SteerSearch: Interactive Search Reasoning with Mid-Response Steering
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.
ProofPaper ↗
Key points
- 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 also investigate Contrastive Context Reward (CCR), an auxiliary signal for aligning post-intervention reasoning with the updated goal.
Sources (1)
- [1]SteerSearch: Interactive Search Reasoning with Mid-Response SteeringarXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 16, 10:38 PM
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.
Search-augmented language models increasingly serve as long-horizon research agents, yet most open training methods assume a fixed user goal throughout search.
Extractive summary: sentences quoted from the sources.
Before this
- Oct 16, 2026axolotl-ai-cloud/axolotl v0.21.0
- Oct 16, 2026Self-Evolving Coding Agents via Harness-Policy Optimization
- Oct 16, 2026DiscoveryRSI: Recursively Learning to Discover Valuable Hypotheses
- Oct 15, 2026Accelerating the search for antimicrobial peptides with Google Antigravity
- Oct 15, 2026Learning to Think and Explore in 3D World
- Oct 15, 2026Data World: Scaling Data Environments from Scratch