DPDM Study Reports Higher Decision Accuracy Across 50 Benchmarks

A study of Decoupled Predictive Decision Making (DPDM) found higher decision accuracy across 50 benchmarks, testing an approach that uses deployment histories without updating language-model parameters.

With Qwen3.5-27B as the semantic encoder, DPDM recorded 84.92% mean accuracy, compared with 78.80% for few-shot prompting, a difference of 6.12 percentage points. In a separate test with identical histories and queries, DPDM reached 82.33% accuracy.

DPDM combines frozen language-model representations of current and historical records with a shared pretrained conditional predictor to estimate task-specific outcomes. The architecture uses deployment evidence without changing the language model’s parameters.

On ToolPRMBench, DPDM’s absolute accuracy gains were 34.26 percentage points on BFCL and 28.87 points on ToolSandbox. Across five reasoning-step assessment tasks and four encoders, mean balanced accuracy improved by 14.51% when using imbalanced histories.

The evaluations covered tool judgments, candidate preferences and reasoning assessments.