DPDM lifts agent accuracy to 84.92% across 50 benchmarks
TL;DR
- DPDM raises mean accuracy across 50 agent benchmarks from 78.80% to 84.92% using a frozen LimiX-2 predictor and Qwen3.5-27B encoder.
- Neither the language model nor the predictor is fine-tuned at deployment; one shared conditional predictor serves all five task categories.
- On ToolPRMBench's BFCL and ToolSandbox tasks, absolute accuracy gains reach 34.26% and 28.87%, respectively.
A single frozen predictor, paired with a frozen language encoder, raised mean agent accuracy from 78.80% to 84.92% across 50 benchmarks. The architecture, called Decoupled Predictive Decision Making, is set out in a paper on arxiv by Lang Mo and co-authors.
DPDM splits agent decisions into two roles. Language models provide semantic representations of current and historical records; a general-purpose conditional predictor predicts task-defined targets from that evidence. Neither model's weights are updated at deployment. "A single frozen LimiX-2 predictor serves all five task categories," the authors write, with Qwen3.5-27B carrying the encoding side.
Some individual tasks move harder than the headline mean. On ToolPRMBench's BFCL and ToolSandbox, "absolute accuracy gains are 34.26% and 28.87%, respectively, at the methods' respective evaluated history budgets." Across five reasoning-step assessment tasks and four encoders, mean balanced accuracy improves by an absolute 14.51% when histories are held identical. With identical histories and queries across the full benchmark set, DPDM lands at 82.33%, still above the 78.80% few-shot baseline.
The abstract reports means, not a per-benchmark breakdown, and does not compare compute cost against few-shot prompting.
Originally reported by paper
Read the original article →Original headline: DPDM Raises Agent Accuracy 6pp Across 50 Benchmarks With One Frozen Predictor