ITER: Interaction-Aware Retrieval for Agentic Search
Haodong Chen, Shuai Wang, Yu Yin, Shengyao Zhuang, Guido Zuccon, Teerapong Leelanupab
cs.IR
2026-08-28
ITER encodes search history into a dense retriever and demotes already-read evidence. Across six agents it beats LRAT by 7.5% on InfoSeek-Eval and 13.5% on BrowseComp-Plus.
Deep-research agents answer hard questions by issuing a sequence of sub-queries, reading documents, and searching again. The retriever behind that loop usually sees only the latest sub-query. It has no record of what the agent already searched for, or which pages it already opened.
The University of Queensland team measured this on 26,482 Tongyi-DeepResearch trajectories released by LRAT. Adjacent sub-queries have mean cosine similarity 0.657 under a current-query-only dense encoder; non-adjacent pairs from the same trajectory still sit at 0.575, versus 0.237 for random cross-trajectory pairs. From the second search onward, 52.1% of the top-10 results have already appeared. By the eighth search, only 3.9 of 10 documents are new. Visited documents come back in a later search 54.9% of the time; 44.3% of those reappearances land at rank 1, and 70.1% in the top three.
A document can stay on-topic for the current sub-query and still add nothing new. The retriever keeps promoting it anyway.
ITER changes two things: the query encoder sees interaction history, and the training labels are relative to where the agent sits in the trajectory.
The default query concatenates three fields. The main question keeps the overall task. The current sub-query names the immediate gap. Previous sub-queries mark directions already tried. The first step has an empty history. The experiments also tried stuffing visited document text, post-visit interpretations, and pre-search reasoning into the query; the first two hurt evidence recall.
Training trajectories were collected under a de-duplicated search interface. The tool retrieves 100 candidates, drops documents returned earlier, and shows the agent 10 unseen hits. Repeated documents sit in a returned-earlier list and remain openable via getdocument. Tongyi-DeepResearch-30B ran this setup on 10,000 InfoSeek training questions against BM25 and Qwen3-Embedding at 0.6B, 4B, and 8B, producing 40,000 trajectories. They kept the 20,893 whose final answers matched the references. Of 67,934 positive pairs, 26.7% are delayed visits: the document appeared earlier, and the agent opened it only after a later search.
Positives and negatives are timed. A document visited after the current search and before the next one, and judged useful by Qwen3-30B-A3B-Thinking-2507, is a positive. Negatives come in three tiers:
Each group has one positive and nine negatives, sampling up to three redundancy and three hard negatives, filling the rest with weak ones. Positives also get an instance weight from post-visit reasoning length. The loss is weighted contrastive learning with temperature 0.02. The backbone is Qwen3-Embedding-0.6B, the same as LRAT, fully fine-tuned for two epochs in FlagEmbedding at 1e-6, batch 32. Queries truncate at 8,192 tokens, documents at 512. A 4B run uses the same recipe.
Two fixed-corpus deep-research benchmarks. InfoSeek-Eval has 300 multi-hop questions on Wiki-25-Dump (11.2 million passages of at most 512 tokens). Questions are disjoint from training, so this is in-domain. BrowseComp-Plus has 830 questions over 100,195 documents; neither questions nor corpus appear in training. Every retriever returns an unfiltered top-10 with 64-token snippets. Agents get at most 50 tool-calling turns.
On the matched agent Tongyi-DeepResearch-30B, default ITER at 0.6B (main question + current + previous sub-queries) reaches 80.0 success on InfoSeek-Eval against 72.7 for LRAT and 76.7 for the current-sub-query-only ITER variant. On BrowseComp-Plus the three numbers are 46.6, 43.4, and 43.7. Evidence search recall moves from 0.602 to 0.636. Adding pre-search reasoning lifts BrowseComp to 47.6 and drops InfoSeek to 79.0.
BM25 already scores 77.3 on InfoSeek, above LRAT. Default ITER beats BM25 by 2.7 points there. On BrowseComp BM25 sits at 31.1 against ITER's 46.6; the gap shows up on the longer, harder trajectories.
Across six backbones from three families, default ITER beats LRAT in all 12 backbone-benchmark pairs, averaging +7.5% relative on InfoSeek-Eval and +13.5% on BrowseComp-Plus, with seven McNemar-significant gains. On the five unseen backbones it also beats scale-matched AgentIR on task success. AgentIR often retrieves more gold evidence; ITER's agents visit more of it and answer more questions. The largest single jump is Qwen3.6-27B: InfoSeek 68.7 to 79.3, BrowseComp 28.1 to 34.5.
Query ablations: adding the main question lifts BrowseComp from 43.7 to 45.4. Adding previous sub-queries reaches 46.6 and InfoSeek 80.0. Putting visited document text into the query drops BrowseComp to 41.1 and recall to 0.541. Dropping redundancy negatives drops BrowseComp from 46.6 to 40.8 and recall from 0.636 to 0.561, a larger hit than dropping hard negatives. Uniform weights (1,1,1) land BrowseComp at 43.9. Pushing the redundancy weight to 5.0 drops InfoSeek to 75.0.
Scaling 0.6B to 4B is mixed. Default InfoSeek falls from 80.0 to 77.0; BrowseComp rises from 46.6 to 48.0. A larger encoder does not replace history conditioning.
Most deep-research stacks still call BM25 or an offline dense retriever as if each sub-query were independent. ITER's claim is narrower: rank at this step should depend on what the agent has already searched and read. The change is small. Concatenate structured history on the query side. Treat already-read useful documents as high-weight negatives on the training side. The serving interface stays ranked top-k.
The practical recipe is equally specific. Do not put visited document text or post-visit summaries into the query; they fight the redundancy negatives. Pre-search reasoning can help BrowseComp with Tongyi and fails to transfer: on gpt-oss-120B, adding it drops BrowseComp from 40.8 to 30.2. For a new agent, use the main question plus previous sub-queries.
This is a progressive improvement, not a new retrieval interface. Gains come from writing interaction state into the same dense encoder. Code and weights are public.
There is no dedicated Limitations section. The conclusion flags three open problems: representing what the agent has learned, transferring interaction signals across agents, and jointly optimizing retrieval and search decisions.
Training trajectories come only from Tongyi-DeepResearch-30B on InfoSeek training questions. Collection used the de-duplicated interface; evaluation uses unfiltered top-10 for every retriever, so train and test search APIs differ. Positives still pass through a Qwen3-30B judge, so the method trades AgentIR's external LLM labels for a different LLM label, not for label-free supervision.
The 0.6B-to-4B drop of 3 points on default InfoSeek is unexplained. BrowseComp-Plus has about 100k documents, two orders of magnitude smaller than the open web, which inflates document repetition. Success metrics also differ: BrowseComp uses a Qwen3 judge, InfoSeek uses exact string match.
Redundancy negatives treat "already useful" as "should rank lower next time." Some multi-hop questions need to revisit the same page. The 26.7% delayed-visit rate already shows that not opening a document immediately does not mean it is useless, yet unvisited hits still enter training as weak negatives at weight 0.3.