When a software agent searches for information in multiple steps, each step's query may be too short to pick the most useful documents. The agent might find a document that matches the words of a sub-query but not the user's original goal.
ACARe is a compact 4-billion-parameter language model that scores documents by considering not just the sub-query, but also the original question and the agent's reasoning about what to search for next. It can work with any search system, including web APIs, and can handle cases where the agent hides its reasoning.
Across three benchmarks, ACARe improved retrieval quality by about 1 to 2 percentage points over a standard reranker while using fewer searches. It also boosted answer accuracy by about 8 to 9 percentage points over using no reranker. This means agents can find better evidence with less effort, helping smaller agents perform nearly as well as larger ones.