Decision Head
An architecture that enables a model to complete classification in a single forward pass by pooling the final hidden state of the sequence and applying a lightweight classification head, avoiding an autoregressive decoding loop.
Timeline (last 90 days)
Jev, Cloudflare's Clef, and the Decision Head integrated by Unsloth use pooling plus a compact Decision Head to do a single forward pass, with no decoding loop.
In an A/B test on the same GPU, the traditional generative prompt had an average inference latency of 3,679.8ms, while the decision model (Decision Head) had 191.4ms.
In multi-task scenarios, the decision model averages 175.28ms for one question at a time and 191.35ms for two questions at once, a marginal cost of about 16.07ms; attaching 5 classification judgments increases latency by at most 60-80ms.
When given input-irrelevant junk messages, the model's maximum confidence drops from 99.9% in normal cases to 54.3%, with the probability spread across the options.
In a customer-service ticket routing scenario, for a ticket involving a duplicate charge, the model gave 99.97% confidence for the billing category and 89.31% confidence for high priority.
In an NGS quality anomaly alert scenario, the model gave only a 0.63% pass probability, 77.21% confidence for human review, and 84.7% confidence for the moderate risk level.
Generative models cannot produce calibrated class probabilities, whereas decision models can output a calibrated probability distribution for threshold-based judgment.
All Facts (7)
Jev, Cloudflare's Clef, and the Decision Head integrated by Unsloth use pooling plus a compact Decision Head to do a single forward pass, with no decoding loop.
50%UnverifiedIn an A/B test on the same GPU, the traditional generative prompt had an average inference latency of 3,679.8ms, while the decision model (Decision Head) had 191.4ms.
50%UnverifiedIn multi-task scenarios, the decision model averages 175.28ms for one question at a time and 191.35ms for two questions at once, a marginal cost of about 16.07ms; attaching 5 classification judgments increases latency by at most 60-80ms.
50%UnverifiedWhen given input-irrelevant junk messages, the model's maximum confidence drops from 99.9% in normal cases to 54.3%, with the probability spread across the options.
50%UnverifiedIn a customer-service ticket routing scenario, for a ticket involving a duplicate charge, the model gave 99.97% confidence for the billing category and 89.31% confidence for high priority.
50%UnverifiedIn an NGS quality anomaly alert scenario, the model gave only a 0.63% pass probability, 77.21% confidence for human review, and 84.7% confidence for the moderate risk level.
50%UnverifiedGenerative models cannot produce calibrated class probabilities, whereas decision models can output a calibrated probability distribution for threshold-based judgment.
50%