More examples do not automatically make an agent better at using experience: the key question is whether it can infer the relationship between past cases and their outcomes. This paper separates that empirical inference from language understanding, letting a shared predictor learn from deployment history without changing either model’s neural parameters.
Key Findings
- A frozen LimiX-2 conditional predictor handles five categories of decisions, while a semantic encoder represents each query and its historical examples. This gives tool judgments, preferences, and reasoning assessments a common evidence-and-query interface.
- Across 50 benchmarks, DPDM with Qwen3.5-27B scores 84.92% mean accuracy versus 78.80% for few-shot prompting at their respective evaluated history budgets. With identical histories and queries, the comparison is 82.33% versus 78.80%, helping separate the architecture’s effect from simply having more examples.
- Gains are particularly notable on tool-use evaluations: absolute accuracy improvements are 34.26% on BFCL and 28.87% on ToolSandbox at the methods’ respective evaluated history budgets. Performance also rises as the amount of history grows.
- The central design choice is to keep pretrained semantic knowledge and structured-data inference complementary: task-specific experience informs predictions through examples, rather than fine-tuning model parameters.
How It Works
Each task defines its records, target, and answer set. A frozen language model encodes historical records and the current query; a feature adapter prepares their representations, and the conditional predictor estimates the query’s target from labeled history. A decision rule then maps those probabilities to an answer. The shared predictor can serve different decision roles as long as their records and targets fit this interface.
Who It’s For—and the Trade-off
A good fit for researchers building agents that can accumulate feedback across tool selection, candidate ranking, or reasoning assessment, especially when updating model weights at deployment is undesirable. The approach depends on having relevant labeled history and a task that can be expressed with consistent records and targets; it does not remove the need to define those targets or establish that past examples represent future cases.