DPDM routes agent decisions through one frozen conditional predictor instead of prompting
A paper proposes DPDM, which pairs a language model encoder with a frozen pretrained predictor so agents learn from deployment history without any weight updates.
The problem the paper targets
The authors of “Decoupled Predictive Decision Making: A Shared Conditional Predictor for Agents” (arXiv 2611.02075, published 2026-11-01 and listed on Hugging Face daily papers on 2026-10-31) start from an observation about agents in deployment. Many agent decisions depend on empirical relationships that deployment observations reveal and that a task description alone may leave unresolved. An agent that reasons only from the description and a few examples can miss what its own history would show.
Their answer is to treat learning from experience as a separate capability, handled by a model pretrained for that job, rather than as something the language model has to do inside its prompt.
How DPDM works
Decoupled Predictive Decision Making (DPDM) formulates heterogeneous agent decisions as conditional prediction with a shared pretrained learner. The work is split between two models:
- A language model acts as the semantic encoder. It produces representations of current and historical records.
- A general-purpose conditional predictor takes that represented evidence and predicts task-defined targets.
A decision rule then maps the predictions to choices. Tool judgments, candidate preferences and reasoning assessments all go through the same evidence-and-query interface, which is what lets one predictor serve different kinds of decisions.
Neither model is trained on the deployment data. According to the abstract, pretraining supplies reusable inference capability, and deployment evidence instantiates problem-specific predictions without updating either model’s neural parameters. The authors describe two complementary sources of capability: semantic knowledge acquired through language pretraining and empirical inference acquired through structured-data pretraining. In the experiments a single frozen LimiX-2 predictor serves all five task categories.
Reported results
All figures below are the authors’ own, as stated in the abstract. The semantic encoder for the headline numbers is Qwen3.5-27B.
- Across 50 benchmarks, mean accuracy rises from 78.80% with few-shot prompting to 84.92% with DPDM, each under its own evaluated history budget.
- With identical histories and queries for both, DPDM reaches 82.33%.
- On the BFCL and ToolSandbox tasks of ToolPRMBench, absolute accuracy gains are 34.26% and 28.87%, again at the methods’ respective evaluated history budgets.
- Accuracy improves as history grows.
- With identical imbalanced histories, mean balanced accuracy improves by an absolute 14.51% across five reasoning-step assessment tasks and four encoders.
The second figure is the one to read alongside the first. The 84.92% headline compares each method at its own history budget, while the matched comparison, with the same histories and queries on both sides, gives 82.33% against the 78.80% few-shot baseline.
The authors frame the result as making learning from experience an independently improvable foundation-model capability: better conditional predictors could improve decisions across agent tasks without retraining the language model.
What the abstract leaves open
The abstract does not list the 50 benchmarks beyond the two ToolPRMBench tasks, and it does not give the history budgets for each method or say how long the histories get. It does not spell out the five task categories or name the four encoders used in the balanced-accuracy comparison. LimiX-2 itself is not described. There is no figure for the latency or compute cost of running a second model next to the language model, and no detail on how the decision rule is set for each task. No code repository is listed on the paper page.
The paper page lists 8 authors and the organization Stable AI.
Source: Decoupled Predictive Decision Making: A Shared Conditional Predictor for Agents (arXiv 2611.02075)