---
title: "Adaptive Instructed-Retriever: Frontier-Quality Search at 2x Lower Latency"
description: "Adaptive Instructed-Retriever addresses the tension between retrieval quality and latency for enterprise data agents, especially on multi-hop questions that benefit from sequential search. It combines parallel single-step retrieval with bounded multi-step search, using a fixed upper limit and an adaptive stopping policy that spends extra steps only when they are likely to improve results. The model is trained with online reinforcement learning using CISPO, a reward that balances trajectory quality against search cost, and synthetic multi-hop data alongside Instructed-Retriever-1 data. Across seven held-out internal and external benchmarks, it matches leading third-party and open-source models while answering in 5.8 seconds, reported as more than twice as fast as the compared alternatives, and substantially improves over single-step search. Varying the step penalty produces checkpoints spanning different quality-latency tradeoffs for production workloads."
---

# Adaptive Instructed-Retriever: Frontier-Quality Search at 2x Lower Latency

[Databricks](https://yomu.fyi/company/databricks) · Cheng Li, Jialu Liu, Sean Kulinski, Arnav Singhvi, Wen Sun, Michael Bendersky, Cindy Wang · Sep 9, 2026

**Type:** Benchmark

## Summary

Adaptive Instructed-Retriever addresses the tension between retrieval quality and latency for enterprise data agents, especially on multi-hop questions that benefit from sequential search. It combines parallel single-step retrieval with bounded multi-step search, using a fixed upper limit and an adaptive stopping policy that spends extra steps only when they are likely to improve results. The model is trained with online reinforcement learning using CISPO, a reward that balances trajectory quality against search cost, and synthetic multi-hop data alongside Instructed-Retriever-1 data. Across seven held-out internal and external benchmarks, it matches leading third-party and open-source models while answering in 5.8 seconds, reported as more than twice as fast as the compared alternatives, and substantially improves over single-step search. Varying the step penalty produces checkpoints spanning different quality-latency tradeoffs for production workloads.

## Context

Enterprise data agents need retrieval that is accurate on complex, multi-hop requests without incurring the latency of sequential search on every request. The work targets agents searching large, changing workspaces containing tables, notebooks, dashboards, and documents.

## Approach / What changed

Adaptive Instructed-Retriever combines parallel retrieval with sequential search under a fixed step limit. Online reinforcement learning with CISPO trains the model to stop when evidence is sufficient or continue when more searching is likely to improve quality; varying the step-penalty weight creates checkpoints with different quality-latency tradeoffs.

## Takeaways

- The model can stop early when evidence is sufficient and use additional sequential search only for requests where further retrieval is likely to help.
- Training reuses Instructed-Retriever-1 data, adds synthetic multi-hop questions, and uses CISPO to penalize search steps that do not produce corresponding performance gains.
- Across seven held-out internal and external benchmarks, the model answers in 5.8 seconds while matching leading compared models and improving substantially over single-step search.

**Tags:** [AI Agents](https://yomu.fyi/topic/ai-agents), [Latency](https://yomu.fyi/topic/latency), [Reinforcement Learning](https://yomu.fyi/topic/reinforcement-learning), [Search](https://yomu.fyi/topic/search)

- Source: [Databricks](https://www.databricks.com/blog/adaptive-instructed-retriever-frontier-quality-search-2x-lower-latency)
- Source URL: https://www.databricks.com/blog/adaptive-instructed-retriever-frontier-quality-search-2x-lower-latency
- Ingested by Yomu: 2026-09-09T16:05:21.215Z

[Read original post](https://www.databricks.com/blog/adaptive-instructed-retriever-frontier-quality-search-2x-lower-latency)