2000 character limit reached
ITER: Interaction-Aware Retrieval for Agentic Search
Published 28 Aug 2026 in cs.IR | (2608.27912v1)
Abstract: Deep-research agents answer complex user questions through an iterative sequence of search steps, where the agent autonomously formulates sub-queries to retrieve the evidence needed at each stage. However, existing retriever training typically relies only on the sub-query and its corresponding search results at the current step as training signals, leaving the information accumulated from previous interactions largely underutilized. We introduce iter, an agent interaction-aware dense retriever trained using agent trajectory learning signals. iter represents each query by incorporating not only the current sub-query, but also the main question and preceding sub-queries, and is trained using trajectory-relative learning signals derived from the agent's interactions. Across six agent backbones from three model families, iter consistently outperforms the existing agent-trajectory-trained dense retriever, LRAT, achieving an average improvement of 7.5% on InfoSeek-Eval and 13.5% on BrowseComp-Plus. iter also demonstrates stronger cross-agent robustness than AgentIR, a deep-research retriever that relies on external LLM-judge signals and the agent's pre-search reasoning. Ablations further show that the main question and previous sub-queries provide the most robust query representation, while previously visited and useful documents, used as redundancy negatives in subsequent searches, provide the strongest trajectory-relative supervision. Code is available at https://github.com/ielab/ITER.
Paper Prompts
Sign up for free to create and run prompts on this paper.
Top Community Prompts
Explain it Like I'm 14
1. What is this paper about?
This paper introduces ITER, a new search system designed for AI agents that answer difficult questions.
A normal search engine usually looks only at the latest search words. However, research agents often search several times:
- They start with a big question.
- They create a smaller search question.
- They read some documents.
- They notice what information is still missing.
- They create another search question.
The authors argue that the search system should remember this history. Otherwise, it may repeatedly show documents the agent has already read instead of finding new evidence.
ITER is designed to help the agent find new and useful information at every step.
2. What questions are the researchers asking?
The paper mainly investigates these questions:
- Can a search system work better if it remembers the main question and earlier searches?
- Can information about the agent’s past actions teach the search system which documents are useful?
- Can ITER avoid showing the same documents again and again?
- Does ITER work well with different AI agents, not just the agent used during training?
- Which kinds of history are most helpful: the main question, earlier searches, the agent’s reasoning, or previously read documents?
The central idea is that a document’s usefulness depends on what the agent has already learned. A document might be relevant to the topic but still not be useful if it repeats information the agent already has.
The authors call the new information provided by a document its marginal gain.
For example, imagine researching basketball players. A document explaining Kobe Bryant’s career may be relevant, but if the agent has already read it, another copy of the same information has little value. A new document about Steve Nash could be much more useful because it helps answer the part of the question that is still unresolved.
3. How did the researchers conduct the study?
Training ITER with search histories
The researchers collected records of AI agents searching for answers. Each record, called a trajectory, shows:
- the original question,
- the smaller questions the agent searched,
- the documents it saw,
- the documents it opened,
- and what the agent learned from those documents.
They used an AI research agent to search through thousands of questions. From these searches, they kept more than 20,000 trajectories that ended with correct answers.
Giving the search system more context
Instead of giving ITER only the latest search query, they usually give it three pieces of information:
- the main question, which explains the overall goal;
- the current sub-query, which describes what the agent needs right now;
- the previous sub-queries, which show what the agent has already tried.
This is similar to asking a student, “What are you trying to solve, what are you working on now, and what methods have you already tried?” That information helps avoid repeating the same approach.
The researchers also tested adding the agent’s pre-search reasoning, meaning the agent’s explanation of why it was creating the next query.
Teaching ITER which documents to prefer
The researchers labelled documents according to how the agent interacted with them.
- A positive document was opened after a search and judged to contain useful information.
- A redundancy negative was useful but had already been read earlier. It was relevant, but it no longer offered much new information.
- A hard negative was opened but later judged to be unhelpful.
- A weak negative appeared in the results but was never opened, so its usefulness was uncertain.
This is like training a student to choose sources for a project. One source might be a good source that has already been used, another might look promising but turn out to be wrong, and another might simply never have been checked.
ITER uses a mathematical training method called contrastive learning. In everyday language, this means the system repeatedly compares:
“Put this useful document near the query, and put these less useful documents farther away.”
The system gives especially strong warnings against ranking previously used documents too highly, because they are often redundant.
Testing ITER
The researchers tested ITER on two difficult question-answering benchmarks:
- InfoSeek-Eval, which was similar to the data used for training.
- BrowseComp-Plus, which used different questions and a different collection of documents, making it an out-of-domain test.
They tested ITER with six different AI agents from three model families. This helped determine whether ITER was useful only for one specific agent or could work more generally.
They compared ITER with systems including:
- BM25, a traditional keyword-based search method;
- a basic pretrained dense retriever;
- LRAT, an earlier retriever trained from agent search records;
- AgentIR, another system that uses the agent’s reasoning to guide search.
4. What did the researchers find?
ITER found better answers
With the main question, current query, and previous queries included, the smaller ITER model achieved:
- 80.0% task success on InfoSeek-Eval, compared with 72.7% for LRAT;
- 46.6% task success on BrowseComp-Plus, compared with 43.4% for LRAT.
On BrowseComp-Plus, ITER also found more of the documents containing the necessary evidence. Its evidence search recall was 0.636, compared with 0.602 for LRAT.
In simple terms, ITER was more likely to locate the information needed to solve the question.
The training method helped even without extra history
The researchers also tested ITER using only the current sub-query. Even in this limited version, it performed better than LRAT:
- InfoSeek-Eval success rose from 72.7% to 76.7%.
- The average number of tool calls fell from 19.6 to 18.3.
This shows that the way ITER learns from search trajectories is useful by itself. The improvement is not only caused by giving ITER more text about the search history.
Previous searches and the main question were especially useful
Adding the main question and previous sub-queries produced further improvements:
- InfoSeek-Eval success increased from 76.7% to 80.0%.
- BrowseComp-Plus success increased from 43.7% to 46.6%.
This suggests that these two types of information help the retriever understand both the big goal and the search paths that have already been explored.
The agent’s reasoning had mixed effects
Adding the agent’s pre-search reasoning sometimes helped and sometimes hurt.
For example, on BrowseComp-Plus, it increased success from 46.6% to 47.6%. However, on InfoSeek-Eval, it reduced success from 80.0% to 79.0%.
This means that longer explanations from the agent are not always better. They may contain useful clues, but they may also distract the retriever or make it focus too narrowly on the agent’s current ideas.
ITER worked with many different agents
ITER performed better than LRAT in all 12 comparisons involving the six agent backbones and the two benchmarks. Its average improvement was:
- about 5.4 percentage points on InfoSeek-Eval;
- about 4.3 percentage points on BrowseComp-Plus.
This is important because it shows that ITER was not useful only with the particular AI agent that produced its training data.
The paper also reports that ITER generally performed better than AgentIR when tested with new, unseen agents. This suggests that ITER’s use of search history may transfer more reliably than relying mainly on one agent’s private reasoning style.
A larger retriever was not always better
The authors tested a larger 4-billion-parameter version of ITER. Bigger models did not consistently produce better results. Performance improved in some cases but decreased in others.
However, one important pattern remained: including the main question and previous searches helped even when using the larger model.
This suggests that better information about the search situation may matter more than simply making the model larger.
Search systems often repeat old results
The researchers found that repeated documents were a serious problem:
- From the second search onward, about 52.1% of the top 10 results had appeared earlier.
- More than half of previously visited documents appeared again in later searches.
- When they reappeared, about 70.1% were placed within the top three results.
A search system that looks only at the current query cannot tell whether a document is new or has already been examined. ITER uses the interaction history to make this distinction.
5. Why are these findings important?
Complex questions often cannot be answered with one search. They require collecting facts from several sources and connecting those facts together.
Without memory, an AI research agent may:
- waste time opening the same documents;
- miss evidence that it has not seen before;
- become stuck on an incorrect possible answer;
- use more search steps than necessary.
ITER helps by treating search as a continuing conversation rather than a series of unrelated searches. It tries to find documents that are not only relevant, but also useful given everything found so far.
6. Possible impact of the research
If systems like ITER are developed further, they could make AI research agents:
- more accurate when answering complicated questions;
- better at finding evidence from several sources;
- less likely to repeat the same searches;
- faster and more efficient;
- more reliable when used with different AI models.
The research may be useful for tools that perform literature reviews, investigate news, answer technical questions, or prepare research reports.
However, ITER does not solve every problem. It still depends on the agent making sensible searches and correctly judging whether documents are useful. Also, the experiments were performed on particular benchmarks and fixed document collections, so more testing would be needed on real-world web searches.
Overall, the paper’s main message is simple:
An AI search system should remember what has already been searched and learned, so that each new search adds fresh evidence instead of repeating the past.
Knowledge Gaps
Knowledge gaps, limitations, and open questions
The paper leaves the following issues unresolved:
- Generalization beyond two fixed-corpus benchmarks is unknown. ITER is evaluated only on InfoSeek-Eval and BrowseComp-Plus, so its effectiveness on open-web search, other domains, languages, temporal information needs, and non-Wikipedia corpora remains untested.
- The contribution of de-duplicated retrieval is not isolated. The training trajectories are collected with a de-duplicated search interface, but the paper does not provide a controlled comparison showing whether ITER’s gains arise from the retriever, the de-duplicated candidate presentation, or their interaction.
- Performance when deployed with standard, non-de-duplicated search is unclear. Training uses a specialized interface that separates previously returned documents from new results, whereas evaluation is described as using unfiltered top- rankings. The robustness of ITER to conventional retrieval interfaces and repeated results is not established.
- The definition of “marginal gain” is not directly measured. The paper formulates utility as , but operationalizes it mainly through visits and an LLM-generated Relevant/Not Relevant judgment. It does not quantitatively validate whether the learned ranking correlates with actual incremental information or answer improvement.
- LLM-based relevance labels may introduce systematic bias. Positive and hard-negative labels are produced by Qwen3-30B-A3B-Thinking-2507, and the paper does not report human-validation agreement, label error rates, sensitivity to verifier prompts, or whether the verifier favors particular writing styles or document formats.
- The use of final-answer correctness to filter training trajectories may create selection bias. Only trajectories whose final answers match references are retained, excluding unsuccessful trajectories that could contain valuable hard negatives, recovery patterns, or evidence of retrieval failure.
- The effect of training on unsuccessful trajectories is unexplored. It remains unknown whether incorporating failed searches, incorrect agent reasoning, or trajectories with incomplete evidence would improve robustness or instead introduce harmful noise.
- The supervision may be confounded by agent browsing behavior. Documents are treated differently based on whether the agent visits them, but visitation depends on the agent’s query formulation, ranking position, browsing budget, and reasoning quality. The paper does not disentangle document utility from the agent’s selection policy.
- Previously visited relevant documents are assumed to be redundancy negatives. A document that was useful earlier may still be necessary later for cross-checking, resolving contradictions, or supporting a different subclaim. The paper does not evaluate when demoting such documents harms evidence completeness.
- The distinction between useful and redundant evidence is coarse. Documents are labeled as relevant or irrelevant, but the method does not model partial overlap, complementary facts, contradictory evidence, source reliability, or the number and type of claims supported by each document.
- Unvisited documents are treated as weak negatives without estimating their true utility. Documents never opened by the agent may be highly relevant but overlooked, and the paper does not quantify the resulting false-negative rate or test alternative methods for handling unlabeled documents.
- Documents returned earlier but never visited remain unlabeled. This creates a potentially important missing-data category, yet the paper does not investigate semi-supervised, positive-unlabeled, or counterfactual approaches for exploiting these documents.
- The chosen negative weights are insufficiently justified. The values are selected for redundancy, hard, and weak negatives, but there is no systematic search, calibration analysis, or demonstration that these weights transfer across datasets, agent backbones, or training scales.
- The positive-instance weighting relies on reasoning length as a proxy for usefulness. Longer post-visit reasoning may reflect verbosity, model uncertainty, or prompt behavior rather than greater document utility. The paper does not compare length-based weighting with factuality, claim coverage, calibrated confidence, or human utility judgments.
- The history representation is limited mainly to textual query history. Although the formal history includes retrieved documents, visited documents, and interpretations, the default model uses only the main question and sub-queries. The paper does not determine whether structured representations of claims, unresolved entities, or evidence graphs would outperform serialized text.
- The method’s sensitivity to long interaction histories is not reported. Query inputs can be up to 8,192 tokens, but the paper does not analyze truncation frequency, performance as trajectory length grows, or which historical interactions are lost when the context limit is exceeded.
- The ordering and serialization of previous sub-queries are not studied. It remains unclear whether recency, chronological order, semantic clustering, or a learned history summarizer is the most effective way to represent prior search directions.
- The mixed effect of pre-search reasoning is unexplained. Reasoning augmentation improves some BrowseComp-Plus results but substantially harms certain cross-agent settings. The paper does not identify whether the problem is reasoning quality, distribution shift, prompt-format mismatch, verbosity, or contamination by incorrect intermediate beliefs.
- Cross-agent robustness is evaluated with a limited set of models. Five unseen backbones provide useful evidence, but the study does not test agents with different tool protocols, planning strategies, browsing policies, context windows, prompting styles, or proprietary APIs.
- The comparison with AgentIR is not fully controlled. ITER and AgentIR differ in encoder scale, query inputs, training signals, and possibly data and implementation details. The reported superiority therefore does not isolate the effect of interaction-aware supervision from other methodological differences.
- Retriever scale is investigated only at 0.6B and 4B parameters. The non-monotonic scaling result cannot establish a general scaling law, and larger encoders, parameter-efficient tuning, distillation, or equal-compute comparisons are not examined.
- The impact of retrieval latency and computational cost is underreported. Dense encoding, long history inputs, and larger retrievers may impose substantial inference and memory costs. Average tool calls alone does not measure wall-clock latency, GPU usage, indexing cost, or total energy/computational efficiency.
- The paper does not analyze how retrieval changes agent reasoning and answer composition. Improvements in search recall and task success are reported, but the causal path from ranking changes to query reformulation, document visitation, evidence synthesis, and final-answer correctness is not examined.
- Evidence search recall and evidence visit recall remain relatively low. On BrowseComp-Plus, many annotated evidence documents are still not retrieved or visited. The paper does not characterize which evidence types, reasoning hops, or question categories account for these failures.
- The relationship between retrieval metrics and end-to-end success is not fully established. Some configurations improve recall without improving task success, and the paper does not provide per-question correlation or causal analyses explaining when additional retrieved evidence helps or distracts the agent.
- Question-level heterogeneity is not analyzed. The results are averaged over benchmarks, leaving open whether ITER benefits mainly entity-linking, multi-hop, temporal, comparative, or compositional questions, and whether it fails systematically on particular query structures.
- Statistical testing is limited. The paper reports unadjusted McNemar tests for selected paired comparisons, but does not provide confidence intervals, corrections for multiple comparisons, effect sizes, power analyses, or significance tests for retrieval-recall and efficiency metrics.
- Reproducibility of agent trajectories may be limited. Generation settings are fixed, but the paper does not report the number of random seeds, trajectory variability, or whether conclusions remain stable across repeated agent runs and nondeterministic serving conditions.
- The method’s robustness to noisy, stale, adversarial, or conflicting documents is untested. Because the training signal rewards documents judged useful by the agent, ITER may inherit or amplify agent susceptibility to misinformation, source-quality differences, and persuasive but incorrect content.
- No explicit source-quality or trustworthiness objective is included. The retriever optimizes novelty and interaction-relative usefulness, but not authority, citation quality, diversity of sources, or resistance to redundant low-quality evidence.
- The method is not evaluated under different browsing budgets. All agents use a maximum of 50 tool-calling turns, so it is unknown whether ITER’s redundancy avoidance remains beneficial under very small budgets or whether it sacrifices verification under generous budgets.
- The effect of the top- candidate pool and top- presentation size is not studied. Different candidate-pool sizes, ranking cutoffs, snippet lengths, or access policies could change both trajectory statistics and the usefulness of interaction-aware training.
- The paper does not test online or continual adaptation. ITER is trained offline from fixed trajectories, leaving open whether it can update safely from live agent feedback, adapt to changing corpora, or avoid reinforcing early retrieval errors.
- Potential information leakage and benchmark overlap are not deeply examined. Although question splits are stated to be disjoint, the paper does not report checks for paraphrased questions, duplicated evidence, corpus overlap, or pretraining contamination across the evaluated models and benchmarks.
Practical Applications
Immediate Applications
- Deep-research and enterprise question-answering systems — Software, knowledge management
- Integrate ITER as the retrieval component in agents that answer multi-step questions, produce literature reviews, investigate competitors, or compile evidence across many documents.
- The workflow would pass the main question, current sub-query, and previous sub-queries to the retriever, while down-ranking documents already examined and promoting documents likely to provide new evidence.
- This is deployable now using the released implementation and model resources, particularly for fixed or controlled document collections.
- Dependencies and assumptions: The agent must maintain a trajectory log containing query history and document identifiers. Performance may depend on corpus quality, document deduplication, context-window limits, and the agent’s ability to interpret retrieved evidence.
- More efficient retrieval-augmented generation (RAG) pipelines — Software and cloud AI
- Replace one-shot or query-only dense retrieval in iterative RAG systems with history-conditioned retrieval.
- Applications include legal research assistants, technical-support copilots, internal policy assistants, and report-generation tools that repeatedly search for missing evidence.
- By reducing repeated results, ITER could lower unnecessary document reads and tool calls; the paper reports fewer average steps in several evaluations.
- Dependencies and assumptions: Efficiency gains require the agent to recognize when information is redundant. The system should retain access to previously visited documents because a previously seen document may still need to be reopened.
- Evidence discovery and citation-support tools — Academia, journalism, policy analysis
- Build research assistants that track which sources have already been examined and deliberately search for complementary evidence, counterarguments, or unresolved facts.
- A practical product could display a trajectory map showing the main question, explored sub-queries, visited sources, and newly discovered sources, helping users audit how an answer was assembled.
- ITER’s emphasis on marginal information gain is especially relevant to multi-hop fact checking and evidence aggregation.
- Dependencies and assumptions: The approach does not itself guarantee source truthfulness, unbiased coverage, or correct citation attribution. Human review, source-quality ranking, and provenance tracking remain necessary.
- Internal enterprise search with exploration memory — Corporate knowledge systems
- Apply ITER to repositories containing manuals, tickets, design documents, compliance policies, and project records.
- The retriever could prevent an employee-facing assistant from repeatedly returning the same familiar documents and instead surface less obvious documents addressing the remaining information gap.
- Potential products include investigation dashboards, incident-response search assistants, and project-history copilots.
- Dependencies and assumptions: Document identifiers must remain stable, access-control filtering must be applied before ranking, and the system must avoid using interaction histories to expose confidential information across users or teams.
- Research and development literature exploration — Academia and industrial R&D
- Use ITER in systems that search scientific papers, patents, standards, and technical reports over multiple rounds.
- After an initial search identifies a topic, subsequent searches can focus on unexplored methods, datasets, competing results, or missing implementation details rather than repeatedly returning the same highly cited papers.
- A workflow could combine ITER with metadata filters, citation graphs, and structured note-taking.
- Dependencies and assumptions: The paper evaluates fixed corpora and benchmark questions; deployment on dynamic scholarly databases requires robust indexing, updated embeddings, metadata integration, and safeguards against missing or poorly represented research areas.
- Customer-support and troubleshooting agents — Service operations
- Use interaction-aware retrieval to search manuals, knowledge-base articles, previous tickets, and diagnostic procedures as a support agent progressively narrows a problem.
- The system can demote troubleshooting steps already attempted and prioritize documentation addressing the unresolved symptoms.
- This could reduce repetitive responses and improve escalation summaries.
- Dependencies and assumptions: The history must distinguish between documents merely shown and documents actually read or used. Incorrect interaction logs could cause the system to suppress relevant information prematurely.
- Personal research and daily-life information assistants — Consumer software
- Implement the method in assistants that help users compare products, plan travel, understand regulations, research health information, or investigate purchases across multiple searches.
- The assistant could maintain a visible “already checked” source list and seek new evidence rather than looping over the same pages.
- Dependencies and assumptions: Consumer deployment requires privacy-preserving history storage, transparent source presentation, freshness checks, and strong disclaimers in high-risk areas such as medicine, finance, and law. ITER improves retrieval behavior but does not replace expert advice or fact verification.
- Search-engine and agent-evaluation instrumentation — Industry and academia
- Add trajectory-aware metrics such as evidence search recall, evidence visit recall, repeated-result rates, marginal document gain, and tool calls per answer to existing evaluation pipelines.
- These metrics can diagnose whether an agent fails because it cannot find evidence, fails to open useful evidence, or repeatedly revisits already examined material.
- Dependencies and assumptions: Reliable evaluation requires document-level relevance judgments and complete logs of retrieval, visits, and agent interpretations. The paper’s reported gains should not be assumed to transfer unchanged to every corpus, agent, or live web environment.
Long-Term Applications
- Adaptive, self-improving research agents — Software and autonomous knowledge work
- Train retrieval systems continuously from production trajectories, using successful document visits as positives and previously consumed or unhelpful documents as trajectory-relative negatives.
- Future agents could learn not only topical relevance but also which evidence is useful at different stages of an investigation.
- This may support autonomous market intelligence, scientific surveying, regulatory monitoring, and strategic analysis.
- Dependencies and assumptions: Production feedback is noisy: a document may be visited because of curiosity, interface position, or agent error. Human validation, debiasing, privacy controls, and safeguards against feedback loops are required before using raw trajectories for continual training.
- Multi-agent collaborative research — Software, robotics, and distributed systems
- Multiple agents could share a common interaction-aware retrieval state, recording which queries and sources have already been explored by other agents.
- One agent might investigate historical evidence, another technical feasibility, and another counterarguments, while the retriever allocates new searches to unexplored regions.
- A future tool could maintain a shared evidence workspace with deduplicated results and agent-specific access histories.
- Dependencies and assumptions: This requires conflict resolution between agents, provenance-aware state management, synchronization at scale, and methods for distinguishing genuinely complementary evidence from merely different copies of the same source.
- Active information acquisition and decision support — Healthcare, finance, energy, and public policy
- Extend the marginal-gain formulation so that retrieval chooses documents, reports, or measurements that most reduce uncertainty about a decision.
- Potential applications include clinical literature review, financial due diligence, energy-market analysis, climate-policy assessment, and emergency-response planning.
- Rather than merely retrieving relevant material, the system could prioritize evidence that fills known gaps or tests competing hypotheses.
- Dependencies and assumptions: High-stakes use requires calibrated uncertainty, validated domain-specific relevance judgments, current and authoritative data, explainability, and human approval. The current paper measures answer success and retrieval recall, not decision quality or safety.
- Interactive educational tutors and learning systems — Education
- Build tutors that remember which explanations, examples, and sources a learner has already encountered and retrieve progressively different material suited to unresolved misconceptions.
- The main question could represent the learning objective, while previous sub-queries represent concepts already explored.
- A future system might deliberately retrieve alternative explanations, worked examples, or diagnostic exercises instead of repeating the same lesson.
- Dependencies and assumptions: Learner interactions are not equivalent to agent judgments of relevance. Educational deployments would need learner modeling, pedagogical evaluation, age-appropriate safeguards, and mechanisms to avoid hiding foundational material merely because it was previously displayed.
- Robotic agents operating in partially known environments — Robotics
- Adapt interaction-aware retrieval from documents to multimodal memories, maps, object databases, and task demonstrations.
- A household or warehouse robot could condition future searches on prior observations and actions, reducing repeated inspection of known locations while seeking information about unexplored areas.
- The analogous training signal would treat successfully used observations as positives and redundant or failed observations as differentiated negatives.
- Dependencies and assumptions: This requires research beyond text retrieval, including temporal state representation, sensor uncertainty, real-time latency, embodied exploration, and safe handling of stale or contradictory observations.
- Dynamic-web deep-research systems — Search engines and public information services
- Deploy ITER-like retrieval over changing web indexes, where the agent must continually discover new pages while retaining access to earlier sources.
- Search interfaces could separate newly discovered results from previously returned results and provide a “new evidence” ranking mode.
- This could improve investigative journalism, public-service information portals, and automated policy monitoring.
- Dependencies and assumptions: The experiments use fixed corpora, whereas the live web introduces spam, duplication, personalization, link rot, adversarial SEO, changing content, and legal or licensing constraints. Robust source-quality and temporal-relevance models would be necessary.
- Retrieval models trained with richer interaction feedback — Long-term IR research
- Extend ITER beyond binary relevance and reasoning-length proxies by incorporating explicit user feedback, dwell time, document comparison, answer citations, contradiction detection, and downstream task outcomes.
- This could yield retrievers optimized for usefulness, novelty, diversity, and evidence coverage rather than similarity alone.
- Dependencies and assumptions: The paper shows that post-visit interpretations and negative tiers can be informative, but also finds that some history components—such as visited documents and interpretations in the query representation—can reduce evidence search recall. Future systems therefore need careful feature selection, better labels, and evaluations across domains.
- Standardized trajectory-aware retrieval infrastructure — Academia, industry, and policy
- Develop shared benchmarks, logging schemas, and governance standards for evaluating agentic retrieval across models, corpora, and sectors.
- Such standards could require reporting answer accuracy, new-evidence discovery, source coverage, repeated-result rates, tool usage, latency, cost, and robustness across unseen agent backbones.
- This would make it easier to compare retrievers such as ITER and identify whether improvements arise from better ranking, better agent behavior, or benchmark-specific adaptation.
- Dependencies and assumptions: Standardization depends on releasing sufficiently detailed trajectories without exposing personal, proprietary, or copyrighted data. Evaluation should also include adversarial and high-stakes scenarios, since the paper’s benchmarks do not establish broad real-world reliability.
Glossary
- Agent backbone: The underlying LLM that controls an agent’s reasoning and tool use. “Across six agent backbones from three model families”
- Agent trajectory: The recorded sequence of an agent’s queries, document interactions, and reasoning steps. “ITER, an agent-trajectory-trained dense retriever”
- Agentic search: An iterative search process in which an autonomous agent formulates queries, examines evidence, and decides what to search next. “Agentic search produces a new, more structured form of interaction within the search loop.”
- Ablation: An experiment that removes or changes one component of a system to measure its contribution. “Ablations further show that the main question and previous sub-queries provide the most robust query representation”
- BM25: A probabilistic sparse-retrieval ranking algorithm based primarily on term frequency and document frequency. “BM25 and the zero-shot Qwen3-Embedding models”
- Candidate coverage: The extent to which a retrieval process exposes potentially useful documents to the agent. “To increase candidate coverage without removing access to earlier results”
- Context-conditioned retrieval: Retrieval that uses contextual information beyond the current query to select documents. “Context-conditioned retrieval uses information beyond the current query to guide retrieval.”
- Contrastive objective: A machine-learning training objective that encourages desired query–document pairs to score higher than competing pairs. “ITER is optimized with a weighted contrastive objective.”
- Cosine similarity: A similarity measure based on the angle between two vectors, commonly used to compare embeddings. “adjacent sub-queries have a mean cosine similarity of .657.”
- Cross-agent robustness: The ability of a retriever trained with one agent to perform effectively with different agent architectures or models. “ITER also demonstrates stronger cross-agent robustness than AgentIR”
- Dense encoder: A neural model that maps text into continuous vector representations for semantic comparison. “Adjacent sub-queries remain semantically close under a dense encoder”
- Dense retriever: A retrieval model that represents queries and documents as dense vectors and ranks them by vector similarity. “ITER, an agent-trajectory-trained dense retriever for this setting.”
- De-duplicated search: A retrieval setting that separates previously returned documents from newly retrieved documents to emphasize unseen candidates. “We therefore collect trajectories in a de-duplicated search setting”
- Evidence aggregation: The process of combining information from multiple retrieved sources to answer a complex question. “requiring multi-step reasoning and evidence aggregation”
- Evidence search recall: The proportion of annotated evidence documents retrieved at least once during an agent trajectory. “Evidence search recall is the fraction of annotated evidence documents retrieved at any point in an agent trajectory.”
- Evidence visit recall: The proportion of relevant evidence documents that the agent actually opens or examines. “Evidence visit recall is the fraction of those documents that the agent visits through the get_document tool.”
- Exact McNemar test: A statistical test for comparing paired binary outcomes, such as correctness on the same questions under two systems. “we report exact McNemar -values on per-question answer outcomes”
- Hard negative: A non-relevant training example that appears plausible or relevant enough to have attracted the agent’s attention. “Hard negatives: documents visited before step and judged irrelevant.”
- History-conditioned query: A query representation constructed using the current query together with information from earlier interactions. “ITER constructs a history-conditioned query representation”
- Information need: The specific information required to resolve a user’s question or a subproblem within it. “the current sub-query specifies the immediate information need”
- In-batch negative: A negative training document taken from another example in the same mini-batch. “documents from other training instances also serve as in-batch negatives.”
- Interaction-aware retrieval: Retrieval that conditions document ranking on an agent’s prior search and document-interaction history. “We formulate interaction-aware retrieval”
- Learning to rank: Methods that train models to order retrieved items according to their relevance or utility. “techniques such as query reformulation, query expansion, and learning to rank”
- Marginal gain: The additional useful information contributed by a document beyond what has already been collected. “We refer to this history-dependent utility as the marginal gain of the document.”
- Macro-averaged: Averaged by first computing a metric separately for each question or unit and then giving each unit equal weight. “Both are macro-averaged over questions.”
- Multi-hop information seeking: Question answering that requires connecting evidence across multiple retrieval or reasoning steps. “InfoSeek-Eval contains 300 multi-hop information-seeking questions”
- Negative tier: A category of negative training examples distinguished by the strength or type of evidence that they should rank below a positive example. “ITER organizes negative documents into three tiers according to the agent's interactions with them”
- Out-of-domain (OOD): Referring to evaluation data that differs from the data distribution used for training. “making it our out-of-domain benchmark.”
- Pre-search reasoning: The reasoning an agent generates before issuing a search query. “AgentIR augments the current sub-query with pre-search reasoning”
- Query expansion: The process of adding terms or information to a query to improve retrieval results. “techniques such as query reformulation, query expansion, and learning to rank”
- Query reformulation: Rewriting a search query to better express the information need. “Agents typically refine an earlier sub-query to address a remaining information gap.”
- Query representation: The textual or vector form of a query supplied to a retrieval model. “The default query representation combines the main question , current sub-query , and previous sub-queries”
- Relevance feedback: Information derived from user judgments or interactions that indicates which retrieved results are useful. “web search, relevance feedback and click logs reveal which results users find helpful”
- Retrieval-augmented generation (RAG): A method that supplies a LLM with externally retrieved evidence before or during text generation. “Retrieval-augmented generation supplements language-model generation with evidence retrieved from external corpora”
- Retrieval backend: The retrieval implementation or model used to produce candidate documents. “collected with Tongyi-DeepResearch across four retrieval backends”
- Retrieval granularity: The level at which a retrieval system selects textual units, such as documents, passages, or sections. “Other studies examine how retrieval granularity, ranking models, query transformations, and lexical or dense matching affect deep-research agents”
- Sparse retriever: A retrieval model that represents text using mostly zero-valued lexical features, such as term-frequency vectors. “We compare ITER against sparse and dense retrievers.”
- Task success rate: The fraction or percentage of questions for which the agent produces a correct answer. “End-to-end effectiveness is measured by task success rate (SR)”
- Trajectory-relative supervision: Training supervision whose labels or weights depend on an item’s position and interactions within an agent trajectory. “trajectory-relative supervision comprising step-specific positives and tiered negatives”
- Weak negative: A document treated as a negative example despite limited evidence about its usefulness because the agent never examined it. “Weak negatives: documents returned at step but never visited anywhere in the complete trajectory.”
- Zero-shot: Describing a model used without task-specific training examples for the target task. “the zero-shot Qwen3-Embedding models”
