Perplexity Releases Decider v1.3: A Specialized Decision Model for Efficient Classification

Perplexity Decider v1.3 achieves a 91.6% mean accuracy across 63 classification tasks, outperforming the company's own production classifiers by 6.54 percentage points. Unlike standard generative models that produce free-form text, Decider is a 'decision model'—a general-purpose classifier that returns probability distributions over a set of possible answers for binary, multi-class, or scored questions.

The chart illustrates the performance gains achieved by the new decision model compared to previous production standards.

Architecture of a Decision Model

Decision models sit between standard task-specific classifiers and full-scale generative LLMs. They are designed to handle tasks like judging output quality, routing requests to different models, and detecting prompt injections by providing probabilities for multiple questions in a single forward pass.

By avoiding token-by-token generation, these models share computation across all questions in a request. For example, a single request can simultaneously ask if a summary is faithful to a document and whether that summary is intended for an expert audience, using a shared context (the state) to inform both answers.

“Decision models sit between LLMs and task-specific classifiers. They return probabilities for multiple questions about the same input in a single forward pass, avoiding token-by-token answer generation and sharing computation across questions.”

— Perplexity Decider v1.3: Training, eval, and use cases

Bidirectional Attention and Isolated Questions

To improve robustness, Perplexity adopted bidirectional attention, allowing each token to incorporate context from the entire input rather than just preceding tokens. This design choice was paired with an 'isolated questions' pattern where the state attends to all questions, but each question attends only to the state and its own tokens.

This visual explains the specific architectural choice used to improve the robustness of the decision model.

This specific configuration was chosen over 'everything sees everything' or 'fully cacheable' patterns because it provides the best balance of performance and answer stability. While 'isolated' questions can shift slightly (1.5-2.2% difference) when the set of questions changes, they offer more reliable results than models where questions heavily influence one another.

Training and the 'Locate' Feature

The models are trained using a three-stage recipe: first, a generalist fine-tuning on large-scale corpora; second, a LoRA-based specialization for specific domains like healthcare or law; and third, a vision-specific step. For the smaller Decider Flash v1.3, the larger model acts as a teacher, using KL-based distillation to transfer the teacher's output distribution and confidence levels.

A new 'locate' feature was added via a dedicated training stage. It uses a span head and a special <|locate|> token to extract verbatim text from the state. Unlike standard classification, which selects from predefined options, 'locate' identifies exact character positions for tasks like named entity recognition or extractive question-answering.

This image demonstrates how the model extracts verbatim text, such as named entities, from the internal state.

Economic and Performance Trade-offs

Perplexity reports that Decider v1.3 is significantly more economical than competing models. For one million production requests, Decider v1.3 costs $26.47, while Jev 1.14 costs $68.10 and GLM-5.3-Flash costs $312.21.

The chart highlights the economic efficiency of the new model relative to other production-ready options.

While the model excels in zero-shot evaluations across seven domains—including cybersecurity and finance—there are technical trade-offs. For instance, merging the 'locate' adapters into the main backbone currently causes regressions on decision requests, meaning the adapters must remain unmerged and activated only for specific extraction tasks.