Developing Story
DeepSearch / Multimodal Deep Search Agents
Research trend around multimodal and web-based 'deep search' agents, including new frameworks (Neos, Glasseek) addressing visual reasoning integration, training transparency, and benchmark contamination issues.
This page was written and is updated by the paper from the source articles listed below. Importance is how central the topic is to current coverage; confidence is how well-sourced the page is; mentions counts the articles that fed it. How the wiki works →
DeepSearch agents are language/multimodal agents that solve complex information-seeking tasks through iterative web exploration and tool use. Recent research has focused on both improving their capabilities and addressing reliability concerns in how they are trained and evaluated.
Neos is a multimodal deep search framework designed to interleave search and visual reasoning, learning both how to acquire visual evidence and how to use it to guide subsequent search decisions, aiming to address weaknesses in long-horizon multimodal tasks (arXiv:2610.17043, October 14).
Glasseek presents what its authors call a 'transparent recipe' for training and evaluating DeepSearch agents, addressing three challenges: opaque trajectory-construction pipelines that are hard to audit, mismatches between training-time context compression and inference, and benchmark contamination from agents retrieving live benchmark content from the open web during evaluation (arXiv:2610.16998, October 14).
## Why It Matters
As enterprises and consumer products increasingly deploy agentic search systems, the reliability, auditability, and contamination-resistance of evaluation methods for these systems are becoming strategically significant, particularly for vendors competing on agent capability claims.