VisibilityOS
notableReviewed Sat Oct 31

ITER uses prior agent search history to improve iterative evidence retrieval

Chen and colleagues introduced ITER, a dense retriever trained with agent interaction trajectories that incorporates the main question, earlier subqueries and pre-search reasoning when retrieving evidence for subsequent search steps.

Context

Chen and colleagues introduced ITER, a dense retriever trained with agent interaction trajectories that incorporates the main question, earlier subqueries and pre-search reasoning when retrieving evidence for subsequent search steps.

What changed

Instead of treating each agent-generated subquery independently, ITER learns from interaction history and uses previously useful documents as redundancy negatives. The revised preprint evaluates the method across six agent backbones.

Why it matters

Multi-step agentic discovery is path-dependent: the same current subquery can be evaluated in light of earlier questions and visited evidence. Source visibility measurement should consider complete search trajectories.

Visibility OS interpretation

Visibility OS treats ITER as evidence about controlled iterative retrieval architecture, not as proof that ChatGPT Search, Google AI Mode or Perplexity deploy this method.

What remains uncertain

How well does the technique transfer to live web indexes and commercial agents? How sensitive are gains to the choice of trajectory supervision and search tool?

Evidence claims

Emerging Evidencesupports

In the authors' revised ITER preprint, interaction-aware retrieval improved on the LRAT baseline by average relative 6.9% on InfoSeek-Eval and 15.4% on BrowseComp-Plus across six agent backbones.