SteerSearch improves interactive search with mid-search user redirection
SteerSearch: Interactive Search Reasoning with Mid-Response Steering
Artificial Intelligence
Summary
When people search for answers using AI models, they sometimes change their questions while the AI is still thinking. The researchers noticed that many models fail to update their answers properly after such changes. They created SteerSearch, a new method that helps the AI adjust its reasoning during the search to match the updated question. Their experiments showed that SteerSearch significantly improves the accuracy of such real-time user corrections without losing the ability to answer normal questions.
What this means in practice
- •For chatbot developers: Build chatbots that can adjust answers dynamically based on user corrections during a conversation, improving interaction quality.
- •For information retrieval teams: Enhance search engines to handle user query changes mid-search without losing context, enabling more flexible and accurate research tools.
- •For customer support platforms: Improve live support systems where agents or customers refine queries mid-exchange, leading to more precise information retrieval.$Commercial implications: This paper enables adaptive AI-driven query refinement features that can be marketed as advanced customer support enhancements.
Authors
Udit Thakur, Baohao Liao, Matteo Nulli, Stefan Vasilev, Seyyed Hadi Hashemi, Shahram Khadivi, Christof Monz
Abstract
Search-augmented language models increasingly serve as long-horizon research agents, yet most open training methods assume a fixed user goal throughout search. We study mid-response steering, where users redirect a model during ongoing reasoning or retrieval. We identify answer-target anchoring, a failure mode in which fixed-query search agents retrieve evidence for the updated request but still answer an earlier question. We introduce SteerSearch, an RL framework that trains on interactive search trajectories with injected interventions to enable context-preserving target reassignment, and SteerSearch-Bench, a benchmark adapted from multi-hop QA datasets. We also investigate Contrastive Context Reward (CCR), an auxiliary signal for aligning post-intervention reasoning with the updated goal. Without CCR, SteerSearch improves average steering accuracy over Search-R1 from 11.1% to 31.3% on Qwen2.5-3B and from 36.6% to 44.6% on Qwen3-4B, while preserving standard single-query performance. Gains also extend to larger Qwen2.5 models and natural multi-turn conversations.