnessLoss` | `FirstTokenNLLLoss`, `ResponseHarmfulnessLoss` |
| **RandomSearchOptimizer** | `SimilarityLoss` | `MisclassCELoss` | `SimilarityLoss` | `SimilarityLoss` | `SimilarityLoss` | `SimilarityLoss` | `FirstTokenNLLLoss`, `PrefillCELoss`, `PrefillCWLoss`, `PrefillDistillationLoss`, `PrefillMellowMaxLoss`, `ResponseHarmfulnessLoss` | `FirstTokenNLLLoss`, `ResponseHarmfulnessLoss` |
| **SoftPromptOptimizer** | `ExternalTriggerPerplexityLoss`, `SimilarityLoss` | `ExternalTriggerPerplexityLoss`, `MisclassCELoss` | **Unsupported** | `ExternalTriggerPerplexityLoss`, `SimilarityLoss` | **Unsupported** | **Unsupported** | `AttentionEnhLoss`, `ExternalTriggerPerplexityLoss`, `FirstTokenNLLLoss`, `PrefillCELoss`, `PrefillCWLoss`, `PrefillDistillationLoss`, `PrefillMellowMaxLoss`, `ResponseHarmfulnessLoss`, `SteeringActivationLoss`, `TriggerPerplexityLoss` | **Unsupported** |

## Legend

### Access levels required by each optimizer

| Optimizer | Required Mixins | Flows | Access Level |
|---|---|---|---|
| ARCAOptimizer | `LossTokenAccessMixin`, `GradientTokenAccessMixin` | `token` | White-box |
| AutoPromptOptimizer | `LossTokenAccessMixin`, `GradientTokenAccessMixin` | `token` | White-box |
| BeamSearchOptimizer | `LossTextAccessMixin` | `text` | Black-box |
| GASLITEOptimizer | `LossTokenAccessMixin`, `GradientTokenAccessMixin` | `token` | White-box |
| GASLITEPlusOptimizer | `LossTokenAccessMixin`, `GradientTokenAccessMixin` | `token` | White-box |
| GBDAOptimizer | `LossTokenAccessMixin`, `GradientTokenAccessMixin` | `token` | White-box |
| GCGOptimizer | `LossTokenAccessMixin`, `GradientTokenAccessMixin` | `token` | White-box |
| GCGPlusOptimizer | `LossTextAccessMixin` | `text` | Black-box |
| HotFlipOptimizer | `LossTokenAccessMixin`, `GradientTokenAccessMixin` | `token` | White-box |
| PALOptimizer | `LossTextAccessMixin` | `text` | Black-box |
| PEZOptimizer | `LossTokenAccessMixin`, `GradientEmbedAccessMixin` | `token` | White-box (embedding) |
| QCGOptimizer | `LossTextAccessMixin` | `text` | Black-box |
| RASLITEPlusOptimizer | `LossTextAccessMixin` | `text` | Black-box |
| RandomSearchOptimizer | `LossTextAccessMixin` | `text` | Black-box |
| SoftPromptOptimizer | `GradientEmbedAccessMixin` | `token` | White-box (embedding) |

### Concrete loss functions

| Loss | Base Type | Required Parameters |
|---|---|---|
| `AttentionEnhLoss` | `AttentionBasedLoss` | `full_attentions`, `input_slices` |
| `ExternalTriggerPerplexityLoss` | `BaseLoss` | `input_trigger_strs` |
| `FirstTokenNLLLoss` | `TextBasedLoss` | `response_first_token_logprobs` |
| `InputFluencyLoss` | `BinaryLMJudgeLoss` | `input_texts` |
| `MisclassCELoss` | `ClassificationBasedLoss` | `output_class_logits` |
| `PrefillCELoss` | `PrefillBasedLoss` | `prefill_response_logits`, `target_response_toks` |
| `PrefillCWLoss` | `PrefillBasedLoss` | `prefill_response_logits`, `target_response_toks` |
| `PrefillDistillationLoss` | `PrefillBasedLoss` | `prefill_response_logits`, `target_response_logits` |
| `PrefillMellowMaxLoss` | `PrefillBasedLoss` | `prefill_response_logits`, `target_response_toks` |
| `ResponseHarmfulnessLoss` | `BinaryLMJudgeLoss` | `generated_response_strs` |
| `SimilarityLoss` | `EmbeddingBasedLoss` | `output_embeddings`, `target_vectors` |
| `SteeringActivationLoss` | `HiddenStateBasedLoss` | `full_hidden_states`, `target_directions` |
| `TriggerPerplexityLoss` | `TriggerLogitBasedLoss` | `full_logits`, `input_trigger_ids`, `input_slices` |

### Discovered fields per model (via source AST)

| Model | Flow | ModelOutput fields | ModelInput fields |
|---|---|---|---|
| CLIPTextEncoderHFModel | token | `output_embeddings` | `input_attention_mask`, `input_embeds`, `input_prefix_cache_kwargs`, `input_slices`, `input_trigger_ids`, `input_trigger_strs`, `message_targets` |
| CLIPTextEncoderHFModel | text | `output_embeddings` | *(none)* |
| ClassifierHFModel | token | `output_class_logits` | `input_attention_mask`, `input_embeds`, `input_prefix_cache_kwargs`, `input_slices`, `input_trigger_ids`, `input_trigger_strs`, `message_targets` |
| ClassifierHFModel | text | `output_class_logits` | *(none)* |
| EncoderGeminiModel | token | *(none)* | *(none)* |
| EncoderGeminiModel | text | `output_embeddings` | *(none)* |
| EncoderHFModel | token | `output_embeddings` | `input_attention_mask`, `input_embeds`, `input_prefix_cache_kwargs`, `input_slices`, `input_trigger_ids`, `input_trigger_strs`, `message_targets` |
| EncoderHFModel | text | `output_embeddings` | *(none)* |
| EncoderOpenAIModel | token | *(none)* | *(none)* |
| EncoderOpenAIModel | text | `output_embeddings` | *(none)* |
| EncoderVoyageModel | token | *(none)* | *(none)* |
| EncoderVoyageModel | text | `output_embeddings` | *(none)* |
| LMHFModel | token | `full_attentions`, `full_hidden_states`, `full_logits`, `generated_response_ids`, `generated_response_logits`, `generated_response_strs`, `prefill_response_logits`, `response_first_token_logprobs` | `input_attention_mask`, `input_embeds`, `input_prefix_cache_kwargs`, `input_slices`, `input_trigger_ids`, `input_trigger_strs`, `message_targets` |
| LMHFModel | text | `full_ids`, `full_strs`, `generated_response_ids`, `generated_response_logits`, `generated_response_strs`, `prefill_response_logits`, `response_first_token_logprobs` | *(none)* |
| LiteLLMModel | token | *(none)* | *(none)* |
| LiteLLMModel | text | `generated_response_strs`, `response_first_token_logprobs` | *(none)* |