InfoSeeker: Decision Planner for Information Seeking

• InfoSeeker is a framework that treats information seeking as a sequential decision problem under partial observability, enabling explicit planning of data acquisition.
• It employs hierarchical decomposition, integrated tool selection, and policy factorization to balance exploration with exploitation for robust decision outcomes.
• Practical implementations span domains from healthcare to software development, with empirical benchmarks demonstrating measurable gains in speed, accuracy, and uncertainty management.

Information Seeking Decision Planner (InfoSeeker) denotes a line of systems that treat information acquisition as an explicit component of decision making rather than as a passive prelude to it. In the recent literature, the term names a hierarchical framework for wide-scale web information synthesis, an LLM decision-making framework for partially observable environments, and a longitudinal healthcare companion concept that mediates between patients, clinicians, and information sources over time (Lee et al., 3 Apr 2026, Fang et al., 2 Oct 2025, Cao et al., 16 Feb 2026). This suggests a broader family resemblance: InfoSeeker systems decide what to observe, which tools or interlocutors to consult, how to compress evidence into working memory, and when to stop seeking and commit to action.

1. Conceptual foundations

A central foundation is the treatment of information seeking as a sequential decision problem under partial observability. “Towards Information-Seeking Agents” formalizes a setting in which an agent does not observe the full world $x$ directly, but instead asks a sequence of questions $q_t$, receives answers $a_t = O(x,q_t)$, updates a belief state $f_\theta(h_{:t})$, and optimizes task performance while minimizing information consumed (Bachman et al., 2016). The paper makes the policy, belief state, and value function co-resident in a shared deep architecture, and combines extrinsic rewards with an intrinsic reward based on incremental reconstruction gain. In that formulation, information seeking is not auxiliary; it is the action space.

A second foundation comes from information behavior research. “Automatic Detection of Search Tactic in Individual Information Seeking” models observable actions such as Query, View, Save, Workspace, and Topic as emissions from latent search tactics in a Hidden Markov Model, with the dominant path $$S5 \rightarrow S1 \rightarrow S2 \rightarrow S3 \rightarrow S4$$ aligning with Marchionini’s information seeking process model (Han et al., 2013). This establishes a process view in which concrete search actions are surface traces of hidden tactic-level states. A plausible implication is that InfoSeeker can be framed not only as a planner over documents or tools, but also as a planner over latent tactic transitions.

A third foundation concerns uncertainty. “Risk and Ambiguity in Information Seeking” distinguishes risk, defined as uncertainty about the quality of the currently visited patch, from ambiguity, defined as the opportunity cost of not exploring elsewhere in the search space (Wittek et al., 2016). In that account, exploitation reduces risk and exploration reduces ambiguity, but the sequential structure of information foraging means the two cannot be minimized simultaneously. “Information-Seeking Decision Strategies Mitigate Risk in Dynamic, Uncertain Environments” reaches a complementary conclusion in a dynamic foraging task: reward-seeking strategies produce slightly more reward on average, whereas information-seeking strategies produce more consistent and predictable outcomes, with minimal reward loss (Barendregt et al., 24 Mar 2025). Together, these results position InfoSeeker as a mechanism for regulating exploration, exploitation, and robustness rather than merely maximizing immediate task payoff.

2. Architectural patterns

One major architectural pattern is hierarchical decomposition. “InfoSeeker: A Scalable Hierarchical Parallel Agent Framework for Web Information Seeking” organizes web-scale information synthesis around a strategic Host $\hat{A}$, multiple domain-specialized Managers $\{\tilde{A}\}$, and parallel Workers $\{\bar{A}\}$, explicitly drawing on Herbert Simon’s principle of near-decomposability (Lee et al., 3 Apr 2026). The Host maintains a compressed global context $$\hat{C}_{t-1} = (Q, q_0, y_0, \dots, q_{t-1}, y_{t-1})$$, emits a step-level directive $q_t$, and selects a Manager; the Manager decomposes $q_t$0 into subtasks $q_t$1, dispatches them to Workers, reflects, aggregates, and returns only a concise summary $q_t$2. The design enforces strict context isolation: Workers keep tool traces locally, Managers see Worker outputs but not Host reasoning, and the Host sees only $q_t$3 pairs. This is the architectural core of InfoSeeker as a web decision planner.

A second pattern is explicit integration of information-seeking and task-oriented planning. “Information Seeking for Robust Decision Making under Partial Observability” formulates InfoSeeker as an LLM-based planning framework that alternates between information-seeking actions, information extraction, and task-oriented planning (Fang et al., 2 Oct 2025). The framework treats the LLM planner as a POMDP-like agent with latent beliefs and internal dynamics, and uses exploratory trajectories to validate its understanding, detect environmental changes, or test hypotheses before generating or revising a task plan. In the appendix algorithm, the loop is: seek, extract, plan, execute, and, if necessary, seek again. The paper’s central claim is that prior LLM planners often react to failures without systematically testing their own internal dynamics model; InfoSeeker inserts that testing step into the planning loop itself.

A third pattern is policy factorization. “Learning to Look” defines factorized Contextual Markov Decision Processes in which the action space is partitioned into information-seeking actions $q_t$4 and information-receiving or task-execution actions $q_t$5, with policies $q_t$6 and $q_t$7 and a learned context extractor $q_t$8 (Dass et al., 2024). DISaM trains $q_t$9 first with ground-truth context, then trains the information-seeking side using the task policy as a source of dense intrinsic reward. At test time, an ensemble of encoders estimates uncertainty, and the agent switches between “look” and “act” based on action-distribution disagreement. This suggests an InfoSeeker architecture in which information seeking is neither a preprocessing step nor a monolithic end-to-end policy, but a factorized subsystem whose value is defined by downstream action quality.

A fourth pattern is planner–reasoner–memory composition. “AVIS: Autonomous Visual Information Seeking with LLM Agent” uses an LLM-powered planner $a_t = O(x,q_t)$0, an LLM-powered reasoner $a_t = O(x,q_t)$1, a working memory $a_t = O(x,q_t)$2, and a human-induced transition graph $a_t = O(x,q_t)$3 to orchestrate captioning, VQA, object detection, OCR, image search, web search, and short QA calls (Hu et al., 2023). The planner chooses the next tool $a_t = O(x,q_t)$4 and query $a_t = O(x,q_t)$5; the reasoner compresses raw tool output into $a_t = O(x,q_t)$6 and classifies it as uninformative, useful, or final answer; memory accumulates filtered evidence. This architecture is structurally close to later InfoSeeker formulations: state-constrained tool choice, utility filtering, and explicit memory updates.

3. Decision mechanisms

A defining mechanism of InfoSeeker systems is the explicit treatment of information-seeking actions as first-class actions. “Adaptive Information Seeking for Open-Domain Question Answering” models open-domain QA as a POMDP whose actions are sparse retrieval $a_t = O(x,q_t)$7, dense retrieval $a_t = O(x,q_t)$8, hyperlink retrieval $a_t = O(x,q_t)$9, and answer $f_\theta(h_{:t})$0, with a learned policy selecting which action to execute from the current belief state $f_\theta(h_{:t})$1 (Zhu et al., 2021). The evidence set $f_\theta(h_{:t})$2 is updated by a learned scoring function, and the policy’s stopping rule is simply the choice of the answer action. The paper’s contribution is not just adaptive retrieval, but adaptive switching among retrieval modes and explicit termination.

Another mechanism is self-gated clarification. “Knowing When to Ask” introduces ACTION-RATING, which places need_clarify(q) inside the action space on a shared ordinal scale with navigation actions, so that asking competes directly with acting at each intermediate decision point (Gao et al., 9 Jun 2026). The model scores top-$f_\theta(h_{:t})$3 candidate actions on a $f_\theta(h_{:t})$4–$f_\theta(h_{:t})$5 scale, and clarification is triggered when a need_clarify action exceeds threshold $f_\theta(h_{:t})$6. Two modes then emerge from the rankings themselves: mandatory clarification, where need_clarify is top-ranked because no viable branch exists, and opportunistic clarification, where a navigation action leads but clarification remains above threshold. The paper defines Information-Seeking Effectiveness (ISE) as the fraction of help interactions followed by a correct next navigation step, and reports a rise from $f_\theta(h_{:t})$7 to $f_\theta(h_{:t})$8. It also shows that the mode split and ISE ranking persist when answer quality is degraded, supporting an empirical separation between where an agent seeks help and the quality of the help it receives.

A third mechanism concerns the granularity of feedback during search. “Conversational Decision Support for Information Search Under Uncertainty” introduces SERA, an LLM-based assistant that provides either gist or verbatim feedback during sequential information search (Quan et al., 16 Feb 2026). Gist feedback extracts essential meaning and core relationships between options, whereas verbatim feedback preserves literal details and specific attributes. Across decremental, local optimum, and random information-gain environments, individuals show more accurate decision outcomes and are more confident with SERA support, especially under higher uncertainty. Gist feedback is associated with more efficient integration and shows a descriptive pattern of reduced oversampling; verbatim feedback promotes more extensive exploration. This makes feedback representation itself a decision-planning variable.

A fourth mechanism is ask-or-act calibration in multimodal environments. “StarBench” adds an ask-or-act diagnostic in a turn-based RPG benchmark, where an agent in the tool-assisted regime decides at the start of an episode whether to ask for a brief guidance hint or act immediately (Zhang et al., 21 Oct 2025). The benchmark defines Ask Rate, Effect, and Efficiency, and reports that judicious information seeking correlates with improved success. The distinction between moderate ask rates with substantial uplift and near-always asking with low efficiency is directly relevant to InfoSeeker evaluation: the objective is not maximal help-seeking, but calibrated help-seeking.

4. Domains and instantiations

InfoSeeker appears in healthcare as a longitudinal companion rather than a one-shot advisor. “More than Decision Support: Exploring Patients’ Longitudinal Usage of LLMs in Real-World Healthcare-Seeking Journeys” describes a multi-stage healthcare trajectory—symptom appraisal, information seeking, care navigation decisions, clinical encounter, post-consultation sensemaking and planning, treatment implementation and self-management, and recurrence monitoring and follow-up—and argues that patients use LLMs across behavioral, informational, emotional, and cognitive levels (Cao et al., 16 Feb 2026). On that basis, the paper and its synthesized design notes describe an “Information Seeking Decision Planner (InfoSeeker)” for healthcare as a longitudinal, LLM-powered companion that scaffolds behavior, information, emotions, and cognition over time, carefully managing agency, trust, and risk. The proposed future model is a “longitudinal boundary companion” spanning the patient–clinician boundary, the clinic–everyday life boundary, and the cognitive–emotional boundary.

In software development, the same planning logic appears as source selection under task, risk, and knowledge constraints. “The Evolution of Information Seeking in Software Development” reports that developers increasingly route debugging, code analysis, refactoring, and code comprehension to AI tools, while traditional resources remain preferred for learning new technologies, trustworthy reviewed content, and context-rich issues (Haque et al., 2024). The study states that about $f_\theta(h_{:t})$9 of a software practitioner’s day involves seeking and using information, and its synthesis translates the findings into an inferred decision process over task type, information source, time pressure, risk, organizational policy, and foundational knowledge. This suggests an InfoSeeker for development that does not merely retrieve answers, but chooses among AI, documentation, search, Q&A, and peers, and attaches validation steps such as tests, cross-checking, or peer review.

Collaborative settings supply a further extension. “Toward Collaborative Information Seeking (CIS)” distinguishes communication, contribution, coordination, cooperation, and collaboration, with collaboration defined as joint problem solving that goes beyond the sum of individual contributions and vests authority in the collaborative rather than in an individual entity (0908.0709). Shah’s extended CIS model adds multiple users, communication links, local and shared result spaces, and joint organization of results. A plausible implication is that InfoSeeker in group settings must plan not only what information to obtain, but also when to preserve independent exploration, when to aggregate judgments, and how to support shared authority.

Other instantiations show that the same family of ideas spans web search, visual QA, embodied robotics, tariff classification, and game agents. AISO casts tool choice in QA as POMDP action selection (Zhu et al., 2021); AVIS uses LLM-guided tool sequences in knowledge-intensive visual QA (Hu et al., 2023); DISaM learns when to look and when to act in contextual manipulation (Dass et al., 2024); ACTION-RATING localizes clarification points in hierarchical taxonomy navigation (Gao et al., 9 Jun 2026); StarBench makes ask-or-act a benchmarked multimodal competency (Zhang et al., 21 Oct 2025). This suggests that InfoSeeker is better interpreted as a cross-domain planning pattern than as a single implementation.

5. Empirical performance and benchmarks

The web-oriented InfoSeeker framework reports strong efficiency and effectiveness on wide-scale information synthesis. On WideSearch-en and BrowseComp-zh, the abstract reports approximately $$S5 \rightarrow S1 \rightarrow S2 \rightarrow S3 \rightarrow S4$$0–$$S5 \rightarrow S1 \rightarrow S2 \rightarrow S3 \rightarrow S4$$1 speed-up in inference latency, $$S5 \rightarrow S1 \rightarrow S2 \rightarrow S3 \rightarrow S4$$2 success rate on WideSearch-en, and $$S5 \rightarrow S1 \rightarrow S2 \rightarrow S3 \rightarrow S4$$3 accuracy on BrowseComp-zh (Lee et al., 3 Apr 2026). In the detailed ablation, increasing the Worker pool from 1 to 17 reduces average latency from 911 seconds to 162 seconds, described as an approximately $$S5 \rightarrow S1 \rightarrow S2 \rightarrow S3 \rightarrow S4$$4 speed-up. Tool-call analysis further shows that over $$S5 \rightarrow S1 \rightarrow S2 \rightarrow S3 \rightarrow S4$$5 of tool invocations occur at the Worker layer, while the Host handles only a small number of steps, with average $$S5 \rightarrow S1 \rightarrow S2 \rightarrow S3 \rightarrow S4$$6 for WideSearch-en and $$S5 \rightarrow S1 \rightarrow S2 \rightarrow S3 \rightarrow S4$$7 for BrowseComp-zh. These results substantiate the claim that context isolation and workflow-level parallelization are not merely architectural niceties but measurable performance drivers.

The partially observable LLM InfoSeeker framework reports a large robustness gain. Its abstract states that InfoSeeker achieves a $$S5 \rightarrow S1 \rightarrow S2 \rightarrow S3 \rightarrow S4$$8 absolute performance gain over prior methods on its new benchmark suite without sacrificing sample efficiency, and that it generalizes across LLMs while outperforming baselines on robotic manipulation and web navigation benchmarks (Fang et al., 2 Oct 2025). On the new benchmark, the largest gaps appear in perturbed-dynamics settings, where baselines often collapse and InfoSeeker retains substantially higher success. The paper’s ablations also show that explicitly inserting information-seeking actions matters more than merely adding uncertainty language to prompts.

In open-domain QA and visual QA, adaptive information seeking likewise yields measurable gains. AISO reports $$S5 \rightarrow S1 \rightarrow S2 \rightarrow S3 \rightarrow S4$$9 supporting-passage exact match on HotpotQA with $\hat{A}$0, and on SQuAD Open AISO-large reaches $\hat{A}$1 EM and $\hat{A}$2 F1 with $\hat{A}$3, outperforming baselines with predefined strategies while using far fewer reading steps (Zhu et al., 2021). AVIS reports $\hat{A}$4 and $\hat{A}$5 accuracy on InfoSeek’s Unseen Entity and Unseen Question splits and $\hat{A}$6 on OK-VQA, substantially above sequential baselines using the same tools (Hu et al., 2023). The common empirical pattern is that dynamic tool planning, memory, and output filtering outperform fixed tool pipelines.

Benchmarks that isolate information seeking as a local competency expose further structure. ACTION-RATING on HTS classification reports a regime shift from mandatory to opportunistic clarification and an ISE rise from $\hat{A}$7 to $\hat{A}$8, while a controlled answer channel yields accuracy gains up to $\hat{A}$9 at 10-digit classification depth, explicitly described as an upper bound rather than a deployment estimate (Gao et al., 9 Jun 2026). StarBench, in turn, shows that in tool-assisted combat, moderate ask rates with high uplift can outperform near-default asking, underscoring that the central benchmarked skill is calibrated asking rather than ask volume (Zhang et al., 21 Oct 2025).

6. Limitations and research directions

Across papers, InfoSeeker systems are consistently presented with explicit limitations. The hierarchical web framework depends on backbone LLM APIs, strong models such as gpt-5.1 for Host and Managers, hand-tuned prompts, and finite context windows; its failure cases include entity-type mismatch and context-length overflow, including a BadRequestError: Input tokens exceed limit of 272000 tokens in exhaustive AMD Zen CPU enumeration (Lee et al., 3 Apr 2026). The authors identify learning task decompositions and coordination policies via multi-agent reinforcement learning, and training smaller or specialized Manager and Worker models, as future work.

The partially observable LLM framework reports failure modes in information seeking, information extraction, instruction understanding, and long-horizon planning, especially in more complex block-stacking tasks (Fang et al., 2 Oct 2025). Its benchmark is explicitly described as relatively small and hand-crafted, and the extraction module can produce incorrect or irrelevant insights that then misguide planning. The paper therefore points toward larger uncertain-dynamics benchmarks, more reliable extraction, and more explicit belief modeling or lightweight learned dynamics models.

Healthcare and software studies add socio-technical constraints. The diary study of patient LLM use is limited by a 25-participant, single-country, hospital-centric sample with moderate-to-high AI literacy, no direct analysis of full chat logs, and only the patient side of the patient–provider–AI triad (Cao et al., 16 Feb 2026). The software-development study emphasizes overreliance, shallow understanding, reduced independent thinking, privacy and policy constraints, and reduced peer interaction as salient caveats of AI-assisted information seeking (Haque et al., 2024). In both domains, a plausible implication is that InfoSeeker must be evaluated not only on task outcome but also on trust calibration, agency, skill development, governance, and the distribution of authority between human and machine.

Several papers converge on open questions. These include how to learn or calibrate when-to-ask policies across domains (Gao et al., 9 Jun 2026), how to model richer context and memory in factorized planners (Dass et al., 2024), how to integrate explicit information objectives with reward-sensitive planning under dynamic uncertainty (Barendregt et al., 24 Mar 2025), how to extend prompt-driven planners into more general reasoning tasks with lighter-weight models (Hu et al., 2023), and how to design adaptive systems that match feedback granularity to uncertainty (Quan et al., 16 Feb 2026). Taken together, these directions indicate that InfoSeeker remains an active research program rather than a settled architecture: its central problem is not information retrieval alone, but the joint optimization of observation, interpretation, coordination, and commitment under uncertainty.