Author丨Zheng Jiamei
Editor丨Cen Feng
Perplexity directly changed the way the model handles judgment tasks this time: the same context can be used by multiple questions at the same time, and the model only needs to perform one forward propagation to output the answer probability of each question. The process that used to require repeatedly calling LLM, generating text, and parsing results can now be converged into one model calculation.
In order to allow multiple problems to share calculations without interfering with each other, Perplexity also redesigned the Attention mechanism to allow the shared state to read all problems while limiting the direct exchange of information between problems. v1.3 further adds the Locate capability, which can directly locate the original text span according to the field description, integrating judgment and text extraction into the same architecture.
This set of models has already entered Perplexity’s production system. In an internal security review, it processed 7,000 Agent sessions in a single day, read 1 billion tokens in total, and pushed for 14 rule fixes.
From the perspective of technical architecture, Decider has three closely related design levels: how the input end organizes shared context and independent questions together, how the attention layer controls the information interaction of different questions, and how the training and output layers allow the same model to take into account both classification judgment and text positioning.
Among them, isolated problem attention changes the underlying way of multi-task shared computing, and is also a key to understanding the entire Decider architecture.
01
One context carries multiple judgment tasks
The input of Decider is actually composed of shared state (State) and several question blocks (Question). State saves web pages, documents or Agent execution records, and Question describes specific judgment goals and lists candidate answers. Each candidate is mapped to a predefined label Token, such as A, B, C, and the special <|marker|> Token marks the answer reading position.
During model processing, the hidden state of the corresponding position of the Marker will be extracted, handed over to the decision-making head to calculate the score of each candidate, and then converted into a probability distribution.
The decision head here is modified from the original language model output layer of the basic model: Perplexity retains the weight rows corresponding to candidate labels such as A, B, C, etc., removes the remaining vocabulary output weights, and then trains for the decision-making task. The output layer, originally used to calculate the probability of the next Token, is now specifically used to handle option judgment.
This design separates task definition from output calculation. The caller can describe the question through natural language and limit the answer space through options, and the model is responsible for calculating the probability of each option based on the context.
Sentiment analysis, natural language reasoning, intent detection, entity matching, content moderation, fact checking and scoring can use the same set of input protocols, without the need to build a separate model structure for each task.
Perplexity converts hundreds of public datasets into a unified training format consisting of States, Questions, Options, and Goals. The training target can be either the correct label or the complete answer probability distribution. For situations where multiple judgments need to be performed on the same long text, multiple Questions can be put into the same forward propagation, and the model only needs to process one shared context.
However, calculating the benefits requires an accurate understanding: adding questions will still increase Question Tokens and attention calculations. What Decider saves is mainly the cost of repeated encoding of long text, as well as the cost of token-by-token decoding and external parsing of the generative model. The greater the number of questions and the longer the shared text, the more obvious the reuse value will be.
The output probability also retains room for adjustment of business strategies. The model is responsible for giving the probability distribution of candidates, and the application sets thresholds based on the business costs of false positives and false negatives. Web filtering and agent execution review can use different thresholds without retraining the decision model for each business.
02
How to share information between multiple tasks without interfering with each other
After the input protocol is determined, a new question falls within the model: multiple Questions enter the Transformer at the same time, what kind of information connection should be established between them?
Decider uses generative LLM as its backbone. The original causal attention organizes information in text order. The current position can only read itself and the previous Token. This structure is suitable for generating text step by step, but is not conducive to judgment tasks that require understanding the complete context.
Perplexity therefore uses bidirectional attention instead, allowing Token to read the complete input. Ablation experiments show that this adjustment improves the model's decision-making performance.
Bidirectional attention opens up the full context, but also allows the interaction of information between multiple issues. v1.2 uses full mutual visibility mode, State and all Questions can read each other. In this way, the representation of a certain question may be affected by other questions, and when the caller adjusts the combination of questions, the original judgment may also change accordingly.
The Isolated Question Attention introduced in v1.3 changes this set of connection relationships: State can read all Questions, but each Question can only read State and its own Token, and cannot directly read other questions.
There is a key detail here: State is no longer just a container for saving raw text, it also assumes the role of aggregating information across tasks. It can read the entire set of questions and form a contextual representation containing task conditions in multi-layer attention calculations. Each Question obtains relevant information through State and retains its own task description.
This structure cuts off the direct transfer of local representations between Questions, but does not eliminate all indirect effects. Since State has read the information of other issues, after multi-layer Transformer propagation, the information of Issue A may still enter the subsequent representation of State and then be read by Issue B.
Perplexity controls direct attentional connections rather than making tasks completely independent in a mathematical sense. This difference also explains why fully cacheable mode is not selected in v1.3. The team has tested another structure: State only reads its own content, and Question reads State and its own Token. leifeng.com
This way the same piece of text can be reused in the cache between requests because the State representation is not dependent on the specific question set. But the cost is that State cannot see the task requirements when coding and cannot adjust the contextual representation to the current problem. Perplexity finally chooses isolated question attention, retaining the conditional effect of questions on context.
The team also checked the output stability through label perturbation testing. On 372 multiple-choice questions, changing option order, label mapping, and label names resulted in semantic answer changes for only 0.27% of the questions in Decider v1.3 and from 2.69% to 4.84% in Jev 1.14. The output of Decider v1.3 remains consistent when executed repeatedly with the same input.
Isolated attention has thus become the core architectural design of Decider: it determines how the shared context participates in multiple judgments, and also determines the paths through which information can be exchanged between different problems.
03
From option judgment to text span extraction
The output range of a normal decision header is limited by preset options. When the task requires the model to find a text of unknown length from the original text, labels such as A, B, and C can no longer meet the demand. Decider v1.3 therefore adds Locate, extending the output to text span positioning.
Locate adds <|locate|> Token, Span Head and LoRA adapter. Each field to be extracted has an independent Locate Token, and its embedding is initialized by the Marker representation obtained from previous training. The span header combines the hidden representation of each Token in Locate Token and State to score the starting point, end point and pairing relationship between the two ends of the candidate text fragment, and at the same time generates a unique threshold for the field.
The start and end points scores are responsible for finding text boundaries, while the pairing scores determine whether the two positions can form a valid segment. Finding two high-scoring positions alone does not mean that they must belong to the same reasonable text, so the pairing relationship also needs to be learned independently. The model returns the original character range, and downstream programs can directly read the corresponding content. leifengwan
Training uses a modified version of the InfoNCE objective such that the score of the correct span is higher than the threshold and the threshold is higher than the other candidate spans. When there is no corresponding field in the original text, the model needs to raise the threshold so that the candidate fragments cannot pass the filtering and return an empty list.
Perplexity trained Locate using named entity recognition, PII detection, and extractive question and answer data, with field descriptions ranging from short labels to natural language instructions. Such data organization requires the model to understand the relationship between field meanings and text content, rather than just memorizing fixed field names.
Locate also brings about parameter interference problems. The original decision head learns the probability distribution between candidate options, and the span head learns text boundaries and segment pairings. The two tasks have different requirements for the internal representation of the model.
Perplexity found that merging the Locate adapter into the backbone degraded common decision requests, so it kept the standalone LoRA adapter enabled only for text location tasks.
The main model is trained in stages. The larger version is based on Qwen3.8-27B. It first carries out large-scale general and full fine-tuning and excludes the training division of public evaluation benchmarks. Then it adds domain tasks through LoRA and then performs visual adaptation. The team averaged the text training weights from three different random seeds to improve the overall model performance.
The Flash version is based on Qwen3.5-4B, gaining capabilities through larger Decider distillation. The student model fits the complete probability distribution given by the teacher and learns the teacher's relative judgment of different candidates through KL divergence. Leifeng.com (public account: Leifeng.com)
Compared to learning only the correct label, distribution distillation also preserves the relative confidence of other candidates. Perplexity shows that KL distillation generalizes better than the standard cross-entropy objective on the hold-out task given the same training data.
In the evaluation, Decider v1.3 achieved an average accuracy of 91.6% on 63 classification tasks, 6.54 percentage points higher than Perplexity's original production classifier. In 72 retrieval tasks, the average nDCG@10 of point-wise reranking reached 83.99, 1.81 points higher than Qwen3-Reranker-8B.
The average accuracy across the eight language tasks was 88.16%, and Jev 1.14 was 85.41%. Locate's average score on the 17 out-of-domain tasks was 69.0, and Flash's was 63.6. The production evaluation covers one million requests in the first 18 days after the API went live, and includes 2,318,887 issues after deduplication. Since production requests lacked human annotation, the team used a cutting-edge LLM judging model to generate reference answers.
The consistency of Decider v1.3 with the reference answer is 93.98%, slightly higher than v1.2's 93.68% and Jev 1.14's 93.35%. This data reflects the degree of agreement between the model and the evaluation model, but cannot be directly equated to the true accuracy rate under manual annotation.
04
Judgment, start to have an independent computing architecture
Decider’s technical route revolves around a clear problem: how to make the model efficiently handle a large number of structured judgments while retaining the ability to understand context. It establishes a complete capability system from option judgment to text positioning through input protocols, isolated attention and Locate extensions.
But this architecture still has boundaries. The current public evaluation results show its performance in classification, retrieval rearrangement and text extraction, and production cases also show that Perplexity has used it for internal workflows.
However, the actual benefit of the model still depends on the task type, context length, number of parallel problems, and deployment cost. The existing data are not enough to prove that this architecture is suitable for all AI applications, especially open-ended reasoning and complex planning, and general LLM still has its own advantages.
For the Agent system, Decider indeed provides an idea of division of labor that is worth continuing to verify: let the general LLM handle complex understanding and generation, and hand over clear-bounded classification, screening, and evaluation to specialized models. However, whether the two can form a more suitable collaboration method needs to be further tested based on the latency, resource consumption and task performance in real workloads.
Reference link: https://www.perplexity.ai/en-GB/hub/blog/perplexity-decider
Original articles from Leifeng.com, unauthorized reproduction is prohibited. See the reprint instructions for details.
edit
Resend Email Modify Email




