Decoupled Predictive Decision Making: A Shared Conditional Predictor for Agents

Decoupled Predictive Decision Making. A Shared Conditional Predictor for Agents. By Lang Mo and seven others, from Stable AI and Tsinghua University. Published on the first of November.

Many of an agent's decisions rest on empirical relationships: which tool tends to work, which candidate tends to be right. Those relationships are revealed by what the agent observes during deployment, and a task description alone may leave them unresolved. An agent that reasons only from the description is missing the evidence that would settle the question. The point of comparison in the results is few-shot prompting, where the evidence simply goes into the prompt as examples for the language model to read.

Decoupled Predictive Decision Making, or DPDM, recasts heterogeneous agent decisions as conditional prediction, served by one shared pretrained learner. The work is split in two. A language model provides semantic representations of current and historical records. A general-purpose conditional predictor then predicts task-defined targets from that represented evidence. Two complementary sources of capability result: semantic knowledge from language pretraining, and empirical inference from pretraining on structured data.

Pretraining supplies an inference capability that can be reused, and deployment evidence turns it into problem-specific predictions without updating either model's neural parameters. Tool judgments, candidate preferences and reasoning assessments all share one interface of evidence plus query, and a decision rule maps each prediction to a choice. A single frozen LimiX-2 predictor serves all five task categories. Nothing is finetuned per task. What changes from one task to the next is the evidence handed to the predictor and the query put to it, while the learner itself stays fixed. Experience accumulates as records, not as weight updates.

With Qwen3.5-27B as the semantic encoder, DPDM raises mean accuracy across fifty benchmarks from seventy-eight point eight percent with few-shot prompting to eighty-four point nine two percent, each under its own evaluated history budget. Given identical histories and queries, it reaches eighty-two point three three percent. On the BFCL tasks of ToolPRMBench, the absolute accuracy gain is thirty-four point two six percent.

On the ToolSandbox tasks of the same benchmark, the gain is twenty-eight point eight seven percent. Accuracy improves as history grows, which is the behaviour the design is meant to produce: more deployment evidence, better predictions, with no retraining in between. And with identical, imbalanced histories, mean balanced accuracy rises by an absolute fourteen point five one percent, across five reasoning-step assessment tasks and four encoders.

What the decoupling buys is a separate axis of progress. Learning from experience stops being something each prompt has to coax out of a language model, and becomes a foundation-model capability that can be improved on its own. Any advance in conditional prediction then has a shared route into better decisions, across every agent task that uses the same interface.