An Interactive Paradigm for Deep Research

# An Interactive Paradigm for Deep Research
Source: [https://arxiv.org/html/2605.24266](https://arxiv.org/html/2605.24266)
###### Abstract

Recent advances in large language models \(LLMs\) have enabled deep research systems that synthesize comprehensive, report\-style answers to open\-ended queries by combining retrieval, reasoning, and generation\. Yet most frameworks rely on rigid workflows with one\-shot scoping and long autonomous runs, offering little room for course correction if user intent shifts mid\-process\. We presentSteER, a framework forSteerable deEpResearch that introduces interpretable, mid\-process control into long\-horizon research workflows\. At each decision point,SteERuses a cost–benefit formulation to determine whether to pause for user input or to proceed autonomously\. It combines diversity\-aware planning with utility signals that reward alignment, novelty, and coverage, and maintains a live persona model that evolves throughout the session\.SteERoutperforms state\-of\-the\-art open\-source and proprietary baselines by up to 22\.80% on alignment, leads on quality metrics such as breadth and balance, and is preferred by human readers in 85%\+ of pairwise alignment judgments\. We also introduce a persona–query benchmark and data\-generation pipeline\. To our knowledge, this is the first work to advance deep research with an interactive, interpretable control paradigm, paving the way for controllable, user\-aligned agents in long\-form tasks\.

Machine Learning, ICML

## 1Introduction

SystemMid\-process steeringAdaptive pause decisionLive persona modelingDeep Research Framework \(Open\-Source\)DeepResearcher\(zheng2025deepresearcher\)✗ ✗ ✗ Search\-r1\(jin2025search\)✗ ✗ ✗ ManuSearch\(huang2025manusearch\)✗ ✗ ✗ Search\-o1\(li2025search\)✗ ✗ ✗ GPT\-Researcher\(gptresearcher\)✗ ✗ ✗ Open Deep Research\(opendeepresearch\)◆ ✗ ✗ ERD\(prabhakar2025enterprisedeepresearchsteerable\)✓ ✗ ✗ Deep Research Framework \(Proprietary Web\-Based\)OpenAI Deep Research\(openaideepresearch\)◆ ✗ ✗ Google Gemini Deep Research\(googledeepresearch\)◆ ✗ ✗ Interactive/Persona\-Aware Reasoning FrameworksPersonaAgent\(zhang2025personaagent\)✗ ✗ ✓ ReasonGraph\(li\-etal\-2025\-reasongraph\)✓ ✗ ✗ HITL CoT MCS\(cai2023human\)✓ ✗ ✗ SteER✓ ✓ ✓

Table 1:Comparison of deep research frameworks \(open\-source and proprietary\) and interactive/persona\-aware reasoning frameworks in terms of mid\-process steering via adaptive pause decisions and live persona modeling\. To our knowledge,SteERis the first benchmarkable deep\-research framework that jointly offers all\.Recent advances in large language models \(LLMs\) have shifted information access from ranked retrieval to systems that generate comprehensive, report\-style answers to complex and open\-ended queries\(du2025deepresearch\)\. These*deep research*systems, spanning proprietary platforms\(googledeepresearch;openaideepresearch;xaideepresearch\)and open\-source frameworks\(gptresearcher;opendeepresearch\), combine iterative retrieval with multi\-step reasoning to synthesize well\-supported outputs\(coelho2025deepresearchgym\)\.Benchmarks such as DeepResearchGym\(coelho2025deepresearchgym\)and DeepResearch Bench\(du2025deepresearch\)have begun to standardize this setting, providing realistic research\-style questions and automated evaluation protocols for long\-form, citation\-heavy reports\. However, these benchmarks evaluate research\-style queries and long\-form reports against generic quality criteria; they do not provide*persona\-conditioned, aspect\-grounded targets*needed to measure alignment and focus when a system is meant to adapt to an individual user\. This leaves a gap for evaluating interactive, user\-steerable research agents, which we address with a persona\-query suite that pairs deep\-research\-worthy queries with query\-conditioned personas and actionable aspect checklists\. On the system side, current deep research agents largely fall into two paradigms: multi\-agent pipelines that divide planning, search, and synthesis\(huang2025manusearch;alzubi2025open;li2025search;zhang2025evolvesearch\), and RL\-trained agents that learn to search and reason effectively\(zheng2025deepresearcher;jin2025search;song2025r1\)\. Yet, regardless of architecture, most systems follow a rigid workflow: one\-time scoping \(often with a single clarification\), followed by a long autonomous run\. If user intent shifts mid\-process, there is little room to course\-correct, resulting in wasted cost and misaligned reports\. This highlights the need for an alternative design where mid\-process interaction is central, not optional\.

Two research threads closely relate to our work\.Personalization and alignmentexamine how to tailor LLM outputs to user intent, from profile\-conditioned generation\(Wuet al\.,[2025](https://arxiv.org/html/2605.24266#bib.bib6)\)to long\-form checklists\(Salemiet al\.,[2024](https://arxiv.org/html/2605.24266#bib.bib3);salemi2025lamp\)and interactive preference elicitation\. While these works show the value of personalization, most assume fixed personas or separate preference modeling from system control, lacking a principled way to determine when to seek input\.Interactive reasoninginvestigates how LLMs ask clarifying questions\(andukuristar;ren2023robots; Wuet al\.,[2024](https://arxiv.org/html/2605.24266#bib.bib5)\), model future turns\(ICLR2025\_97e2df4b\), or learn clarification policies\(DBLP:conf/iclr/Chen0PA25\)\. These methods address alocal clarificationbottleneck — detecting ambiguity in the current request and deciding whether to answer it directly or ask first — rather than an end\-to\-end policy over where to pause in a long\-horizon trajectory, how pausing trades off against exploration, and how user goals should evolve mid\-process\. A complementary thread exposes reasoning traces for human inspection:Interactive Reasoning\(pang2025interactive\)renders chain\-of\-thought as an editable topic hierarchy, andReasonGraph\(li2025reasongraph\)visualizes sequential or tree\-based reasoning paths\. These tools improve transparency and oversight, but providestatic visualizationrather than an adaptive pause mechanism or live persona model that conditions planning, branch utility, and synthesis within a single long\-horizon research tree\. Existing approaches thus either optimize autonomous agents or isolate clarification as a narrow skill\. In contrast, we aim to offer an integrated control paradigm that jointly governs pausing, exploration, and personalization mid\-process\.

We introduceSteER, a framework forSteerable deEpResearch that brings interactive control to long\-horizon research workflows \(Figure[1](https://arxiv.org/html/2605.24266#S1.F1)\)\. The key intuition is that deep research should occasionally*ask*, not just*answer*:SteERuses a cost–benefit formulation at each decision point to determine whether to pause for user input or to proceed autonomously\. To remain both user\-aligned and exploratory, it combines diversity\-aware planning with utility signals that reward alignment, novelty, and coverage\. A live persona is continuously updated based on interactions and conditions all downstream planning, scoring, and synthesis, enabling the system to adjust as user needs evolve\.

Table[1](https://arxiv.org/html/2605.24266#S1.T1)situatesSteERamong deep research frameworks and interactive/persona\-aware reasoning systems\. Most deep research systems still do one\-shot scoping followed by a long autonomous run: once a query is issued, a fixed pipeline executes to completion with no interpretable way to intervene in the research tree or to decide when to solicit guidance \(◆denotes a single upfront scoping/plan\-confirmation step\)\. Concurrent work, EDR\(prabhakar2025enterprisedeepresearchsteerable\), adds user\-initiated mid\-process steering via an exposed task plan, but lacks an adaptive pause policy and live persona modeling\. Interactive/persona\-aware systems address complementary aspects: they maintain evolving user representations or expose reasoning for inspection and manual correction, but they do not implement an adaptive pause policy that decides*where*to pause and*how*to balance autonomy and control\. In contrast,SteERis, to our knowledge, the first benchmarkable deep research framework that jointly supports mid\-process steering, an adaptive pause mechanism, and live persona modeling\.

Our contributions are as follows:

- •We proposeSteER, an interactive deep research framework that supports interpretable, mid\-process control and dynamic user alignment throughout the research loop\.
- •Extensive experiments show thatSteERoutperforms the strongest open\-source and proprietary OpenAI baselines on persona\-tailoredness and report quality, while offering fine\-grained control to tune trade\-offs between alignment and user burden, as well as between under\-exploration and over\-personalization\. A human study further confirms its preference among readers, with significant gains in alignment, focus, and usability\.
- •We introduce a persona–query evaluation suite and a reusable data generation pipeline grounded in prior benchmarks, suitable for future evaluation and training of interactive deep research agents\.

In summary,SteERconsistently outperforms strong open and proprietary baselines, achieving7\.83%​–​22\.80%7\.83\\%\\text\{–\}22\.80\\%higher alignment and leading on general quality metrics such as breadth and balance\. Human readers preferSteERin over 85% of alignment and 83% of focus pairwise comparisons\. To our knowledge,*this is the first work toadvance deep research with an interactive, interpretable control paradigm*\. We believe that this paradigm shift will shape future long\-horizon research agents, enabling decision policies that adapt to individual users and their evolving needs, rather than relying on a single upfront clarification\.

![Refer to caption](https://arxiv.org/html/2605.24266v1/figures/steer.png)Figure 1:Overview ofSteER\. The upper panel shows the end\-to\-end pipeline\. The lower panels zoom into the three core modules:*Planning*,*Pause Decision*, and*Persona Modeling*\.
## 2SteER

### 2\.1Problem Setup and Objectives

We formulate steerable deep research as an interactive planning task\. Given a user queryQQ, the system incrementally constructs a research tree and produces a cited synthesis reportRR\. The goal is to generate a report that is both high\-quality and aligned with the evolving preferences of the user, while keeping the number of interruptions minimal and well\-timed\.

Each user is represented by a personaP=\(ptext,𝒜\)P=\(p\_\{\\text\{text\}\},\\mathcal\{A\}\), whereptextp\_\{\\text\{text\}\}is a natural\-language description combining profile and personality traits \(followingWuet al\.\([2025](https://arxiv.org/html/2605.24266#bib.bib6)\)\), and𝒜\\mathcal\{A\}is a set of aspects the user expects to see addressed in the final report\. We evaluate reports along two complementary dimensions:\(1\) Alignment: the extent to which the report covers the aspects in𝒜\\mathcal\{A\}; and\(2\) Focus: the proportion of content that remains on\-topic with respect to𝒜\\mathcal\{A\}\.

### 2\.2System Overview

Our frameworkSteERtransforms monolithic deep\-research pipelines into an interactive process\. The system is structured around three core components:diversity\-aware exploration,pause decision, andpersona modeling\. At a high level, the framework incrementally builds a research tree that represents possible exploration paths and selectively engages the user at key checkpoints\.

We denote the research tree asT=\(N,E\)T=\(N,E\), where each noden∈Nn\\in Nrepresents a sub\-problem query with partial research results and each edge\(n,n′\)∈E\(n,n^\{\\prime\}\)\\in Eindicates a decomposition into sub\-directions\. The tree expands level by level up to a maximum depthDD, a hyper\-parameter controlling how many layers are explored before synthesis\. At each step, the system operates at afrontier node𝐧⋆\\mathbf\{n^\{\\star\}\}and performs the following actions \(Figure[1](https://arxiv.org/html/2605.24266#S1.F1)\):

1. 1\.Diversity\-aware exploration:Generate candidate follow\-up directions fromn⋆n^\{\\star\}and select a diversified subset of sizeKKto serve as potential expansions \(see thePlanningpanel in Figure[1](https://arxiv.org/html/2605.24266#S1.F1)\)\.
2. 2\.Pause decision and expansion:Compute branch utilities, execution costs, and the expected gain of asking, and compare this to the pause cost\. If a pause occurs, present the diversified subset to the user and then expand the user\-selected items together with any newly suggested directions\. Otherwise, expand the system\-proposed diversified subset directly\. Sub\-agents then perform retrieval and reasoning at each expanded child to produce node\-level reports \(see thePause Decisionpanel in Figure[1](https://arxiv.org/html/2605.24266#S1.F1)\)\.
3. 3\.Persona modeling:Update the inferred personaP^\\hat\{P\}with signals from the query, initial profile, and any user feedback gathered during pauses\. The updated persona conditions planning, utility scoring, and synthesis in subsequent steps \(see thePersona Modelingpanel in Figure[1](https://arxiv.org/html/2605.24266#S1.F1)\)\.

The process terminates once all nodes at depthDDhave been expanded, at which point the accumulated node reports are aggregated into the final reportRR\. This interactive loop enables reports that are better aligned with user goals while minimizing redundant or off\-topic exploration\.

### 2\.3Diversity\-Aware Exploration

As described in Section[2\.2](https://arxiv.org/html/2605.24266#S2.SS2), at each frontier noden⋆n^\{\\star\}, the system generates a set of follow\-up questions as potential next steps\. To promote exploration and reduce redundancy, we explicitly prompt fordistinct facetsand include onewild\-carddirection \(see Appendix[K](https://arxiv.org/html/2605.24266#A11)for prompt details\)\. From this candidate set, we select a diversified subset of sizeKKto either present to the user \(if a pause is triggered\) or expand automatically\.

To select this subset, we apply a greedy Maximal Marginal Relevance \(MMR\) strategy\(carbonell1998use;wang2025diversity\), which balances confidence scores with dissimilarity to previously chosen directions\. MMR is particularly well\-suited to our setting: it is simple, efficient, and interpretable, while effectively encouraging topical coverage across different aspects\. In contrast, alternative diversity methods \(e\.g\., clustering or determinantal point processes\) introduce additional complexity and hyper\-parameters without clear gains in this context\. Appendix[C](https://arxiv.org/html/2605.24266#A3)provides the full algorithmic details\.

### 2\.4Pause Decision and Expansion

After the proposal stage has produced a diversified set of candidates, the system must decide whether to involve the user or continue autonomously\. Asking everywhere is undesirable: user tolerance for interruptions is limited and varies widely\. Some users prefer high\-level guidance while trusting the system to handle details; others are more detail\-oriented but want control only in specific themes\. Preferences also shift across the depth of the research tree and over time\. A well\-calibrated system must respect these preferences while steering the exploration toward the user’s goals\. Below, we present the pause decision mechanism in a top\-down structure: we begin with the overall decision rule and then unpack its components, including pause cost, expected gain, and branch utility\.

#### Decision rule

At each frontier noden⋆n^\{\\star\}, the system evaluates whether pausing to ask the user is beneficial\. This decision is framed as a cost–benefit comparison:

a​\(n⋆\)=\{PauseAsk,Δ​E​V​\(n⋆\)\>C​\(n⋆\),Proceed,otherwise\.a\(n^\{\\star\}\)=\\begin\{cases\}\\textsc\{PauseAsk\},&\\Delta EV\(n^\{\\star\}\)\>C\(n^\{\\star\}\),\\\\\[2\.0pt\] \\textsc\{Proceed\},&\\text\{otherwise\.\}\\end\{cases\}Here,Δ​E​V​\(n⋆\)\\Delta EV\(n^\{\\star\}\)denotes the expected utility gain from pausing — by allowing the user to refine or redirect the next steps — whileC​\(n⋆\)C\(n^\{\\star\}\)denotes the cost of interruption, scaled by user\-specific tolerance\.

#### Pause cost

Not all users interact in the same way\. To model this, we assume two things:\(1\)a user’s tolerance for interruptions decreases over time, and\(2\)users differ in how much interruption they are willing to tolerate in total, and how fast that tolerance depletes\.

To capture this, we introduce two hyper\-parameters:

- •C0∈\[0,1\]C\_\{0\}\\in\[0,1\]: the*base pause cost*\. This reflects a user’s general sensitivity to interruptions\. A lowerC0C\_\{0\}implies the user is open to frequent interaction; a higherC0C\_\{0\}indicates a preference for minimal disruption\.
- •Tol∈ℕ\\mathrm\{Tol\}\\in\\mathbb\{N\}: the*tolerance budget*\. This governs how quickly the pause cost increases with the number of questions asked\. Intuitively,Tol\\mathrm\{Tol\}represents the approximate number of clarification questions the user is comfortable answering across the entire session\.

A user may tolerate multiple clarifications within a single topic but become frustrated by interruptions scattered across too many unrelated ones\. To reflect this, we distribute the global tolerance budgetTol\\mathrm\{Tol\}across all activetop\-level directions, defined as the root’s immediate children\. While users may have different preferences across themes, we simplify by evenly dividing the tolerance budget across top\-level directions\. Each nodennbelongs to a top\-level directionj∈K′j\\in K^\{\\prime\}, whereK′K^\{\\prime\}denotes the number of currently active directions\. If the system proceeds automatically,K′=KK^\{\\prime\}=K\(the full diversified set\)\. If a pause occurs,K′K^\{\\prime\}equals the number of user\-selected plus user\-added directions\. The pause cost at a frontier noden⋆n^\{\\star\}is then computed as:

C​\(n⋆\)=C0⋅\(1\+pausesjTolj\),C\(n^\{\\star\}\)=C\_\{0\}\\cdot\\left\(1\+\\frac\{\\mathrm\{pauses\}\_\{j\}\}\{\\mathrm\{Tol\}\_\{j\}\}\\right\),wherepausesj\\mathrm\{pauses\}\_\{j\}is the number of times the system has previously paused in directionjj\. As the number of pauses grows within a direction, the cost of pausing again increases proportionally\.

#### Pause gain

The*gain of pausing*should reflect two factors: the utility we forgo by pruning branches and the execution cost we save by not pursuing them\.

Let\{nk⋆\}k=1K\\\{n^\{\\star\}\_\{k\}\\\}\_\{k=1\}^\{K\}be the candidate children at the frontier node, with branch utilitiesU​\(nk⋆\)U\(n^\{\\star\}\_\{k\}\)and normalized execution costsCexec​\(nk⋆\)C^\{\\mathrm\{exec\}\}\(n^\{\\star\}\_\{k\}\)\. If we proceed automatically, we pursue allKK, and the expected value of the frontier node without pausing isE​Vno​\-​ask​\(n⋆\)=∑k=1KU​\(nk⋆\)−∑k=1KCexec​\(nk⋆\)EV^\{\\mathrm\{no\\text\{\-\}ask\}\}\(n^\{\\star\}\)=\\sum\_\{k=1\}^\{K\}U\(n^\{\\star\}\_\{k\}\)\-\\sum\_\{k=1\}^\{K\}C^\{\\mathrm\{exec\}\}\(n^\{\\star\}\_\{k\}\)\. If we pause, the user keeps a subsetS⊆\{1,…,K\}S\\subseteq\\\{1,\\dots,K\\\}, so the expected value of the frontier node with pause isE​Vask​\(n⋆\)=∑k∈SU​\(nk⋆\)−∑k∈SCexec​\(nk⋆\)EV^\{\\mathrm\{ask\}\}\(n^\{\\star\}\)=\\sum\_\{k\\in S\}U\(n^\{\\star\}\_\{k\}\)\-\\sum\_\{k\\in S\}C^\{\\mathrm\{exec\}\}\(n^\{\\star\}\_\{k\}\)\. To estimateSS, we retain candidates whose upper utility bound overlaps the leader’s lower bound, capturing all options that are plausibly optimal\. Equivalently, this decision rule prunes all branches whose best\-case utility still falls below the worst\-case value of the current leader\. See Appendix[D](https://arxiv.org/html/2605.24266#A4)for bound construction and filtering\.

A pause only changes which branches we do*not*execute\. The gain of pausing at the frontier node is the saved cost minus the lost utility of those pruned branches:

Δ​E​V​\(n⋆\)\\displaystyle\\Delta EV\(n^\{\\star\}\)=E​Vask​\(n⋆\)−E​Vno​\-​ask​\(n⋆\)\\displaystyle=EV^\{\\mathrm\{ask\}\}\(n^\{\\star\}\)\-EV^\{\\mathrm\{no\\text\{\-\}ask\}\}\(n^\{\\star\}\)=∑k∈Sc\(−U​\(nk⋆\)\+Cexec​\(nk⋆\)\)\.\\displaystyle=\\sum\_\{k\\in S^\{c\}\}\\bigl\(\-U\(n^\{\\star\}\_\{k\}\)\+C^\{\\mathrm\{exec\}\}\(n^\{\\star\}\_\{k\}\)\\bigr\)\.
Branch utility\.We score each candidate childnk⋆n^\{\\star\}\_\{k\}using a weighted combination of three factors:

U​\(nk⋆\)=Δ​Align​\(nk⋆\)\+λexp​Explore​\(nk⋆\)\+λinfo​InfoGain​\(nk⋆\),U\(n^\{\\star\}\_\{k\}\)=\\Delta\\mathrm\{Align\}\(n^\{\\star\}\_\{k\}\)\+\\lambda\_\{\\mathrm\{exp\}\}\\,\\mathrm\{Explore\}\(n^\{\\star\}\_\{k\}\)\\\\ \+\\lambda\_\{\\mathrm\{info\}\}\\,\\mathrm\{InfoGain\}\(n^\{\\star\}\_\{k\}\),where each component is scaled to\[0,1\]\[0,1\]for direct comparability with the pause cost\. \(See Appendix[D](https://arxiv.org/html/2605.24266#A4)for exact computations and normalization\.\)

- •Alignment gain\(Δ​Align\\Delta\\mathrm\{Align\}\) computes predicted increase in persona alignment relative to the parent under the current inferred aspectsA^s\\hat\{A\}\_\{s\}\. It rewards branches that cover more of what the user actually cares about\.
- •Exploration bonus\(Explore\\mathrm\{Explore\}\) adds a small reward for under\-explored facets to discourage repeatedly selecting the same angle\. We capture this “reward under\-explored, penalize over\-explored” behavior with a lightweight*count\-based bonus over facet tags*, inspired by the Upper Confidence Bound \(UCB\) family\(auer2002using;auer2002finite;li2010contextual\): rarely used facets receive larger bonuses, and the bonus decays naturally as they are chosen more frequently\. We use UCB here as an*optimism\-based exploration prior*rather than a classical stationary\-bandit algorithm\. The arms \(facets\) are non\-stationary and conditioned on an evolving persona, so we do not claim standard UCB regret guarantees\. Instead, Explore provides a simple, interpretable inductive bias against repeated collapse onto the same facet, complementing MMR \(local diversity\) and InfoGain \(semantic novelty\) by promoting longer\-horizon facet coverage\.
- •Information gain\(InfoGain\\mathrm\{InfoGain\}\) measures the content\-level novelty of a candidate’s expected evidence relative to accumulated learnings\. WhileExplore\\mathrm\{Explore\}encourages facet\-level diversity,InfoGain\\mathrm\{InfoGain\}focuses on semantic\-level novelty, prioritizing branches that are more likely to yield genuinely new information from the web\.

λexp\\lambda\_\{\\mathrm\{exp\}\}andλinfo\\lambda\_\{\\mathrm\{info\}\}balance breadth and novelty against alignment\. BothExplore\\mathrm\{Explore\}andInfoGain\\mathrm\{InfoGain\}complement the diversify\-aware exploration described in Section[2\.3](https://arxiv.org/html/2605.24266#S2.SS3): while the latter ensures that the*initial question set*spans distinct facets, it does not guarantee that the resulting content will be diverse\.Explore\\mathrm\{Explore\}andInfoGain\\mathrm\{InfoGain\}help mitigate this by promoting long\-term diversity at the facet and content levels, respectively\. While our process uses a minimal three\-factor utility for clarity and stability, the framework is easily extensible — additional criteria \(e\.g\., risk, credibility\) can be incorporated as needed\.

Execution cost\.Cexec​\(nk⋆\)C^\{\\mathrm\{exec\}\}\(n^\{\\star\}\_\{k\}\)estimates remaining work if we expandnk⋆n^\{\\star\}\_\{k\}\. It is also normalized to\[0,1\]\[0,1\]so it is commensurate with utilities\. We approximate the cost by the tokens of a saturated subtree beneathnk⋆n^\{\\star\}\_\{k\}, as tokens provide a consistent, model\-agnostic proxy, and correlate with both latency and spend\. See Appendix[D](https://arxiv.org/html/2605.24266#A4)for computation details\.

### 2\.5Persona Modeling

Beyond deciding*when*to ask \(Section[2\.4](https://arxiv.org/html/2605.24266#S2.SS4)\), the system must also know*who*it is optimizing for\. In deep research, users often do not know exactly what they want at the start\. Their goals shift as they encounter new information, and partial results may reveal new priorities\. Fixing a full persona upfront risks overfitting to stale assumptions or flooding the system with irrelevant detail\. To address this, we maintain a*live*persona that evolves dynamically as the research progresses\.

At eachn⋆n^\{\\star\},SteERmaintains an updated persona estimateP^​\(n⋆\)=\(p^text​\(n⋆\),𝒜^​\(n⋆\)\)\\hat\{P\}\(n^\{\\star\}\)=\\big\(\\hat\{p\}\_\{\\text\{text\}\}\(n^\{\\star\}\),\\,\\hat\{\\mathcal\{A\}\}\(n^\{\\star\}\)\\big\), wherep^text​\(n⋆\)\\hat\{p\}\_\{\\text\{text\}\}\(n^\{\\star\}\)captures the user’s profile and𝒜^​\(n⋆\)\\hat\{\\mathcal\{A\}\}\(n^\{\\star\}\)represents the current inferred set of aspects the user cares about\. When a pause occurs, we updateP^​\(n⋆\)\\hat\{P\}\(n^\{\\star\}\)based on user\-selected directions and any new suggestions, and implicitly incorporate recent research findings\. This evolving persona conditions all downstream modules: it guides research and follow\-up question generation, shapes the branch utility score via alignment to𝒜^​\(n⋆\)\\hat\{\\mathcal\{A\}\}\(n^\{\\star\}\)\(decision\), and steers final report synthesis\. See Appendix[K](https://arxiv.org/html/2605.24266#A11)for full details on how the persona is inferred and updated using LLM prompts \(Persona Checklist InferenceandPersona Modelingprompts\), and how the evolvingP^​\(n⋆\)\\hat\{P\}\(n^\{\\star\}\)is used across the planning, research, and synthesis pipeline\.

A live persona keeps the interaction tightly aligned with the user’s current interests\. It prevents drift caused by outdated assumptions, reduces unnecessary questions by filtering irrelevant directions, and adapts to new priorities that emerge during exploration\.

## 3Experiments

### 3\.1Experimental Setup

#### Evaluation data

We synthesize query–persona pairs by adapting established datasets and methods, with light modifications to better suit our goals\. We begin with 1k queries from the*Researchy Questions*dataset\(rosset2024researchy\), as used in*DeepResearchGym*\(coelho2025deepresearchgym\)\. For each query, we generate a plausible user personaptextp\_\{\\text\{text\}\}by adapting the ALOE profile–personality paradigm\(Wuet al\.,[2025](https://arxiv.org/html/2605.24266#bib.bib6)\): we seed from ALOE profiles and prompt an LLM to propose new profiles that would reasonably ask the given query\. To ensure diversity, we apply SBERT\-based filtering\(Reimers and Gurevych,[2019](https://arxiv.org/html/2605.24266#bib.bib2)\)and keep only distinct, plausible personas, following prior work\(Wuet al\.,[2025](https://arxiv.org/html/2605.24266#bib.bib6); Wanget al\.,[2023](https://arxiv.org/html/2605.24266#bib.bib4)\)\.

Given eachptextp\_\{\\text\{text\}\}, we generate 5–8 evaluation aspects𝒜\\mathcal\{A\}using prompts inspired bysalemi2025lamp, following their checklist format to ensure that the aspects are actionable, measurable, and grounded in the persona\. This enables robust alignment and focus evaluation, avoiding the ambiguity of more generic rubrics\.

Compared toWuet al\.\([2025](https://arxiv.org/html/2605.24266#bib.bib6)\)andsalemi2025lamp, our adaptations are minimal but tailored to deep research:\(i\)persona generation is query\-conditioned to ensure relevance,\(ii\)diversity filtering is stricter to avoid near\-duplicates, and\(iii\)aspects are framed for long\-form, cited outputs\. We evaluate on a held\-out set of 200 queries\. Full details of data generation are in Appendix[E](https://arxiv.org/html/2605.24266#A5)\.

#### *User Agent*simulation

To enable scalable, repeatable evaluation, we simulate user interactions with a*User Agent*conditioned on the full personaP=\(ptext,𝒜\)P=\(p\_\{\\text\{text\}\},\\mathcal\{A\}\)\. The agent selects directions that best align with𝒜\\mathcal\{A\}and proposes a new follow\-up when uncovered aspects remain, yielding realistic steering signals without human\-in\-the\-loop variability\. \(See Appendix[N](https://arxiv.org/html/2605.24266#A14)for the full prompt\.\)

#### Metrics

We evaluate persona\-tailored quality using two proposed metrics:AlignmentandFocus, both judged by*gpt\-4\.1\-mini*following*DeepResearchGym*\. \(Prompts used to obtain these metrics are listed in Appendix[M](https://arxiv.org/html/2605.24266#A13)\.\) We present the meta\-evaluation results for the LLM judge in Appendix[I](https://arxiv.org/html/2605.24266#A9)\.

- •Alignment:Given aspect set𝒜\\mathcal\{A\}and reportRR, we compute:Align​\(R,𝒜\)=12​\|𝒜\|​∑a∈𝒜align​\(R,a\),align​\(R,a\)∈\{0,1,2\}\.\\mathrm\{Align\}\(R,\\mathcal\{A\}\)=\\tfrac\{1\}\{2\|\\mathcal\{A\}\|\}\\sum\_\{a\\in\\mathcal\{A\}\}\\mathrm\{align\}\(R,a\),\\quad\\mathrm\{align\}\(R,a\)\\in\\\{0,1,2\\\}\.Here, 0 means that the aspect is not addressed, 1 means that it is partially addressed \(e\.g\., mentioned or vaguely covered\), and 2 means that it is fully addressed with sufficient detail and evidence, all scored by the LLM\-judge\. This gives an interpretable, per\-aspect measure of user alignment\.
- •Focus:We extract a set of keypoints𝒦​𝒫\\mathcal\{KP\}— short, evidence\-bearing spans — fromRRusing an LLM, and then ask the judge whether each keypoint \(k∈𝒦​𝒫k\\in\\mathcal\{KP\}\) maps to at least one user aspect:Focuskp​\(R,𝒜\)=1\|𝒦​𝒫\|​∑k∈𝒦​𝒫𝐈​\[map​\(k\)≠∅\]\.\\mathrm\{Focus\}\_\{\\mathrm\{kp\}\}\(R,\\mathcal\{A\}\)=\\tfrac\{1\}\{\|\\mathcal\{KP\}\|\}\\sum\_\{k\\in\\mathcal\{KP\}\}\\mathbf\{I\}\[\\mathrm\{map\}\(k\)\\neq\\varnothing\]\.While alignment is akin to*recall*over aspects, focus acts as a form of*precision*, rewarding dense, on\-target content\.

In addition, we report*DeepResearchGym*’s quality metrics, including clarity, depth, breadth, and insight, to evaluate general writing quality beyond persona targeting\.

#### Baselines

We compareSteERto two strong open\-source frameworks:*GPT\-Researcher*\(gptresearcher\)and*Open Deep Research*\(opendeepresearch\), both evaluated as top\-performing frameworks\(coelho2025deepresearchgym\)\. On the proprietary model side, we benchmark against OpenAI’so4\-mini\-deep\-researchmodel\.

We compare systems under a controlled setting: forSteERand the open\-source frameworks, all agents use GPT\-4o, the research tree is fixed \(depth 3, breadth 3\), outputs share the same token cap, and the only variable is persona information\. For fairness, all baselines are run under three input settings: \(1\) query only, \(2\) query \+ initial persona \(first sentence ofptextp\_\{\\text\{text\}\}\), and \(3\) query \+ full persona\. This allows us to assess how well each baseline adapts to different levels of user information\. Note thatSteERalways operates with only the initial persona, and must infer preferences dynamically throughout the interaction\.

### 3\.2How Much DoesSteERImprove Persona\-Tailored Quality?

Metric→\\rightarrowPersona\-TailoredQualitySystem↓\\downarrowAlignFocuskp\{\}\_\{\\text\{kp\}\}ClarityDepthBreadthInsightBalanceGPT\-Researcher66\.6378\.42\\cellcolorgreen\!7081\.8086\.3088\.4076\.6081\.25GPT\-Researcherinitial\-persona\{\}\_\{\\text\{initial\-persona\}\}74\.5981\.6879\.0587\.3788\.7175\.5281\.71GPT\-Researcherfull\-persona\{\}\_\{\\text\{full\-persona\}\}\\cellcolorgreen\!2079\.4883\.8377\.9387\.09\\cellcolorgreen\!2090\.3179\.0582\.58OpenDeepResearch62\.7483\.7274\.9082\.4088\.8568\.3981\.25OpenDeepResearchinitial\-persona\{\}\_\{\\text\{initial\-persona\}\}69\.7985\.4572\.5181\.6484\.1268\.9877\.44OpenDeepResearchfull\-persona\{\}\_\{\\text\{full\-persona\}\}77\.20\\cellcolorgreen\!2086\.1074\.0283\.4287\.6273\.1879\.44o4\-mini\-deep\-researchinitial\-persona\{\}\_\{\\text\{initial\-persona\}\}72\.7386\.0975\.76\\cellcolorgreen\!7089\.1089\.51\\cellcolorgreen\!7086\.74\\cellcolorgreen\!2082\.76o4\-mini\-deep\-researchfull\-persona\{\}\_\{\\text\{full\-persona\}\}75\.7286\.0275\.5487\.1987\.36\\cellcolorgreen\!2085\.0182\.63SteER\\cellcolorgreen\!7085\.70\\cellcolorgreen\!7086\.45\\cellcolorgreen\!2079\.97\\cellcolorgreen\!2088\.67\\cellcolorgreen\!7091\.2983\.04\\cellcolorgreen\!7084\.19

Table 2:Performance comparison betweenSteERand baselines\. ForSteER, we report performance atC0=0\.7C\_\{0\}=0\.7\(see Section[3\.3](https://arxiv.org/html/2605.24266#S3.SS3)for selection rationale\)\.From Table[2](https://arxiv.org/html/2605.24266#S3.T2), we see thatSteERachieves the strongest persona\-tailored performance on both metrics across all systems \(e\.g\. 7\.83% higher alignment than the runner\-up GPT\-Researcherfull\-persona\{\}\_\{\\text\{full\-persona\}\}\), even though some of those baselines are given the full persona, whileSteERonly receives the first sentence\. This highlights the effectiveness ofSteER’s interactive pausing and live persona modeling, which enable accurate mid\-process adaptation without relying on full upfront persona input\. This has practical appeal: real\-world deployments often face privacy constraints, onboarding friction, or noisy user profiles\.SteER’s ability to achieve strong alignment under minimal initial input makes it more robust in such settings\.

SteERalso leads in breadth and balance, reflecting the role ofSteER’s diversity\-aware exploration and utility components,Explore\\mathrm\{Explore\}andInfoGain\\mathrm\{InfoGain\}, in promoting semantic novelty and facet diversity\.SteERalso significantly outperforms the open\-source baselines in depth and insight, though it falls slightly short of the proprietary OpenAI model on these metrics\.

### 3\.3How DoesSteERProvide Interpretable Controls for Optimal Pausing?

![Refer to caption](https://arxiv.org/html/2605.24266v1/figures/pauseask_distribution.png)Figure 2:Distribution of number of pauses per run across base pause cost values\.![Refer to caption](https://arxiv.org/html/2605.24266v1/figures/align_per_pause.png)Figure 3:Alignment per pause across base pause cost values\.![Refer to caption](https://arxiv.org/html/2605.24266v1/figures/bpc_ablation.png)Figure 4:Effect of base pause cost on alignment \(left\) and focus \(right\)\. Baseline scores are shown as horizontal reference lines for comparison\.As introduced in Section[2\.4](https://arxiv.org/html/2605.24266#S2.SS4),SteERoffers two interpretable knobs to control pausing behavior: the base pause costC0C\_\{0\}, which sets the system’s aversion to interruptions, and the tolerance budgetTol\\mathrm\{Tol\}, which controls how quickly pause cost grows within a top\-level direction\. In this study, we varyC0C\_\{0\}while keepingTol=3\\mathrm\{Tol\}=3fixed\. This is because, in shallow trees \(depth 3\), the effect ofTol\\mathrm\{Tol\}is limited\.Tol\\mathrm\{Tol\}is more impactful in long\-horizon tasks where user fatigue may accumulate across levels\. Conceptually,Tol\\mathrm\{Tol\}captures a user\-specific interaction limit\. We tuneC0C\_\{0\}to match the pause budget \(Tol\\mathrm\{Tol\}\) while maximizing gain per pause\.

To benchmark against intuitive alternatives, we introduce aPauseAgentbaseline that uses an LLM agent to predict pause vs\. proceed at each frontier node \(prompt in Appendix[N](https://arxiv.org/html/2605.24266#A14)\)\. As shown in Figure[2](https://arxiv.org/html/2605.24266#S3.F2)\(and Appendix[F\.2](https://arxiv.org/html/2605.24266#A6.SS2)\),PauseAgentpauses excessively, far exceeding theTol=3\\mathrm\{Tol\}=3budget\. In contrast,SteERwithC0≥0\.4C\_\{0\}\\geq 0\.4remains within budget, averaging fewer than 3 pauses\.

Frequent pausing also hurts efficiency\. Figure[3](https://arxiv.org/html/2605.24266#S3.F3)shows that alignment per pause drops sharply at lowC0C\_\{0\}, while higherC0C\_\{0\}yields fewer but more impactful interventions\. This trade\-off is evident in Figure[4](https://arxiv.org/html/2605.24266#S3.F4): while absolute alignment declines asC0C\_\{0\}increases, alignment and focus reach local maxima aroundC0=0\.7C\_\{0\}=0\.7, indicating a practical sweet spot\.

In summary,SteERsupports calibrated control of interaction\.C0C\_\{0\}adjusts interruption cost directly, andTol\\mathrm\{Tol\}governs how that cost compounds over time\. This formulation provides both interpretability and personalization, outperforming thePauseAgentbaseline in effectiveness and flexibility\.

### 3\.4How DoesSteERAvoid Under\-Exploration Driven by Personalization?

Method↓\\downarrowAlignmentFocuskpDepthBreadthSteER85\.8287\.7990\.2793\.15\(w/o\)Explore\\mathrm\{Explore\}\\cellcolorred\!584\.98↓0\.98%\\cellcolorred\!1585\.17↓2\.98%\\cellcolorred\!1589\.86↓0\.45%\\cellcolorred\!1592\.60↓0\.59%\(w/o\)InfoGain\\mathrm\{InfoGain\}\\cellcolorred\!2582\.81↓3\.51%\\cellcolorred\!586\.40↓1\.58%\\cellcolorgreen\!1090\.41↑0\.15%\\cellcolorred\!592\.73↓0\.45%\(w/o\) Div Explore\\cellcolorred\!1584\.57↓1\.46%\\cellcolorred\!2584\.29↓3\.99%\\cellcolorred\!2588\.63↓1\.82%\\cellcolorred\!2591\.09↓2\.21%

Table 3:Ablation study on novelty and exploration components\. Darker red indicates a larger performance drop relative toSteER\.A potential failure mode is overfitting to personalization: when optimization focuses solely on aspect alignment, the system quickly collapses to a narrow trajectory, branch utilities flatten asΔ​Align\\Delta\\mathrm\{Align\}approaches zero, and exploration stalls\. To prevent this,SteERintegrates three complementary signals at different axes\.

First, diversity\-aware exploration ensures that research directions span distinct facets at each step, avoiding early myopia\. As shown in Table[3](https://arxiv.org/html/2605.24266#S3.T3), removing it causes the largest drops in depth, breadth, and focus, along with a significant decline in alignment, underscoring its role in maintaining structural and semantic diversity throughout the session\.

In addition, two utility terms guide exploration:Explore\\mathrm\{Explore\}encourages rotation across underrepresented facets, whileInfoGain\\mathrm\{InfoGain\}prioritizes semantic novelty\. AblatingExplore\\mathrm\{Explore\}leads to a large focus drop and notable declination in depth and breadth, with only a small impact on alignment, showing its importance in sustaining report\-wide diversity\. In contrast, removingInfoGain\\mathrm\{InfoGain\}yields the largest alignment drop but only relatively modest effects on other metrics\. This suggests that without semantic novelty, the system tends to dig deeper into already\-favored lines, satisfying more user aspects while producing redundant evidence\. These complementary behaviors introduce an interpretable trade\-off:λinfo\\lambda\_\{\\mathrm\{info\}\}prioritizes aspect satisfaction, whileλexp\\lambda\_\{\\mathrm\{exp\}\}favors breadth; we set both to 0\.5 for balance\.

While our experiments focus on novelty and exploration, the utility function is extensible\. Additional signals, such as factuality or plausibility, can be integrated into the same calibrated framework\. Our contribution lies not in these specific factors, but in the interaction paradigm that supports modular, interpretable control over research behavior\.

### 3\.5User Study Evaluation

MetricRaw Agreement \(%\)Gwet’s AC1Alignment82\.00\.639Focus73\.80\.475CoverageSteER65\.90\.318Baselines65\.20\.303FindabilitySteER75\.40\.508Baselines65\.60\.311

Table 4:Inter\-annotator agreement for user study annotations\.To complement the automated LLM\-judged metrics, we conducted a user study to evaluate whetherSteERis preferred by human users\. We comparedSteERwith GPT\-Researcher ando4\-mini\-deep\-researchon 20 query–persona pairs\. 12 annotators \(all NLP/CS graduate students\) viewed two reports for the same pair \(one fromSteER, one from a baseline\) in randomized order on our custom annotation platform\. Annotators saw only the two final reports, side by side in randomized order — no clarification questions, interaction traces, or intermediate system behaviors were exposed — and were unfamiliar with our system or the baselines’ output formats\. We did not observe consistent structural cues that would make a system identifiable \(e\.g\., tables of contents appeared across systems\)\. We nonetheless acknowledge that perfect blinding is difficult in any open\-ended generation setting; see Appendix[G](https://arxiv.org/html/2605.24266#A7)and Limitations \(Appendix[J](https://arxiv.org/html/2605.24266#A10)\) for further discussion\. For each comparison, annotators judgedAlignment\(better coverage of persona aspects\),Focus\(more on\-topic with less redundancy\),Coverage\(aspect\-level0–22, averaged\), andFindability\(report\-level0–22for ease of locating relevant information\)\.

This design captures both quality and usability:AlignmentandFocusreflect perceived persona\-fit111Note that theAlignmentandFocusmetrics used in the user study are based on pairwise human preferences and are not directly comparable to the automatic metrics defined in Section[3\.1](https://arxiv.org/html/2605.24266#S3.SS1)\.;Coveragemeasures how thoroughly user interests are addressed, andFindabilityassesses how easily users can locate what matters\. Full platform design and annotator instructions are in Appendix[G](https://arxiv.org/html/2605.24266#A7)\.

We collected 58 valid pairwise annotations\. To validate the quality of the annotations, we computed agreement using pairwise metrics across all annotator pairs for each evaluation dimension\. Specifically, we report raw pairwise agreement and Gwet’s AC1\(gwet2008computing\), a prevalence\-resistant chance\-corrected agreement coefficient that avoids the artificial deflation often observed with Fleiss’κ\\kappa\(fleiss1971measuring\)when the label distribution is skewed\. As summarized in Table[4](https://arxiv.org/html/2605.24266#S3.T4), forAlignmentandFocuspairwise preferences, annotators achieve raw pairwise agreement of82\.0%82\.0\\%and73\.8%73\.8\\%, with AC1 values of0\.6390\.639and0\.4750\.475, respectively, indicating substantial and moderate agreement\. ForCoverageandFindability, we observe similar patterns of fair to moderate agreement\. ForCoverage, raw agreement is65\.9%65\.9\\%forSteERand65\.2%65\.2\\%when aggregating baselines \(GPT\-Researcher ando4\-mini\-deep\-research\)\. ForFindability, raw agreement is75\.4%75\.4\\%forSteERand65\.6%65\.6\\%for the aggregated baselines\. These values reflect consistent, non\-trivial consensus across annotators on all dimensions, especially given the inherent subjectivity of report quality\.

![Refer to caption](https://arxiv.org/html/2605.24266v1/figures/win_loss_analysis.png)Figure 5:Pairwise human preference win rates onAlignmentandFocus\.As shown in Figure[5](https://arxiv.org/html/2605.24266#S3.F5),SteERis preferred in about8686–90%90\\%of cases for Alignment and about83%83\\%for Focus across GPT\-Researcher ando4\-mini\-deep\-research\.

![Refer to caption](https://arxiv.org/html/2605.24266v1/figures/coverage_findability_analysis.png)Figure 6:Human ratings onCoverageandFindability\.Left:Average aspect\-levelCoveragescores ofSteERand baselines\.Right:AverageFindabilityscores ofSteERand baselines\.Figure[6](https://arxiv.org/html/2605.24266#S3.F6)shows significant gains in*Coverage*and*Findability*forSteER\. On a0–22aspect\-coverage scale,SteERimproves the average by\+0\.623\+0\.623\(from0\.8280\.828to1\.4511\.451,p=3\.05​e−12p=3\.05e\-12\), a relative improvement of about75%75\\%which indicates a shift from below “somewhat covered” toward between “somewhat” and “fully” covered\. On the0–22Findability scale,SteERimproves by\+0\.690\+0\.690\(from0\.8450\.845to1\.5341\.534,p=1\.64​e−11p=1\.64e\-11\), moving readers from mostly difficult\-to\-medium retrieval to comfortably above medium and closer to “easy to find\.” Together, these results indicate thatSteERproduces reports that are both better aligned with persona needs and easier to navigate\.

## 4Discussion

#### *User Agent*simulation

To understand how our pause policy translates into user\-facing behavior, we analyzed the*User Agent*used in offline evaluation \(Appendix[F\.4](https://arxiv.org/html/2605.24266#A6.SS4)\)\. The*User Agent*maintains very high precision across base\-pause costs \(\>0\.97\>0\.97\), while recall declines asC0C\_\{0\}increases, and report alignment closely tracks*User Agent*recall \(r≈0\.81r\\approx 0\.81\)\. This indicates that pausing affects outcomes primarily by changing how many promising directions are retained and developed, offering a controllable alignment–effort frontier viaC0C\_\{0\}\. We view the*User Agent*as a diagnostic tool for sweeping policies and stress\-testing settings, but acknowledge that real users may be noisier and value exploration differently; future work will calibrate the*User Agent*with human logs and run counterfactual replays to quantify gaps between simulated and actual behavior\.

#### Persona modeling

We also examine how wellSteER’s live persona tracks report quality\. A useful takeaway from Appendix[F\.3](https://arxiv.org/html/2605.24266#A6.SS3)is thatSteERnot only pauses effectively but also recovers and maintains an accurate persona during a run\. Even with only the first persona sentence as input, the inferred persona’s alignment with the ground\-truth aspect set strongly tracks final report alignment \(r≈0\.85r\\approx 0\.85,p<10−3p<10^\{\-3\}\), indicating that the learned persona is informative rather than decorative\. AsC0C\_\{0\}increases, pauses become fewer, the inferred persona is less specified, and downstream alignment declines\. In practice, persona–report agreement is a useful diagnostic for selectingC0C\_\{0\}: choosing the smallestC0C\_\{0\}that achieves a target agreement while balancing the alignment–effort trade\-off\.

#### Broader application

Beyond our experiments,SteERsuggests a general pattern for long\-horizon, high\-stakes tasks that must balance personalization with exploration under interpretable control\. For instance, scientific\-discovery agents and research stacks could benefit from pausing and live\-persona steering to curb drift while preserving exploration\(team2025novelseek;schmidgall2025agentrxiv;zheng2025automation\)\. Likewise, high\-stakes domains such as financial advising and trading\(zhang2024multimodal;yu2024fincon\)and law and policy research\(li2024legalagentbench;pipitone2024legalbench\)are natural application areas forSteER’s interpretable, user\-steerable control\. Because ofSteER’s modularity, domains can add factors such as factuality, citation quality, or safety alongside novelty and exploration\. We view validating these extensions as promising future work\.

## 5Conclusion and Future Work

We have presentedSteER, proposing a new*interactive paradigm*for deep research\.SteERcouples a cost–benefit pause policy with interpretable controls, a live persona that adapts mid\-process, and diversity–novelty utility signals that keep exploration purposeful\. Our experiments show thatSteERimproves persona\-tailored quality by7\.83%​–​22\.80%7\.83\\%\\text\{–\}22\.80\\%over strong open\-source and proprietary systems, leads on generic quality metrics, and is preferred by human readers in over85%85\\%of alignment and83%83\\%of focus pairwise judgments\. We also release a persona–query evaluation suite and data pipeline to support reproducible testing and future model development\.

Looking ahead, several directions appear especially promising\. On the system side, exploring speculative pre\-execution to reduce latency, a dynamic breadth–depth planner, and policy learning for pause and branch selection could further strengthen real\-time usability\. On the evaluation side, end\-to\-end user studies evaluating full interactions, including task success, time to insight, perceived control and trust, and cognitive load, would better capture real\-world value\. We also see studyingSteER’s robustness to noisier real\-user behavior — contradictory feedback, partial or incomplete responses, and outright abandonment of the clarification process — as an important next step, naturally tied to preference\-conflict handling in the live persona model \(see Appendix[J](https://arxiv.org/html/2605.24266#A10)\)\.

## Impact Statement

This work presentsSteER, a framework for interactive, human\-guided deep research agents\. Our primary contribution is moving away from monolithic autonomous workflows toward a paradigm where model reasoning is transparent and steerable by the user\.

This research has the potential to significantly enhance productivity in knowledge\-intensive sectors by reducing the cognitive load of information synthesis\. By enabling agents to adapt to user intent in real\-time, we democratize access to complex research capabilities, allowing non\-experts to navigate dense technical or domain\-specific literature effectively\. Furthermore, our proposed adaptive pause mechanism improves computational efficiency by preventing agents from wasting resources on misaligned or irrelevant search trajectories\.

We acknowledge two primary risks associated with highly personalized research agents:

- •Automation Bias and Over\-Reliance:As agents become more capable of long\-horizon reasoning, users may uncritically accept generated summaries\.SteERmitigates this by design; the interactive nature of the framework forces periodic user engagement, keeping the human in the loop and encouraging critical evaluation of the agent’s intermediate findings rather than passive consumption of a final output\.
- •Echo Chambers and Confirmation Bias:Our system’s live persona modeling optimizes for user alignment, which carries the risk of reinforcing existing biases by prioritizing information that aligns with the user’s preconceptions\. To address this, future deployment of such systems should include diversity\-aware exploration in the frontier selection algorithm \(as described in Section[2\.3](https://arxiv.org/html/2605.24266#S2.SS3)to ensure opposing viewpoints are surfaced, even when they conflict with the modeled persona\.

Ultimately, this work aims to align powerful autonomous agents with human values\. By prioritizing steerability over pure autonomy, we provide a blueprint for safer AI systems that remain responsive to human oversight during complex tasks\.

## References

- I\. Frisch and M\. Giulianelli \(2024\)LLM agents in interaction: measuring personality consistency and linguistic alignment in interacting populations of large language models\.InProceedings of the 1st Workshop on Personalization of Generative AI Systems \(PERSONALIZE 2024\),A\. Deshpande, E\. Hwang, V\. Murahari, J\. S\. Park, D\. Yang, A\. Sabharwal, K\. Narasimhan, and A\. Kalyan \(Eds\.\),St\. Julians, Malta,pp\. 102–111\.External Links:[Link](https://aclanthology.org/2024.personalize-1.9/),[Document](https://dx.doi.org/10.18653/v1/2024.personalize-1.9)Cited by:[Appendix A](https://arxiv.org/html/2605.24266#A1.SS0.SSS0.Px2.p1.1)\.
- N\. Reimers and I\. Gurevych \(2019\)Sentence\-BERT: sentence embeddings using Siamese BERT\-networks\.InProceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing \(EMNLP\-IJCNLP\),K\. Inui, J\. Jiang, V\. Ng, and X\. Wan \(Eds\.\),Hong Kong, China,pp\. 3982–3992\.External Links:[Link](https://aclanthology.org/D19-1410/),[Document](https://dx.doi.org/10.18653/v1/D19-1410)Cited by:[Appendix E](https://arxiv.org/html/2605.24266#A5.p2.1),[§3\.1](https://arxiv.org/html/2605.24266#S3.SS1.SSS0.Px1.p1.1)\.
- A\. Salemi, S\. Mysore, M\. Bendersky, and H\. Zamani \(2024\)LaMP: when large language models meet personalization\.InProceedings of the 62nd Annual Meeting of the Association for Computational Linguistics \(Volume 1: Long Papers\),L\. Ku, A\. Martins, and V\. Srikumar \(Eds\.\),Bangkok, Thailand,pp\. 7370–7392\.External Links:[Link](https://aclanthology.org/2024.acl-long.399/),[Document](https://dx.doi.org/10.18653/v1/2024.acl-long.399)Cited by:[Appendix A](https://arxiv.org/html/2605.24266#A1.SS0.SSS0.Px2.p1.1),[§1](https://arxiv.org/html/2605.24266#S1.p2.1)\.
- Y\. Wang, Y\. Kordi, S\. Mishra, A\. Liu, N\. A\. Smith, D\. Khashabi, and H\. Hajishirzi \(2023\)Self\-instruct: aligning language models with self\-generated instructions\.InProceedings of the 61st Annual Meeting of the Association for Computational Linguistics \(Volume 1: Long Papers\),A\. Rogers, J\. Boyd\-Graber, and N\. Okazaki \(Eds\.\),Toronto, Canada,pp\. 13484–13508\.External Links:[Link](https://aclanthology.org/2023.acl-long.754/),[Document](https://dx.doi.org/10.18653/v1/2023.acl-long.754)Cited by:[Appendix E](https://arxiv.org/html/2605.24266#A5.p2.1),[§3\.1](https://arxiv.org/html/2605.24266#S3.SS1.SSS0.Px1.p1.1)\.
- C\. Wu, Z\. R\. Tam, C\. Wu, C\. Lin, H\. Lee, and Y\. Chen \(2024\)I need help\! evaluating LLM’s ability to ask for users’ support: a case study on text\-to\-SQL generation\.InProceedings of the 2024 Conference on Empirical Methods in Natural Language Processing,Y\. Al\-Onaizan, M\. Bansal, and Y\. Chen \(Eds\.\),Miami, Florida, USA,pp\. 2191–2199\.External Links:[Link](https://aclanthology.org/2024.emnlp-main.131/),[Document](https://dx.doi.org/10.18653/v1/2024.emnlp-main.131)Cited by:[Appendix A](https://arxiv.org/html/2605.24266#A1.SS0.SSS0.Px3.p1.1),[§1](https://arxiv.org/html/2605.24266#S1.p2.1)\.
- S\. Wu, Y\. R\. Fung, C\. Qian, J\. Kim, D\. Hakkani\-Tur, and H\. Ji \(2025\)Aligning LLMs with individual preferences via interaction\.InProceedings of the 31st International Conference on Computational Linguistics,O\. Rambow, L\. Wanner, M\. Apidianaki, H\. Al\-Khalifa, B\. D\. Eugenio, and S\. Schockaert \(Eds\.\),Abu Dhabi, UAE,pp\. 7648–7662\.External Links:[Link](https://aclanthology.org/2025.coling-main.511/)Cited by:[Appendix A](https://arxiv.org/html/2605.24266#A1.SS0.SSS0.Px2.p1.1),[Appendix E](https://arxiv.org/html/2605.24266#A5.p2.1),[§1](https://arxiv.org/html/2605.24266#S1.p2.1),[§2\.1](https://arxiv.org/html/2605.24266#S2.SS1.p2.5),[§3\.1](https://arxiv.org/html/2605.24266#S3.SS1.SSS0.Px1.p1.1),[§3\.1](https://arxiv.org/html/2605.24266#S3.SS1.SSS0.Px1.p3.1)\.

## Appendix ARelated Work

#### Deep research

LLM\-based research agents combine retrieval and multi\-step reasoning to produce long\-form answers\(coelho2025deepresearchgym\)\. Among the open\-source frameworks, two dominant paradigms are multi\-agent pipelines that split planning, browsing, and reading across roles\(huang2025manusearch;alzubi2025open;li2025search;zhang2025evolvesearch\)and RL\-trained agents that learn to search and reason\(zheng2025deepresearcher;jin2025search;song2025r1\)\. On the evaluation side, benchmarks such as DeepResearchGym\(coelho2025deepresearchgym\)and DeepResearch Bench\(du2025deepresearch\)have begun to standardize this setting, providing realistic research\-style questions and automated evaluation protocols for long\-form, citation\-heavy reports\. Despite progress, most systems still follow one\-shot scoping with at most a single clarification, then a long autonomous run and a monolithic report, offering little mid\-process control when user needs evolve\. Concurrent work, EDR\(prabhakar2025enterprisedeepresearchsteerable\), adds user\-initiated mid\-process steering via an exposed task plan, but lacks an adaptive pause policy and live persona modeling\.

#### Personalization and alignment

A growing line of work pursues personalization for LLM agents, moving from static profile–personality conditioning and long\-form checklists\(Wuet al\.,[2025](https://arxiv.org/html/2605.24266#bib.bib6); Salemiet al\.,[2024](https://arxiv.org/html/2605.24266#bib.bib3);salemi2025lamp\)toward interactive, test\-time adaptation and multi\-stakeholder alignment\(xie2025survey\)\. Recent trends probe persona behavior in interaction \(e\.g\., consistency and drift under dialogue\)\(Frisch and Giulianelli,[2024](https://arxiv.org/html/2605.24266#bib.bib1)\)and build agent mechanisms that adapt actions to user preferences at inference time\(zhang2025personaagent\)\. While this work establishes that preferences should be updated during use, most approaches still lack*interpretable, end\-to\-end controls*for deciding*when*to seek input and*how*to steer long\-horizon generation as goals evolve\.

#### Interactive reasoning and control

Another closely related line of work equips LLMs with interactive reasoning via clarification\. Prior studies train models to ask when information is missing\(andukuristar;ren2023robots; Wuet al\.,[2024](https://arxiv.org/html/2605.24266#bib.bib5)\), model future turns to decide ask vs\. answer\(ICLR2025\_97e2df4b\), and learn clarification policies with contrastive objectives\(DBLP:conf/iclr/Chen0PA25\)\. Visualization tools improve transparency and user steering over chains of thought\(pang2025interactive;li2025reasongraph\)\. However, these efforts mostly address local interactions or static control, rather than providing interpretable, end\-to\-end controls for*when*to pause,*what*to explore, and*how*to adapt personalization mid\-process in long\-horizon research\.

Table[1](https://arxiv.org/html/2605.24266#S1.T1)positionsSteERagainst both deep research frameworks and interactive/persona\-aware reasoning systems, clarifying which capabilities each class actually offers\. Open\-source deep research frameworks and proprietary web\-based services are all built around long autonomous run paradigm: once the user issues a query, a fixed pipeline executes to completion with no exposed, interpretable control over where in the research tree to intervene or when to ask for guidance\. The◆marks in the mid\-process steering column indicate the limited behavior: these systems sometimes allow a single upfront scoping or plan\-confirmation step before the full autonomous run, but provide no further steering within the trajectory; beyond that point they are fully autonomous\. In addition, for proprietary systems, public documentation suggests that fixed personas or user memory may be used, but there is no evidence of live persona modeling, and their control policies and internal mechanisms are not accessible or benchmarkable; we therefore restrict ourselves to conceptual comparison\.

Meanwhile, interactive/persona\-aware frameworks occupy the complementary side of the space: PersonaAgent\(zhang2025personaagent\)maintains a live user representation and adapts over time but has no steering capability, while ReasonGraph\(li\-etal\-2025\-reasongraph\)and HITL CoT MCS\(cai2023human\)expose mid\-process reasoning for visualization, inspection, or human correction but implement no adaptive pause mechanism and do not perform live persona modeling\. Taken together, the table shows that existing systems provide at most one of the three capabilities we target — mid\-process steering, adaptive pause decisions, or live persona modeling — and never all three in a benchmarkable deep research setting\.SteERis the only system that supports all three simultaneously, coupling a cost–benefit pause policy with a live persona that conditions planning, branch utility, and synthesis within a single research tree\.

## Appendix BUse of LLMs for Writing Assistance

We used ChatGPT\-4o*only*for language\-level editing\. Concretely:

- •Polishing prose, tightening sentences, fixing grammar and LaTeX wording, reordering or shortening paragraphs, and suggesting alternative titles or section headers\.
- •No ideas, methods, claims, proofs, experiments, numbers, figures, tables, code, prompts, or citations were produced by the model\. All technical content, analyses, and results were authored and verified by the authors\.
- •We supplied already written passages or outlines and requested editing \(for example, “polish wording, keep all technical details unchanged”\)\.
- •The model was not given proprietary data, code, or unpublished results beyond the text to be edited\. All outputs were reviewed by the authors for accuracy and tone\.

## Appendix CDiversified Subset Selection

For completeness, we include the pseudocode of the greedy MMR selection used in our framework\. Given a candidate set of follow\-up questions𝒞=\{q1,…,qM\}\\mathcal\{C\}=\\\{q\_\{1\},\\dots,q\_\{M\}\\\}with confidence scoresconf​\(qi\)\\mathrm\{conf\}\(q\_\{i\}\)and embeddings𝐞i\\mathbf\{e\}\_\{i\}, the algorithm selects a diversified subset𝒞′\\mathcal\{C\}^\{\\prime\}of sizeKK:

0:Candidate list

𝒞=\{q1,…,qM\}\\mathcal\{C\}=\\\{q\_\{1\},\\dots,q\_\{M\}\\\}with confidences

conf​\(qi\)\\mathrm\{conf\}\(q\_\{i\}\), embeddings

𝐞i\\mathbf\{e\}\_\{i\}; desired subset size

KK
0:Diversified subset

𝒞′\\mathcal\{C\}^\{\\prime\}
1:

𝒞←\\mathcal\{C\}\\leftarrowsort

𝒞\\mathcal\{C\}in non\-increasing order of

conf​\(qi\)\\mathrm\{conf\}\(q\_\{i\}\)
2:

𝒞′←∅,I𝒞′←∅\\mathcal\{C\}^\{\\prime\}\\leftarrow\\emptyset,\\;I\_\{\\mathcal\{C\}^\{\\prime\}\}\\leftarrow\\emptyset
3:while

\|𝒞′\|<K\|\\mathcal\{C\}^\{\\prime\}\|<Kdo

4:

C←\{i∣i∉I𝒞′\}C\\leftarrow\\\{\\,i\\mid i\\notin I\_\{\\mathcal\{C\}^\{\\prime\}\}\\\}
5:if

I𝒞′=∅I\_\{\\mathcal\{C\}^\{\\prime\}\}=\\emptysetthen

6:

i⋆←min⁡Ci^\{\\star\}\\leftarrow\\min C\{top\-confidence question\}

7:else

8:for

i∈Ci\\in Cdo

9:

di←maxj∈I𝒞′⁡sim⁡\(𝐞i,𝐞j\)\+εd\_\{i\}\\leftarrow\\displaystyle\\max\_\{j\\in I\_\{\\mathcal\{C\}^\{\\prime\}\}\}\\operatorname\{sim\}\(\\mathbf\{e\}\_\{i\},\\mathbf\{e\}\_\{j\}\)\\;\+\\;\\varepsilon
10:endfor

11:

i⋆←arg⁡mini∈C⁡dii^\{\\star\}\\leftarrow\\arg\\min\_\{i\\in C\}d\_\{i\}\{least similar to current set \(MMR criterion\)\}

12:endif

13:

𝒞′←𝒞′∪\{qi⋆\},I𝒞′←I𝒞′∪\{i⋆\}\\mathcal\{C\}^\{\\prime\}\\leftarrow\\mathcal\{C\}^\{\\prime\}\\cup\\\{\\,q\_\{i^\{\\star\}\}\\\},\\quad I\_\{\\mathcal\{C\}^\{\\prime\}\}\\leftarrow I\_\{\\mathcal\{C\}^\{\\prime\}\}\\cup\\\{i^\{\\star\}\\\}
14:endwhile

15:return

𝒞′\\mathcal\{C\}^\{\\prime\}

## Appendix DDetails for Gain of Pausing Implementation

#### Alignment gain

Letr​\(n\)r\(n\)denote the chunk report at nodenn, formed by concatenating the learnings\{ℓi\}i=1m\\\{\\ell\_\{i\}\\\}\_\{i=1\}^\{m\}\(if there aremmlearnings at the node\), and letA^n\\hat\{A\}\_\{n\}be the inferred aspect set at that node\. For thekk\-th child node of a frontier nodenk⋆n^\{\\star\}\_\{k\},

Δ​Align​\(nk⋆\)=Align​\(r​\(nk⋆\),A^n⋆\)−Align​\(r​\(n⋆\),A^n⋆\)\.\\Delta\\mathrm\{Align\}\(n^\{\\star\}\_\{k\}\)=\\mathrm\{Align\}\\bigl\(r\(n^\{\\star\}\_\{k\}\),\\hat\{A\}\_\{n^\{\\star\}\}\\bigr\)\\\\ \-\\;\\mathrm\{Align\}\\bigl\(r\(n^\{\\star\}\),\\hat\{A\}\_\{n^\{\\star\}\}\\bigr\)\.

#### Exploration bonus

For each chunk report, we prompt an LLM to assign short tags \(see Appendix[K](https://arxiv.org/html/2605.24266#A11),*Search Result Processing*prompt\)\. We maintain the global tag set𝒯\\mathcal\{T\}and a cumulative usage countcount​\(T\)\\mathrm\{count\}\(T\)for each tagTTup to the current step\. With a small constantϵ\>0\\epsilon\>0, the exploration bonus is

Explore​\(nk⋆\)=1\|𝒯\|​∑T∈𝒯ϵ1\+count​\(T\)\.\\mathrm\{Explore\}\(n^\{\\star\}\_\{k\}\)=\\frac\{1\}\{\|\\mathcal\{T\}\|\}\\sum\_\{T\\in\\mathcal\{T\}\}\\frac\{\\epsilon\}\{1\+\\sqrt\{\\mathrm\{count\}\(T\)\}\}\.This UCB\-style term grants larger bonus to under\-tried tags and decays as a tag is reused\.

#### Information gain

To reward novelty relative to what has already been learned, we compare a candidate’s node embedding to the centroid of accumulated learnings\. Let𝐞ℓ\\mathbf\{e\}\_\{\\ell\}be the embedding of a learningℓ\\ell\. For nodennwith number of learningsL​\(n\)=\{ℓi\}i=1m​\(n\)L\(n\)=\\\{\\ell\_\{i\}\\\}\_\{i=1\}^\{m\(n\)\}, define its embedding𝐞n=1m​\(n\)​∑i=1m​\(n\)𝐞ℓi\\mathbf\{e\}\_\{n\}=\\frac\{1\}\{m\(n\)\}\\sum\_\{i=1\}^\{m\(n\)\}\\mathbf\{e\}\_\{\\ell\_\{i\}\}\(whenm​\(n\)\>0m\(n\)\>0\)\. Letℒ\\mathcal\{L\}be the set of all learnings gathered so far,M=\|ℒ\|M=\|\\mathcal\{L\}\|, andμ=1M​∑ℓ∈ℒ𝐞ℓ\\mu=\\frac\{1\}\{M\}\\sum\_\{\\ell\\in\\mathcal\{L\}\}\\mathbf\{e\}\_\{\\ell\}\. Then

InfoGain​\(nk⋆\)=\{1−sim⁡\(𝐞nk⋆,μ\),m​\(nk⋆\)\>0​and​M\>0,0,m​\(nk⋆\)=0,1,otherwise\.\\mathrm\{InfoGain\}\(n^\{\\star\}\_\{k\}\)=\\begin\{cases\}1\-\\operatorname\{sim\}\\\!\\bigl\(\\mathbf\{e\}\_\{n^\{\\star\}\_\{k\}\},\\mu\\bigr\),&m\(n^\{\\star\}\_\{k\}\)\>0\\text\{ and \}M\>0,\\\\ 0,&m\(n^\{\\star\}\_\{k\}\)=0,\\\\ 1,&\\text\{otherwise\.\}\\end\{cases\}

#### Execution cost

LetDDbe the max depth,d​\(n\)d\(n\)the depth of nodenn, andKKthe branching factor\. For childnk⋆n^\{\\star\}\_\{k\}, the remaining depth isdrem=D−d​\(nk⋆\)d\_\{\\mathrm\{rem\}\}=D\-d\(n^\{\\star\}\_\{k\}\)\. The number of nodes in a saturatedKK\-ary subtree is

Nrem=\{Kdrem\+1−1K−1,K\>1,drem\+1,K=1\.N\_\{\\mathrm\{rem\}\}=\\begin\{cases\}\\dfrac\{K^\{\\,d\_\{\\mathrm\{rem\}\}\+1\}\-1\}\{K\-1\},&K\>1,\\\\\[4\.0pt\] d\_\{\\mathrm\{rem\}\}\+1,&K=1\.\\end\{cases\}With a running average token costTokavg\\mathrm\{Tok\}\_\{\\mathrm\{avg\}\}per node, the estimated tokens areTkest=Tokavg​NremT^\{\\mathrm\{est\}\}\_\{k\}=\\mathrm\{Tok\}\_\{\\mathrm\{avg\}\}\\,N\_\{\\mathrm\{rem\}\}, and the normalized execution cost is

Cexec​\(nk⋆\)=TkestTkest\+Tokavg=NremNrem\+1\.C^\{\\mathrm\{exec\}\}\(n^\{\\star\}\_\{k\}\)=\\frac\{T^\{\\mathrm\{est\}\}\_\{k\}\}\{T^\{\\mathrm\{est\}\}\_\{k\}\+\\mathrm\{Tok\}\_\{\\mathrm\{avg\}\}\}=\\frac\{N\_\{\\mathrm\{rem\}\}\}\{N\_\{\\mathrm\{rem\}\}\+1\}\.

#### Filtering candidates when pausing

LetUk=U​\(nk⋆\)U\_\{k\}=U\(n^\{\\star\}\_\{k\}\)andc​o​n​fk∈\[0,1\]conf\_\{k\}\\in\[0,1\]be a confidence score generated by the LLM \(see Appendix[K](https://arxiv.org/html/2605.24266#A11),*Search Result Processing*prompt\)\. Define the uncertainty radius

rk=\(1−c​o​n​fk\)​\(maxi∈K⁡Ui−mini∈K⁡Ui\),\\displaystyle r\_\{k\}=\(1\-conf\_\{k\}\)\\bigl\(\\max\_\{i\\in K\}U\_\{i\}\-\\min\_\{i\\in K\}U\_\{i\}\\bigr\),Ukupper=Uk\+rk,Uklower=Uk−rk\.\\displaystyle U^\{\\mathrm\{upper\}\}\_\{k\}=U\_\{k\}\+r\_\{k\},\\quad U^\{\\mathrm\{lower\}\}\_\{k\}=U\_\{k\}\-r\_\{k\}\.The*could\-be\-the\-best*set is

S=\{k\|Ukupper≥maxi∈K⁡Uilower\}\.S=\\bigl\\\{k\\,\\big\|\\,U^\{\\mathrm\{upper\}\}\_\{k\}\\geq\\max\_\{i\\in K\}U^\{\\mathrm\{lower\}\}\_\{i\}\\bigr\\\}\.This mirrors upper and lower confidence bounds for best\-arm filtering\(pmlr\-v35\-jamieson14\)\.

## Appendix EData Construction Details

Data Split→\\rightarrowAllEvalTotal Queries1000200Total Query\-Persona Pairs1381286Queries with 1 Persona646\(64\.6%\)125\(62\.5%\)Queries with 2 Personas327\(32\.7%\)64\(32\.0%\)Queries with 3 Personas27\(2\.7%\)11\(5\.5%\)

Table 5:Data StatisticsTo evaluate our method, we need a dataset with deep research worthy questions paired with realistic personas, where personas are, as defined in Section[2\.1](https://arxiv.org/html/2605.24266#S2.SS1),\(ptext,𝒜\)\(p\_\{\\text\{text\}\},\\mathcal\{A\}\), whereptextp\_\{\\text\{text\}\}is a string, combining the user’s profile and personality, and𝒜\\mathcal\{A\}is a set ofaspectsthat the user is interested to see in a high\-quality, well\-aligned final report\. We construct our dataset on top of the subset of 1,000 queries from Researchy Questions dataset\(rosset2024researchy\)used in DeepResearchGym\(coelho2025deepresearchgym\)\.

For each query, we first generate one or more\(ptext\(p\_\{\\text\{text\}\}that would be reasonable to ask the query\. For this, we adopt a two\-step approach\. In the first step, inspired byWuet al\.\([2025](https://arxiv.org/html/2605.24266#bib.bib6)\)and\(Wanget al\.,[2023](https://arxiv.org/html/2605.24266#bib.bib4)\), we use an iterative self\-generation and filtering pipeline\. In each round, 3 profiles are randomly selected from the profiles in the ALOE dataset\(Wuet al\.,[2025](https://arxiv.org/html/2605.24266#bib.bib6)\)and used as input to an off\-the\-shelf LLM \(GPT\-4o\) to generate 3 new profiles that would be reasonable to ask the query per iteration\. Then we introduce an automatic filtering process based on semantic similarity to ensure the distinctiveness and diversity of the generated profiles\. Same asWuet al\.\([2025](https://arxiv.org/html/2605.24266#bib.bib6)\), we use Sentence Transformers\(Reimers and Gurevych,[2019](https://arxiv.org/html/2605.24266#bib.bib2)\)to compute embedding of the generated profiles and measure the cosine similarity among the generated new profiles\. For each new profile, if the highest similarity score compared to the other profiles exceeds 0\.65, the profile is considered too similar to at least one of the other profiles and discarded\. Otherwise, it will be accepted as a successful new profile to pair with the query\. We repeat the process until 3 new accepted profiles are generated\. In step 2, for each accepted profile, we generate a reasonable personality with GPT\-4o to pair with it\. For this, we randomly sample personalities from personality pool of the ALOE dataset as sample personalities fed into the LLM for generation\.

Once we have generated one or moreptextp\_\{\\text\{text\}\}for each query, we then generate the set of aspects𝒜\\mathcal\{A\}for eachptextp\_\{\\text\{text\}\}\. We adopt the same approach as insalemi2025lampto generate 5\-8 specific aspects that a user \(described byptextp\_\{\\text\{text\}\}\) would expect to see in a comprehensive and helpful report to the query, along with an evidence and a reasoning for each aspect, attributed fromptextp\_\{\\text\{text\}\}\.

Table[5](https://arxiv.org/html/2605.24266#A5.T5)details the statistics of our generated dataset\. All prompts for persona/profile/aspect generation are provided in Appendix[L](https://arxiv.org/html/2605.24266#A12)\.

## Appendix FAdditional Experiment Details

SystemFocusst\{\}\_\{\\text\{st\}\}↑\\uparrowRelevancesup\{\}\_\{\\text\{sup\}\}↑\\uparrowRelevancecon\{\}\_\{\\text\{con\}\}↓\\downarrowGPT\-Researcher67\.0761\.391\.02GPT\-Researcherinitial\-persona\{\}\_\{\\text\{initial\-persona\}\}69\.1860\.821\.11GPT\-Researcherfull\-persona\{\}\_\{\\text\{full\-persona\}\}70\.7859\.941\.04OpenDeepResearch73\.1560\.360\.69OpenDeepResearchinitial\-persona\{\}\_\{\\text\{initial\-persona\}\}75\.6157\.070\.81OpenDeepResearchfull\-persona\{\}\_\{\\text\{full\-persona\}\}78\.8457\.060\.81o4\-mini\-deep\-researchinitial\-persona\{\}\_\{\\text\{initial\-persona\}\}78\.4167\.361\.74o4\-mini\-deep\-researchfull\-persona\{\}\_\{\\text\{full\-persona\}\}80\.6066\.451\.94SteER\[C0=0\.7\]\{\}\_\{\[C\_\{0\}=0\.7\]\}78\.5160\.471\.13SteER\[C0=0\.1\]\{\}\_\{\[C\_\{0\}=0\.1\]\}80\.6760\.191\.10

Table 6:Performance comparison betweenSteERand baseline frameworks on sentence\-level focus score and report relevance scores\.### F\.1Additional Metrics

Table[6](https://arxiv.org/html/2605.24266#A6.T6)reports sentence\-level focus and DeepResearchGym relevance \(support↑\\uparrowand contradiction↓\\downarrow\)\. We do not use sentence\-level focus as a primary metric because it is length sensitive: the scoreFocusst\\mathrm\{Focus\}\_\{\\mathrm\{st\}\}is the fraction of sentences mapped to any aspect, so longer reports with a few connective or background sentences are penalized, whereas terse styles can inflate the ratio\. Still,SteERachieves competitive values \(e\.g\.,80\.6780\.67atC0=0\.1C\_\{0\}\{=\}0\.1\), on par with the proprietary model and higher than the open\-source baselines, indicating that personalization does not come at the cost of sentence\-level topicality\.

DeepResearchGym relevance compares a report to a pre\-extracted, task\-generic keypoint list; becauseSteERsteers into personalized directions, it is expected to score lower on relevancesupthan a system optimized for the generic keypoints \(e\.g\.,o4\-mini\-deep\-research\), while maintaining moderate relevancecon\. In our results,SteER’s relevancesupis similar to GPT\-Researcher and OpenDeepResearch, with contradiction around1\.101\.10–1\.131\.13; the proprietary model attains higher support but also substantially higher contradiction, whereas OpenDeepResearch shows low contradiction but lower support\. Taken together, these metrics are complementary diagnostics: sentence\-level focus confirms topicality at the sentence granularity, and DeepResearchGym relevance reflects overlap with generic keypoints rather than user\-specific goals\.

### F\.2Base Pause Cost vs\. Pause Behavior

To better understand system behavior, Figure[2](https://arxiv.org/html/2605.24266#S3.F2)shows the distribution of number of pauses per run\. As expected, lowering base pause cost increases the number of pauses, with median pauses dropping from around 10 \(C0=0\.0C\_\{0\}=0\.0\) to fewer than 2 \(C0≥0\.8C\_\{0\}\\geq 0\.8\)\. Compared to an LLM\-based PauseAgent baseline, which issues many more questions,SteER’s cost\-sensitive mechanism achieves tighter control over the frequency of interruptions\. This suggests that base pause cost provides a direct and interpretable knob for regulating user burden\.

### F\.3SteER’s Persona Modeling Analysis

To assess the effectiveness ofSteER’s dynamic persona modeling, we examine how well the inferred persona aligns with the system’s final report over the course of interaction\. Specifically, we track the alignment score between the generated report and the inferred persona at different base pause cost \(C0C\_\{0\}\) settings, alongside the alignment between the report and the ground\-truth persona provided at the start\.

![Refer to caption](https://arxiv.org/html/2605.24266v1/figures/persona_analysis.png)Figure 7:Analysis of User Agent and Persona Modeling Performance across Base Pause Cost \(C0C\_\{0\}\)\.Left:User Agent precision, recall, andSteERreport alignment scores plotted across varying base pause cost values\.Right:Alignment scores ofSteER’s inferred persona and final report, both evaluated against the ground\-truth aspect set𝒜\\mathcal\{A\}, plotted across varying base pause cost values\.#### Report Alignment Tracks Persona Alignment

As shown in the right panel of Figure[7](https://arxiv.org/html/2605.24266#A6.F7), there is a strong positive correlation betweenSteER’s report alignment and the alignment of its inferred persona to the ground\-truth aspect set\. The Pearson correlation isr=0\.85r=0\.85\(p=8\.7×10−4p=8\.7\\times 10^\{\-4\}\), indicating that improvements in inferred persona accuracy are tightly coupled with improvements in report quality\. This supports the intuition thatSteER’s performance stems not only from architectural advances like mid\-process pausing, but also from its ability to incrementally build an accurate model of user goals\.

#### Impact of Base Pause Cost

We observe a general downward trend in both inferred persona alignment and report alignment as base pause cost increases\. This confirms that higher interruption costs reduce the frequency of clarifying interactions, resulting in less accurate persona estimates and, consequently, less aligned outputs\. In contrast, lowC0C\_\{0\}values allowSteERto query the user more frequently, leading to refined persona inference and stronger downstream alignment\.

These results highlight the central role of interactive refinement in personalized research workflows\. Rather than relying solely on upfront persona injection,SteERlearns about the user incrementally — and this process is empirically shown to improve alignment\. The correlation between inferred and actual persona alignment validates the design of our live persona model and its integration into the decision\-making process\.

### F\.4User Agent Performance Analysis

To enable scalable, automatic evaluation ofSteER, we employ a User Agent that simulates a real user interacting with the system\. This User Agent is responsible for selecting preferred research directions based on a target persona and proposing new follow\-up questions when relevant aspects remain uncovered\. Its effectiveness directly impacts the utility of our offline evaluation framework\.

As shown in the left panel of Figure[7](https://arxiv.org/html/2605.24266#A6.F7), the User Agent maintains consistently high precision across a wide range ofC0C\_\{0\}values, with scores above 0\.97\. This suggests that when the agent chooses to retain a direction, it is highly likely to align with the user’s intended aspects\. In contrast, recall is more sensitive to the pausing configuration\. At lowerC0C\_\{0\}\(e\.g\., 0\.1\), the User Agent achieves peak recall near 0\.95, but recall steadily declines asC0C\_\{0\}increases, falling to approximately 0\.85 byC0C\_\{0\}= 1\.0\. This reflects the agent’s conservative behavior under higher interruption costs, where it refrains from selecting additional directions that could be beneficial\.

We also observe that the alignment score of the final report generated bySteER\(in purple\) closely tracks the recall curve of the User Agent\. The Pearson correlation between the two is strong and statistically significant \(r=0\.81r=0\.81,p=2\.49×10−3p=2\.49\\times 10^\{\-3\}\), as annotated in the plot\. This indicates that the breadth of information the agent retains during interaction is highly predictive of the alignment quality of the final report\. The stronger the agent’s coverage of relevant aspects \(recall\), the more aligned the report tends to be with the user’s needs\.

These results confirm that the simulated User Agent is not only a faithful proxy for real user behavior but also a critical driver ofSteER’s alignment performance\. Its high precision ensures quality, while its recall effectively governs how much of the user’s goals are ultimately realized in the research output\.

## Appendix GUser Study Details

To complement automated evaluation, we conducted a human annotation study to directly assess how wellSteERreports align with user personas compared to baseline systems\. We developed a custom web\-based annotation platform \(Figure[8](https://arxiv.org/html/2605.24266#A7.F8)\) that guides annotators through a structured evaluation procedure with clear instructions and embedded report viewers\.

![Refer to caption](https://arxiv.org/html/2605.24266v1/figures/user_study_step03.png)
![Refer to caption](https://arxiv.org/html/2605.24266v1/figures/user_study_step12.png)

Figure 8:User study interface\.### G\.1Setup

Annotators were provided with apersona cardcontaining \(i\) the query, \(ii\) a short persona description, and \(iii\) the persona’sinterested aspects—the specific information needs that the final report should cover\. These interested aspects formed the primary basis of evaluation\. Annotators then evaluated two reports for the same query—persona pair: one generated bySteERand one by a baseline system \(either GPT\-Researcher or Open Deep Research\)\. Report order was randomized to reduce bias\.

### G\.2Evaluation Procedure

- •Step 1: Aspect Coverage\.Annotators skimmed both reports and rated, for each aspect, how well the report addressed it on a 3\-point scale:0 = not covered,1 = somewhat covered,2 = fully covered\. When assigning a score of 1 or 2, annotators were instructed to copy\-paste a short supporting quote \(1—2 sentences\) from the report to ground their judgment\. This ensured ratings were evidence\-backed rather than impressionistic\.
- •Step 2: Findability\.Annotators rated how easy it was to locate content relevant to each aspect in the report on a 3\-point scale:0 = difficult,1 = medium,2 = easy\. This step captured not only whether the aspect was present, but also whether it was readily discoverable by a reader\.
- •Step 3: Report Comparison\.Based on their coverage and findability assessments, annotators selected a winner between the two reports along two dimensions:Alignment\(which report better served the persona’s aspects\) andFocus\(which report stayed more on\-topic versus digressing into irrelevant content\)\.

#### Interface Design\.

The interface \(Figure[8](https://arxiv.org/html/2605.24266#A7.F8)\) displayed both reports side by side in embedded PDF viewers, alongside the persona’s aspects in a draggable panel for quick reference\. Each evaluation step was clearly separated into dedicated panels, with concise instructions and tips \(e\.g\., “You don’t need to read every word—scan section titles and opening sentences for relevant content”\)\. Progress indicators guided annotators through the sequence, ensuring consistency\. Importantly, the platform emphasized that judgments should be madefrom the persona’s perspective, not based on annotators’ personal preferences\.

#### Instructions and Quality Control\.

![Refer to caption](https://arxiv.org/html/2605.24266v1/figures/user_study_instruction.png)Figure 9:User study instructions\.The study followed a three\-step protocol:

As displayed in Figure[9](https://arxiv.org/html/2605.24266#A7.F9), annotators were instructed to:

1. 1\.Read the persona aspects carefully, treating them as the ground truth for evaluation\.
2. 2\.Provide evidence quotes for all non\-zero aspect coverage ratings\.
3. 3\.Complete all steps in sequence \(coverage→\\rightarrowfindability→\\rightarrowcomparison\)\.
4. 4\.Judge strictly by persona relevance, not by report verbosity, formatting, or personal opinion\.

These safeguards helped ensure high\-quality, reproducible annotations grounded in persona\-aligned judgments\.

#### Blinding considerations\.

Pairwise comparison in open\-ended generation cannot guarantee perfect blinding, and we treat this as a known limitation rather than a solved problem\. We took several steps to reduce identifiability risk: \(i\) annotators were shown only the two final reports, side by side in randomized order, with no access to clarification questions, user\-system interaction traces, persona updates, or any other intermediate system behavior; \(ii\) all systems produced report\-style outputs with broadly similar structure — e\.g\., tables of contents and section headings appeared across bothSteERand the baselines — so we did not observe consistent structural giveaways; and \(iii\) annotators were unfamiliar with our system or the baselines’ formatting conventions and were instructed to judge from the persona’s perspective rather than from stylistic preference\. Residual risk remains: a sufficiently distinctive lexical or structural fingerprint could in principle be guessed, and we cannot rule out that some annotators occasionally formed such a guess\. We therefore interpret the human\-study results as strong but not unconfounded evidence, and view fully blinded protocols \(e\.g\., post\-hoc style normalization or third\-party rewriting\) as useful future work\.

## Appendix HSteERWorking Prototype

![Refer to caption](https://arxiv.org/html/2605.24266v1/figures/demo.png)Figure 10:Interface ofSteERweb application\.To illustrate the functionality ofSteER, we build an interactive web\-based prototype \(Figure[10](https://arxiv.org/html/2605.24266#A8.F10)\) that visualizes theSteERframework in action\. The interface consists of three synchronized panels:\(i\)a conversation pane for clarification prompts and user feedback,\(ii\)a dynamically expanding research tree that reflects the research status and partial research results, and\(iii\)a live persona tracker that displays the evolving inferred personaP^\\hat\{P\}and monitors the updating alignment between cumulative research results and the inferred user aspects𝒜^\\hat\{\\mathcal\{A\}\}\. This prototype supports interactive research sessions, allowing users to guide the exploration by selecting preferred subtopics or introducing new follow\-up questions mid\-process\.

## Appendix ILLM\-as\-Judge Evaluation

To validate the effectiveness of the LLM judge used throughout evaluation, we conduct a small\-scale meta\-evaluation of the LLM\-as\-judge\. Specifically, we take the alignment score per aspect produced by the LLM judge \(*gpt\-4\.1\-mini*\) and the Coverage score per aspect produced by human annotators in the user study \(both in the scale of 0 \- 2\), and compute the Pearson correlationrrbetween LLM\-assigned and human\-assigned scores over all overlapping aspects where both annotations are available\. We obtainr=0\.34r=0\.34,p​\-value<0\.0001p\\text\{\-value\}<0\.0001, indicating a statistically significant, moderate positive correlation between the LLM\-as\-judge and human evaluations\. At the annotator level, all annotators exhibit positive correlations with the LLM judge, ranging from 0\.19 to 0\.45 with small variability \(standard deviation = 0\.094\), suggesting that the*gpt\-4\.1\-mini*’s scoring is consistently aligned with different human raters rather than being driven by any single annotator\. While imperfect, these results indicate that the LLM\-judge is directionally consistent with human judgments at the aspect level and is suitable as a scalable proxy for our large\-scale evaluation, especially when interpreted alongside the human study that directly validates our main claims\.

## Appendix JLimitations and Future Work

#### Simulated personas and queries vs\. users’ own queries\.

Our user study asks annotators to*mimic*a target persona and query rather than using their own information needs\. This is intentional: our goal is to evaluate system\-level behavior under controlled, persona\-conditioned information needs and to compareSteERagainst baselines on*exactly*the same persona–query pairs\. This requires\(i\)a fixed set of personas and queries that all systems answer and\(ii\)a shared reference for annotators, so that cross\-system differences in alignment and coverage can be attributed to the system rather than to heterogeneous user goals\. If each participant were to choose their own query and implicit persona, different systems would be evaluated on tasks of varying difficulty, domain, and specificity, making it difficult to perform clean system\-level comparisons and to interpret differences in outcomes as stemming from the model rather than from the task\. Moreover, obtaining enough repeated measurements per system–persona–task condition under fully free\-form user queries would require a substantially larger number of participants and interactions, constituting a much larger user\-study effort in terms of human resources, annotation time, and cost\. We acknowledge that this controlled design does not fully capture all aspects of real\-world usage, particularly long\-term adaptation to an individual’s genuine information needs\. As outlined in our limitations and future work, a natural next step is to conduct more comprehensive, end\-to\-end user studies in which participants bring their own personas and tasks, interact with the system over longer sessions, and are evaluated on richer metrics such as task success, time to insight, perceived control and trust, and cognitive load\.

#### Blinding in pairwise human evaluation\.

Our user study compares final reports fromSteERand a baseline side by side in randomized order \(Appendix[G](https://arxiv.org/html/2605.24266#A7)\)\. Annotators see only the final reports — never clarification turns, interaction traces, or intermediate system state — and we did not observe consistent structural cues separating systems\. However, perfect blinding is difficult in any open\-ended generation setting: lexical, formatting, or organizational tendencies of the underlying model and pipeline could in principle leak system identity\. We therefore view the pairwise preferences as supportive but not fully unconfounded evidence, complementing the controlled, persona\-conditioned automatic evaluation\. Stronger blinding \(e\.g\., style\-normalized rewrites, third\-party paraphrasing, or A/B/X protocols\) is a natural next step\.

#### Robustness to noisy and contradictory user feedback\.

Our User Agent is, by construction, a high\-fidelity simulator: it stays consistent with a fixed target persona, responds to every clarification, and does not fatigue\. Real users behave differently — they may give contradictory or self\-revising answers, supply partial or vague responses, ignore some prompts, or abandon the interaction entirely if asked too much\.SteER’s live persona module already handles*new*preferences by updating the inferred persona from incoming feedback, but it does not explicitly resolve*conflicts*between earlier and later signals, nor does it model the cost of disengagement when interactions accumulate\. We see three concrete extensions as important future work: \(i\) preference\-conflict handling that reconciles contradictory feedback over time, drawing on agent memory and preference\-tracking literature; \(ii\) graceful degradation under partial or missing responses, so the system can still steer with low\-confidence signals rather than treating silence as endorsement; and \(iii\) abandonment\-aware pause policies that incorporate user fatigue or drop\-out risk into the cost\-benefit rule\. Importantly, we view these as natural extensions ofSteER’s interactive paradigm rather than departures from it — the same framework that opens the door to mid\-process steering also opens the door to studying how users actually steer\.

#### Runtime and latency analysis\.

A key limitation of this work is that we do not provide a systematic runtime or latency analysis ofSteERrelative to baseline frameworks\. BecauseSteERis explicitly designed as an interactive deep\-research system, defining and comparing “runtime” in a meaningful way is non\-trivial: a naive wall\-clock measure would conflate\(i\)time spent on user interaction \(or User Agent responses in our simulations\) and\(ii\)the highly variable size of the research tree induced by different pause policies and user choices\. Simply averaging end\-to\-end runtimes could therefore be misleading —SteERmay spend more time within a single run while reducing wasted effort by correcting misalignment earlier, whereas non\-steerable baselines may require stacking multiple full deep\-research calls to reach a comparable result\. Moreover, even in a controlled environment like DeepResearchGym, end\-to\-end latency is heavily influenced by external factors such as search endpoint stability, network conditions, and rate limits, especially in multi\-agent, tool\-using pipelines\. As a result, raw wall\-clock differences are hard to attribute cleanly to the pause policy rather than infrastructure noise\. We view a careful efficiency study — using user\-centric measures such as time\- or turns\-to\-satisfactory\-answer, and stratifying by tree size, number of pauses, and backend conditions — as an important direction for future work, and plan to complement our current alignment and quality evaluations with such analyses in follow\-up studies\.

## Appendix KPrompt Templates inSteER

We include here the core prompt templates used in ourSteERframework, organized by functionality\. Each block shows thesystem promptand the correspondinguser prompt\. Placeholders such as\{query\},\{persona\_text\}, and\{checklist\_items\}are substituted at runtime\.

#### Research Planning

[⬇](data:text/plain;base64,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)Youareanexpertresearcherworkingwithaspecificuserpersona\.Yourtaskistoanalyzetheoriginalqueryandsearchresults,thengeneratetargetedquestionsthatexploredifferentdirectionsandtimeperiodsofthetopic,specificallytailoredtotheuser’sinterestsandchecklistitems\.User PromptOriginalquery:\\\{query\\\}Currenttime:\\\{current\_time\\\}Userpersona:\\\{persona\_text\\\}Userchecklist\(aspectstheycareabout\):\\\{checklist\_items\\\}Searchresults:\\\{search\_results\\\}Basedontheseresults,theoriginalquery,theuser’spersona,andtheirchecklist,generate5\-8uniquefollow\-upquestionsthat:1\.Exploredifferentdirectionsrelevanttothisquery2\.Coveragoodwiderangeoftopicsandaspectsofthequery3\.Considerrecentdevelopmentsupto\\\{current\_time\\\}4\.Aresomewhattailoredtotheuser’sbackgroundandneeds,butnotconstrainedbytheuser’spersonaandinterests5\.Eachfollow\-upquestionshouldcoveradistinctthematicfacet\-donotrepeatotherquestionsForeachquestion,provideaconfidencescorebetween0\.0and1\.0indicating:\-Relevanceofthequestiontothemainresearchquery\-Insightfulnessofthequestionthatwouldbeusefulforthefinalreportgeneration\-HowlikelythisquestionistoleadtovaluableinformationforthisuserReturnyourresponseasaJSONobjectwiththefollowingstructure:\\\{"follow\_up\_questions":\[\\\{"follow\_up\_question":"follow\-upquestiontext","confidence":0\.0\-1\.0,"reasoning":"whythisisagoodfollow\-upquestion"\\\}\]\\\}

#### Search Result Processing

[⬇](data:text/plain;base64,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)Youareanexpertresearcheranalyzingsearchresultsforaspecificuserpersona\.Focusonextractinglearningsandfollow\-upquestionsthataremostrelevanttotheuser’sinterestsandchecklistitems\.User PromptGiventhefollowingresearchresultsforthequery’\\\{query\\\}’,extractkeylearningsandsuggest5\-8follow\-upquestionsthatarespecificallyrelevanttotheuser’spersonaandinterests\.Userpersona:\\\{persona\_text\\\}Userchecklist\(aspectstheycareabout\):\\\{checklist\_items\\\}Previouslyseentags:\\\{seen\_tags\\\}Focuson:1\.Learningsthataddresstheuser’schecklistitems2\.Informationrelevanttotheirbackgroundandinterests3\.Follow\-upquestionsthatwouldhelpaddresstheirspecificneeds4\.Eachfollow\-upquestionshouldcoveradistinctthematicfacet\-donotrepeatotherquestionsForeachfollow\-upquestion,provideaconfidencescorebetween0\.0and1\.0indicating:\-Howlikelythisquestionistoleadtovaluableinformationforthisuser\-Alignmentwithuser’spersonaandchecklistitems\-RelevancetotheoriginalresearchqueryAdditionally,ALWAYSgenerateone"wild\-card"questionintheseparatewild\_card\_questionfieldthatgoesoutsidetheinferredpersonabutisplausiblyusefulforbroaderunderstandingofthetopic\.Additionally,assigntagstocategorizewhataspectsthisresearchcontentcovers\.Tagsshouldbeshortphrases\(2\-4words\)thatdescribethekeytopics,themes,ordomainscoveredbythequery,context,andlearnings\.Beverycautiousaboutaddingnewtags:\-FIRST,checkifanyofthepreviouslyseentagsarerelevanttothiscontent\-REUSEexistingtagswheneverpossible\-ONLYaddnewunseentagsifthecontentcoversaspectsnotcapturedbyexistingtags\-KeeptagsconciseanddescriptiveReturnyourresponseasaJSONobjectwiththefollowingstructure:\\\{"learnings":\[\\\{"insight":"keyinsightorfindingrelevanttotheuser","source\_url":"URLofthesource\(ifavailable\)","relevance\_to\_user":"howthislearningrelatestotheuser’sinterests"\\\}\],"follow\_up\_questions":\[\\\{"follow\_up\_question":"follow\-upquestiontext","confidence":0\.0\-1\.0,"reasoning":"whythisisagoodfollow\-upquestion"\\\}\],"wild\_card\_question":\\\{"question":"wild\-cardquestionthatgoesoutsidethepersonabutisplausiblyuseful","confidence":0\.0\-1\.0,"reasoning":"whythiswild\-cardquestioncouldbevaluable"\\\},"tags":\["tag1","tag2","tag3"\]\\\}Researchquery:\\\{query\\\}Searchresults:\\\{context\\\}

#### Follow\-up Questions to Search Queries

[⬇](data:text/plain;base64,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)Youareanexpertsearchqueryoptimizer\.Yourtaskistoconvertfollow\-upresearchquestionsintoeffectivesearchqueriesthatwillyieldrelevantsearchresults\.User PromptConvertthefollowingfollow\-upquestionintoanoptimizedsearchquerythatwillyieldrelevantsearchresults\.Originalresearchquery:\\\{original\_query\\\}Userpersona:\\\{persona\_text\\\}Userchecklist\(aspectstheycareabout\):\\\{checklist\_items\\\}Foreachofthefollowingfollow\-upquestion,createasearchquerythat:1\.Effectivelysearchesforinformationtoanswerthefollow\-upquestion2\.Isoptimizedforsearchengines3\.Maintainsconnectiontotheoriginalresearchquery4\.Considerstheuser’spersonaandinterestsForeachsearchquery,alsoprovideaclearresearchgoalthatdescribes:\-Whatspecificinformationorinsightsthissearchaimstodiscover\-Howitrelatestotheoriginalresearchquestion\-WhatdirectionofthetopicitwillexploreFollow\-upquestions:\\\{followup\_questions\\\}ReturnyourresponseasaJSONobjectwiththefollowingstructure:\\\{"search\_queries":\[\\\{"follow\_up\_question":"inputfollow\-upquestiontext","search\_query":"optimizedsearchquery","research\_goal":"cleardescriptionofwhatthissearchaimstodiscoverandhowitrelatestotheoriginalresearchquestion"\\\}\]\\\}

#### Persona Checklist Inference

[⬇](data:text/plain;base64,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)Youareanexpertatunderstandinguserpersonasandinferringwhataspectstheywouldcareaboutinresearchresponses\.Yourtaskistoanalyzeauser’spersonaandgeneratespecificchecklistitemstheywouldexpecttoseeaddressed\.User PromptGiventhefollowinguserpersonaandtheirresearchquery,inferachecklistofspecificaspectsthatthisuserwouldexpecttoseeaddressedinacomprehensiveresearchresponse\.Userpersona:\\\{persona\_text\\\}Researchquery:\\\{query\\\}Basedonthispersonaandquery,generate5\-8specificchecklistitemsthatthisuserwouldexpecttoseeinahelpfulresponse\.Eachitemshouldbe:1\.Specifictothisuser’sbackgroundandinterests2\.Relevanttotheresearchquery3\.Actionableandmeasurable4\.DistinctfromotheritemsReturnyourresponseasaJSONobjectwiththefollowingstructure:\\\{"checklist\_items":\["specificaspectthisuserwouldexpecttoseeaddressed","anotherspecificaspectrelevanttotheirinterests"\]\\\}

#### Persona Modeling

[⬇](data:text/plain;base64,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)Youareanexpertatunderstandinguserpersonasandupdatingthembasedonuserinteractions\.Yourtaskistoanalyzeauser’sresponseandinferadditionalinformationabouttheirpersonaandinterests\.User PromptGiventhefollowingcurrentpersonaandauser’sresponsetoaresearchproposal,inferadditionalinformationaboutthisuser’spersonaandinterests\.Currentpersona:\\\{current\_persona\\\}Currentchecklistitemstheyalreadycareabout:\\\{current\_checklist\\\}User’sresponse:\\\{user\_response\\\}Basedonthisresponse,identifyadditionalinformationabouttheuser’s:1\.Backgroundandinterests2\.Specificpreferencesandpriorities3\.Communicationstyleandconcerns4\.AnynewaspectstheycareaboutIMPORTANT:DoNOToutputrepetitiveinformation:\-OnlyincludeNEWpersonainformationthatisn’talreadycoveredinthecurrentpersona\-OnlyincludeNEWchecklistitemsthataren’talreadyinthecurrentchecklist\-Ifnothingnewcanbeinferred,returnemptystringsandemptyarraysReturnyourresponseasaJSONobjectwiththefollowingstructure:\\\{"additional\_persona\_info":"newinformationtoappendtothepersona\(emptyifnothingnew\)","new\_checklist\_items":\["newaspecttheymentionedorimpliedtheycareabout\(onlyifnotalreadyinchecklist\)","anothernewaspectifapplicable"\]\\\}

#### Clarification Question Generation

[⬇](data:text/plain;base64,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)Youareanexpertresearchassistant\.Yourtaskistogenerateclear,helpfulclarificationquestionsthatpresentconcisesummariesofresearchdirectionstousersforselection\.User PromptBasedonthefollowingresearchcontext,generateastructuredclarificationquestionthatpresentsconcisesummariesoftheavailableresearchdirectionstotheuserforselection\.Researchquery:\\\{query\\\}Availableresearchdirections\(searchqueries\):\\\{research\_directions\\\}Userpersona:\\\{persona\_text\\\}Createastructuredquestionthat:1\.Startswithanaturalintroduction2\.Listseachresearchdirectionasnumberedbulletpoints\(1\.,2\.,3\.,etc\.\)3\.Foreachdirection,provideaconcisesummary\(1sentence\)thatcapturestheessenceofwhatthatsearchquerywouldexplore,ratherthanshowingtherawsearchquery4\.Providesclearselectioninstructions:\-Toselectdirections:justtypethebulletnumbers\(e\.g\.,"1,3"\)\-Tosuggestnewfollow\-upquestions:startanewlinewith"Newfollow\-upquestions:"followedbyeachnewfollow\-upquestiononseparatelines5\.Matchestheuser’scommunicationstyleReturnyourresponseasaJSONobjectwiththefollowingstructure:\\\{"clarification\_question":"yourstructuredquestiontotheuserwithconcisesummaries"\\\}

#### Report Generation

[⬇](data:text/plain;base64,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)Youareaprofessionalresearchreportwriterspecializinginpersona\-awarereports\.Createcomprehensive,well\-structuredreportsbasedonresearchdatawithpropercitations,tailoredtothespecificuser’sbackgroundandinterests\.User PromptUsingthefollowinghierarchicallyresearchedinformationandcitations:"\\\{context\\\}"Writeacomprehensiveresearchreportansweringthequery:"\{question\}"Userpersona:\\\{persona\_text\\\}Userinterests\(checklistitemstheycareabout\):\\\{checklist\_items\\\}Thereportshould:1\.Synthesizeinformationfrommultiplelevelsofresearchdepth2\.Integratefindingsfromvariousresearchbranches3\.Presentacoherentnarrativethatbuildsfromfoundationaltoadvancedinsights4\.Maintainpropercitationofsourcesthroughout5\.Bewell\-structuredwithclearsectionsandsubsections6\.Haveaminimumlengthof\\\{total\_words\\\}words7\.Follow\\\{report\_format\\\}formatwithmarkdownsyntax8\.Usemarkdowntables,listsandotherformattingfeatureswhenpresentingcomparativedata,statistics,orstructuredinformation9\.Betailoredtotheuser’spersonaandinterestsAdditionalrequirements:\-Prioritizeinsightsthatemergedfromdeeperlevelsofresearch\-Highlightconnectionsbetweendifferentresearchbranches\-Includerelevantstatistics,data,andconcreteexamples\-Focusondirectionsthatalignwiththeuser’sinterestsandchecklist\-Uselanguageandexplanationsappropriatefortheuser’sbackground\-Addresstheuser’sspecificconcernsandpriorities\-YouMUSTdetermineyourownconcreteandvalidopinionbasedonthegiveninformation\.DoNOTdefertogeneralandmeaninglessconclusions\.\-YouMUSTprioritizetherelevance,reliability,andsignificanceofthesourcesyouuse\.Choosetrustedsourcesoverlessreliableones\.\-Youmustalsoprioritizenewarticlesoverolderarticlesifthesourcecanbetrusted\.\-Usein\-textcitationreferencesin\\\{report\_format\\\}formatandmakeitwithmarkdownhyperlinkplacedattheendofthesentenceorparagraphthatreferencesthemlikethis:\(\[in\-textcitation\]\(url\)\)\.\-Writein\\\{language\\\}Citationrequirements:\-YouMUSTwriteallusedsourceURLsattheendofthereportasreferences\-Makesuretonotaddduplicatedsources,butonlyonereferenceforeach\-EveryURLshouldbehyperlinked:\[urlwebsite\]\(url\)\-IncludehyperlinkstotherelevantURLswherevertheyarereferencedinthereport\-Formatexample:Author,A\.A\.\(Year,MonthDate\)\.Titleofwebpage\.WebsiteName\.\[urlwebsite\]\(url\)Pleasewriteathorough,well\-researchedreportthatsynthesizesallthegatheredinformationintoacohesivewhole,tailoredspecificallytothisuser’spersonaandinterests\.Assumethecurrentdateis\\\{current\_date\\\}\.

#### Persona Alignment Evaluation

[⬇](data:text/plain;base64,WW91IGFyZSBhbiBleHBlcnQgZXZhbHVhdG9yIHNwZWNpYWxpemluZyBpbiBhc3Nlc3NpbmcgaG93IHdlbGwgcmVzZWFyY2ggY29udGVudCBhbGlnbnMgd2l0aCB1c2VyIHBlcnNvbmFzIGFuZCBpbnRlcmVzdHMuIFlvdXIgdGFzayBpcyB0byBhbmFseXplIGNvbnRlbnQgYW5kIGRldGVybWluZSBob3cgd2VsbCBpdCBhZGRyZXNzZXMgc3BlY2lmaWMgZGlyZWN0aW9ucyBpbXBvcnRhbnQgdG8gdGhlIHVzZXIuCgooKkBcdGV4dGJme1VzZXIgUHJvbXB0fUAqKQpZb3UgYXJlIGV2YWx1YXRpbmcgaG93IHdlbGwgcmVzZWFyY2ggY29udGVudCBhbGlnbnMgd2l0aCBhIHVzZXIncyBwZXJzb25hIGFuZCBpbnRlcmVzdHMuCgojIFVzZXIgUGVyc29uYTogXHtwZXJzb25hX3RleHRcfQoKIyBSZXNlYXJjaCBDb250ZW50Ogpce2NvbnRlbnRcfQoKIyBLZXkgTGVhcm5pbmdzOgpce2xlYXJuaW5nc1x9CgojIENoZWNrbGlzdCBJdGVtcyB0byBFdmFsdWF0ZToKXHtjaGVja2xpc3RfaXRlbXNcfQoKRm9yIGVhY2ggY2hlY2tsaXN0IGl0ZW0sIGV2YWx1YXRlIGhvdyB3ZWxsIHRoZSByZXNlYXJjaCBjb250ZW50IGFuZCBsZWFybmluZ3MgYWRkcmVzcyBpdC4KUHJvdmlkZSBhIHNjb3JlIGZyb20gMC0yIGZvciBlYWNoIGl0ZW06Ci0gMDogTm90IGFkZHJlc3NlZCBvciBjb3ZlcmVkCi0gMTogU29tZXdoYXQgYWRkcmVzc2VkIG9yIHBhcnRpYWxseSBjb3ZlcmVkCi0gMjogV2VsbCBhZGRyZXNzZWQgb3IgdGhvcm91Z2hseSBjb3ZlcmVkCgpSZXR1cm4geW91ciBldmFsdWF0aW9uIGFzIGEgSlNPTiBvYmplY3Qgd2l0aCB0aGUgZm9sbG93aW5nIHN0cnVjdHVyZToKXHsKICAiZXZhbHVhdGlvbnMiOiBbCiAgICBcewogICAgICAiaXRlbSI6ICJjaGVja2xpc3QgaXRlbSB0ZXh0IiwKICAgICAgInNjb3JlIjogMC0yLAogICAgICAicmVhc29uaW5nIjogImJyaWVmIGV4cGxhbmF0aW9uIG9mIHRoZSBzY29yZSIKICAgIFx9CiAgXQpcfQ==)Youareanexpertevaluatorspecializinginassessinghowwellresearchcontentalignswithuserpersonasandinterests\.Yourtaskistoanalyzecontentanddeterminehowwellitaddressesspecificdirectionsimportanttotheuser\.User PromptYouareevaluatinghowwellresearchcontentalignswithauser’spersonaandinterests\.\#UserPersona:\\\{persona\_text\\\}\#ResearchContent:\\\{content\\\}\#KeyLearnings:\\\{learnings\\\}\#ChecklistItemstoEvaluate:\\\{checklist\_items\\\}Foreachchecklistitem,evaluatehowwelltheresearchcontentandlearningsaddressit\.Provideascorefrom0\-2foreachitem:\-0:Notaddressedorcovered\-1:Somewhataddressedorpartiallycovered\-2:WelladdressedorthoroughlycoveredReturnyourevaluationasaJSONobjectwiththefollowingstructure:\\\{"evaluations":\[\\\{"item":"checklistitemtext","score":0\-2,"reasoning":"briefexplanationofthescore"\\\}\]\\\}

## Appendix LPrompt Templates for Data Generation

We include here the prompt templates used in data generation\.

#### Profile Generation Prompt

[⬇](data:text/plain;base64,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)Generateauserprofileforsomeonewhowouldlogicallyandreasonablyaskthefollowingquestion:"\\\{query\\\}"Theprofileshouldincludedemographicandbackgroundinformationsuchasagerange,occupation,hobbies,familystructure,educationbackground,oranyotherrelevantfacts\.Notethatyoudon’tneedtoincludeallofthesedetailsforeachpersona\.Youcanuseanykindsofcombinationsandpleasethinkaboutotheraspectsotherthanthese\.Youshouldincludesomethingthatcanbeelicitedfromdailyandnaturalconversations\.Youshouldnotincludetoomuchinformationaboutthisperson’sworkcontentandyoushouldnotgiveanydescriptionabouttheuser’spersonalitytraits\.Focusonobjectivefactsabouttheperson\.Herearesomeexampleprofilesforreference:\\\{profile\_examples\\\}Generateasingleuserprofilethatcontainsaround8\-10distinctfactsabouttheperson\.Theprofileshouldlogicallyconnecttowhythispersonwouldaskthegivenquestion\.Youshouldonlyoutputtheprofileinplaintextformat\.IMPORTANT:Trytobecreativeandcomprehensive\.Makesuretheprofilemakesitrealisticforthispersontoaskthespecificquestion\.

#### Personality Generation Prompt

[⬇](data:text/plain;base64,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)Generatepersonalitytraitsforapersonwiththefollowingprofilewhowouldaskthisquestion:"\\\{query\\\}"Profile:\\\{generated\_profile\\\}Basedonthisprofileandthequestiontheywouldask,generateappropriatepersonalitytraits\.Youshouldincludesomethingthatcanbeelicitedfromdailyandnaturalconversations\.Eachdescriptionshouldcontainaround8\-10personalitytraitsabouttheperson\.Herearesomeexamplepersonalitydescriptionsforreference:\\\{personality\_examples\\\}Generatepersonalitytraitsthatareconsistentwiththeprofileandmakeitlogicalforthispersontoaskthegivenquestion\.Youshouldonlyoutputthepersonalitydescriptioninplaintextformat\.IMPORTANT:Youshouldnotincludeanyothercontentthatisbeyondpersonalitytraits,suchasoccupationordemographicinformation\(thosearealreadyintheprofile\)\.Focusonlyonpersonalitycharacteristics,behavioralpatterns,andpsychologicaltraits\.Becreativeandmakesurethepersonalityalignswithboththeprofileandtheresearchquestion\.

#### Aspect Generation Prompt

[⬇](data:text/plain;base64,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)Givenauser’spersonaandtheirquery,generatealistofspecificaspectsthatthisuserwouldexpecttoseeaddressedinahigh\-qualityresponsetotheirquery\.Theseaspectswillserveasevaluationcriteriatoassesshowwellaresponsemeetsthisspecificuser’sneedsandexpectations\.Query:"\\\{query\\\}"UserPersona:\\\{persona\\\}Basedonthispersonaandquery,generate5\-8specificaspectsthatthisuserwouldexpecttoseeinacomprehensiveandhelpfulresponse\.Eachaspectshouldbe:1\.Specifictothisuser’sbackground,needs,andcontext2\.Actionableandmeasurable\(canbeusedtoevaluatearesponse\)3\.Relevanttothequeryandpersona4\.Distinctfromotheraspects\(nooverlap\)FormatyourresponseinJSONformatwhereeachaspectisaclear,specificexpectationthatcanbeusedtoevaluatewhetheraresponseadequatelyaddressesthisuser’sneedsandprovideaclearexplanationofwhyeachaspectissignificantfortheuserandwhatspecificdetailstheywouldexpecttoseeintheresponse\.Focusonwhatcontent,depth,style,orapproachwouldbemostvaluableforthisspecificuser\.Eachaspectshouldhavethefollowingfields:\-aspect:astringthatisthenameoftheaspectthatisimportanttobepresentintheresponse\-evidence:astringthatpointstospecificdetailsfromtheuser’spersonathatindicatethisaspectisimportant\-reason:astringthatexplainswhytheaspectisimportantfortheuserUsethefollowingJSONstructure:\\\{"aspects":\[\\\{"aspect":"Nameoftheaspect","evidence":"Specificdetailsfromtheuser’spersonathatindicatethisaspectisimportant","reason":"Explanationofwhythisaspectisimportantfortheuser"\\\}\]\\\}IMPORTANT:Maketheaspectsspecifictothisuser’suniquesituationandneeds,notgenericaspectsthatwouldapplytoanyuseraskingthisquestion\.

## Appendix MPrompt Templates for Evaluation

We include here the prompt templates used in evaluation scripts\.

#### Alignment Evaluation Prompt

[⬇](data:text/plain;base64,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)Youareafairandinsightfuljudgewithexceptionalreasoningandanalyticalabilities\.Yourtaskistoevaluateauser’squestion,ageneratedresponsetothatquestion,andmultipleaspectsthatareimportanttotheuser\.Basedonthisinformation,assesshowwelleachaspectisaddressedinthegeneratedresponse\.Provideaclearandaccurateassessmentforeachaspect\.\#Yourinput:\-question:thequestionaskedbytheuser\-persona:theuser’spersona\(profileandpersonality\)thattheaspectsarebasedon\-response:ageneratedresponsetotheuser’squestion\-aspects:alistofaspectsthatareimportanttotheuser,eachconsistingof:\-aspect:thetitlefortheaspect\-reason:thereasonthatthisaspectisimportantfortheuser\-evidence:theevidencefromtheuserpersonathattheaspectwasextractedfrom\#Youroutput:YouroutputshouldbeavalidJSONobjectin‘‘‘json‘‘‘formatcontainingthefollowingfields:\-evaluations:Alistofevaluationsforeachaspect,whereeachevaluationcontains:\-aspect:theaspecttitle\-match\_score:Ascorebetween0to2thatindicateshowwellthegeneratedresponseaddressesthisaspect,where:\*0meanstheresponsedoesnotcoverthisaspect\*1meanstheresponsesomewhatcoversthisaspect\*2meanstheresponsecoversthisaspectverywell\-reasoning:Adetailedexplanationofwhythisscorewasassigned,includingspecificexamplesfromtheresponse\#Question:\\\{question\\\}\#Persona:\\\{persona\\\}\#Response:\\\{response\\\}\#Aspects:\\\{aspects\_formatted\\\}Output:

#### Sentence Focus Evaluation Prompt

[⬇](data:text/plain;base64,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)Youareanexpertjudgeevaluatingwhethersentencesinareportcoverspecificuseraspects\.Foreachsentence,determinewhichaspects\(ifany\)itaddresses\.\#Yourinput:\-sentences:numberedsentencesfromareport\-aspects:useraspectswithIDs,titles,andreasons\#Yourtask:Foreachsentence,identifywhetheritcoversanyoftheuseraspects\.\*\*BEEXTREMELYSTRICT\*\*inyourevaluation\.AsentencecoversanaspectONLYifit:1\.Directlyaddressesthespecificconcernorinterestdescribedintheaspect2\.Providessubstantive,detailedinformationthatwouldbevaluabletosomeonewiththatspecificaspect3\.GoesbeyondmerekeywordmentionsorgeneralbackgroundinformationAsentencedoesNOTcoveranaspectifit:\-Onlyprovidesgeneralbackgroundorintroductoryinformation\-Mentionskeywordsrelatedtothetopicbutdoesn’taddressthespecificconcern\-Givesbroadoverviewswithouttargetingtheparticularinterest\-Describesgeneralprincipleswithoutconnectingtothespecificaspect\-Isjustfactualinformationthatdoesn’tservetheuser’sparticularneed\*\*DefaulttoNOTcoveringaspectsunlessthereisclear,direct,substantialrelevancetothespecificuserconcern\.\*\*\#JSONSchemaforoutput:\\\{"type":"object","patternProperties":\\\{"^d\+\\$":\\\{"type":"object","properties":\\\{"cover\_aspects":\\\{"description":"AlistofaspectIDsthatthesentencecovers\.Ifthesentencedoesnotcoveranyoftheaspects,thelistshouldbeempty\.","type":"array","items":\\\{"type":"integer"\\\}\\\},"reasoning":\\\{"type":"string"\\\}\\\},"required":\["cover\_aspects","reasoning"\]\\\}\\\}\\\}\#Sentences:\\\{sentences\_formatted\\\}\#Aspects:\\\{aspects\_formatted\\\}OutputvalidJSON:

#### Key Point Extract Prompt

[⬇](data:text/plain;base64,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)Basedonthereportprovided,identifykeypointsinthereportthatdirectlyhelpinrespondingtothequery\.Thekeypointsarenotsimplysomekeycontentofthetext,butratherthekeypointsthatareimportantfor\*\*answeringthequery\*\*\.IMPORTANT:Ensureeachpointishelpfulinrespondingtothequery\.Keepthepointusingtheoriginallanguageanddonotaddexplanations\.IMPORTANT:Eachspanmustbeasingleconsecutiveverbatimspanfromthecorrespondingpassages\.Copyverbatimthespans,don’tmodifyanyword\!Yourresponseshouldstatethepointnumber,followedbyitscontent,andspansinthetextthatentailthekeypoint\.RespondstrictlyinJSONformat:\\\{"points":\[\\\{"point\_content":point\_content,"spans":\[span1,span2,\.\.\.\]\\\},\.\.\.\]\\\}Remember:\-Keypointscanbeabstractedorsummarized,butthespanmustbeacopyoftheoriginaltext\.ThecontentofthekeypointdoesNOTneedtobethesameasthatofthespan\.\-Thesekeypointsmustbehelpfulinrespondingtothequery\.\-Iftherearemultiplespansforapoint,addalloftheminthespanslist\.Report:\\\{report\\\}Query:\\\{query\\\}Output:

#### Key Point Focus Evaluation Prompt

[⬇](data:text/plain;base64,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)Youareanexpertjudgeevaluatingwhetherkeypointsofareportcoverspecificuseraspectstoansweraquery\.Foreachkeypoint,determinewhichaspects\(ifany\)itaddresses\.\*\*BEEXTREMELYSTRICT\*\*inyourevaluation\.AkeypointcoversanaspectONLYifit:1\.Directlyaddressesthespecificconcernorinterestdescribedintheaspect2\.Providessubstantive,detailedinformationthatwouldbevaluabletosomeonewiththatspecificaspect3\.GoesbeyondmerekeywordmentionsorgeneralbackgroundinformationAkeypointdoesNOTcoveranaspectifitonlyprovidesintroductoryinformationorbroadoverviews\*\*DefaulttoNOTcoveringaspectsunlessthereisclear,direct,substantialrelevancetothespecificuserconcern\.\*\*ResponsestrictlyinJSONformat:\\\{"point\_number":\\\{"cover\_aspects":\[aspect1,aspect2,\.\.\.\],"reasoning":reasoning\\\},\\\}\#Query:\\\{query\\\}\#ReportKeyPoints:\\\{key\_points\_formatted\\\}\#UserAspects:\\\{aspects\_formatted\\\}Output:

#### User Agent Alignment Evaluation Prompt

[⬇](data:text/plain;base64,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)Youareafairandinsightfuljudgewithexceptionalreasoningandanalyticalabilities\.Yourtaskistoevaluateauser’sfollow\-upquestionsinregardtoaquery,andmultipleaspectsthatareimportanttotheuser\.Basedonthisinformation,assesshowwellthefollow\-upquestionstryingtocovertheuser’sinterestedaspects\.Anaspectisconsideredcoverediftherearefollow\-upquestionsaretryingtoinitiateresearchdirectionsthatarerelatedtotheaspect\.Provideaclearandaccurateassessmentforeachaspect\.\#Yourinput:\-query:thequeryaskedbytheuser\-persona:theuser’spersona\(profileandpersonality\)thattheaspectsarebasedon\-follow\-upquestions:alistoffollow\-upquestionsthattheuserasked\-aspects:alistofaspectsthatareimportanttotheuser,eachconsistingof:\-aspect:thetitlefortheaspect\-reason:thereasonthatthisaspectisimportantfortheuser\-evidence:theevidencefromtheuserpersonathattheaspectwasextractedfrom\#Youroutput:YouroutputshouldstrictlybeavalidJSONobject:\\\{"evaluations":\[\\\{"aspect":aspect,"match\_score":match\_score,"reasoning":Adetailedexplanationofwhythisscorewasassigned,includingspecificexamplesfromthefollow\-upquestions\\\},\.\.\.\]\\\}"match\_score"isascorebetween0to2thatindicateshowwellthefollow\-upquestionsaddressesthisaspect,where:\*0meansthefollow\-upquestionsdoesnotcoverthisaspect\*1meansthefollow\-upquestionssomewhatcoversthisaspect\*2meansthefollow\-upquestionscoversthisaspectverywell\#Query:\\\{query\\\}\#Persona:\\\{persona\\\}\#Follow\-upQuestions:\\\{follow\_up\_questions\_formatted\\\}\#Aspects:\\\{aspects\_formatted\\\}Output:

#### User Response Precision Evaluation Prompt

[⬇](data:text/plain;base64,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)Youareanexpertjudgeevaluatingwhetherauser’sfollow\-upquestionsorresponsesaretrulytargetedtospecificuseraspectsforansweringaquery\.Foreachfollow\-up,determinewhichaspects\(ifany\)itsubstantivelytargets\.BEEXTREMELYSTRICT\.Afollow\-upCOVERSanaspectONLYifit:1\)Clearlyaimstogatherinformationdirectlyrelevanttothespecificconcerndescribedbytheaspect;AND2\)Goesbeyondsurfacekeywordsorgenericcuriosity\.Afollow\-updoesNOTcoveranaspectifit:\-Isabroad/backgroundquestionwithouttailoringtothataspect;OR\-Onlymentionsrelatedkeywordsbutlacksatargetedobjectivetiedtotheaspect;OR\-Isunrelatedtotheuser’sstatedconcerns\.RespondstrictlyinJSONformat:\\\{"response\_number":\\\{"cover\_aspects":\[aspect\_id\_1,aspect\_id\_2,\.\.\.\],"reasoning":reasoning\\\},\.\.\.\\\}\#Query:\\\{query\\\}\#UserResponses\(indexedfrom0\):\\\{user\_responses\_formatted\\\}\#UserAspects\(IDsstartat0\):\\\{aspects\_formatted\\\}Output:

#### Final Persona State Evaluation Prompt

[⬇](data:text/plain;base64,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)Youareafairandinsightfuljudgewithexceptionalreasoningandanalyticalabilities\.Yourtaskistoevaluatehowwellitemsfromafinalpersonastatechecklistcoveruseraspects\.Giventheuser’squery,theoriginalpersona,alistofchecklistitems,andtheuseraspects,assessforeachaspecthowwellthechecklistcoversit\.\#Yourinput:\-query:thequeryaskedbytheuser\-persona:theuser’soriginalpersonatext\-checklist:alistofitemsinferredthatmightbeimportantfortheusertoanswerthequery\-aspects:alistofaspectsthatareindeedimportanttotheuserasgroundtruth,eachconsistingofaspect,reason,andevidence\#Youroutput:ReturnstrictlyvalidJSONoftheform:\\\{"evaluations":\[\\\{"aspect":aspect\_title,"match\_score":0\|1\|2,"reasoning":detailed\_reasoning\_referencing\_specific\_checklist\_items\\\},\.\.\.\]\\\}Interpretmatch\_scoreas:\-0:thechecklistdoesnotcoverthisaspect\-1:thechecklistsomewhatcoversthisaspect\-2:thechecklistcoversthisaspectverywell\#Query:\\\{query\\\}\#Persona:\\\{persona\\\}\#ChecklistItems:\\\{checklist\_formatted\\\}\#Aspects:\\\{aspects\_formatted\\\}Output:

## Appendix NAdditional Prompt Templates

We include here the prompt templates used for User Agent and Pause Agent\.

#### User Agent

[⬇](data:text/plain;base64,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)Youaresimulatingarealuserwithaspecificpersonaandinterests\.YourtaskistorespondtoSteER’sresearchproposalsbyselectingrelevantdirectionsandsuggestingnewdirectionsbasedonyourpersonaandresearchinterests\.User PromptYouareactingasauserwiththefollowingpersona:\#UserPersona:\\\{persona\_text\\\}\#AspectsanddirectionsYouCareAbout:\\\{aspects\_text\\\}\#Historyofyourpreviousaskedquestions:\\\{questions\_history\_text\\\}\#ResearchQuery:\\\{query\\\}\#SteER’sProposal:\\\{steer\_proposal\\\}SteERispresentingresearchdirectionsasnumberedbulletpoints\.Basedonyourpersonaandinterests,respondasthisuserwouldby:1\.SelectingONLYthemostrelevantdirectionnumbersthathavethehighestpriorityforthisresearch2\.Suggestingnewfollow\-upquestionsONLYifyoufeelthere’saveryimportantdirectionmissingfromtheproposal3\.Providingnaturalcommentaryasthisuserwouldspeak\*\*IMPORTANTCONSTRAINTS:\*\*\-\*\*DONOTselectdirectionsorsuggestquestionsthatareoutsideyourpersonaandaspects/interests\*\*\-\*\*DONOTsuggestquestionsyouhavealreadyaskedbeforeorthataresimilartothequestionsyouhavealreadyasked\(checkyourhistoryabove\)\*\*\-Onlyfocusonareasthatalignwithyourspecificexpertise,interests,andconcernsasdescribedinyourpersona\-Ifallcurrentdirectionsseemunrelatedtoyourinterests,it’sbettertoselectnoneandsuggestrelevantalternativesFocusonqualityoverquantity\-selectonlythedirectionsthattrulymattermosttoyouandalignwithyourexpertise\.Youshouldrefrainfromsuggestingnewfollow\-upquestionsunlesssomethingcriticalismissinganddirectlyrelatestoyourinterests\.Youshouldatmostsuggest1newfollow\-upquestion\.Theprobabilityofyousuggestinganewfollow\-upquestionis50%\.Yourresponseshouldreflecthowthispersonwouldactuallycommunicatewhendiscussingtheirresearchpreferences\.ReturnyourresponseasaJSONobjectwiththefollowingstructure:\\\{"selected\_directions":\[\\\{"number":1,"direction":"directionnamefromtheproposal","reasoning":"whythisdirectionismostimportanttoyouandalignswithyourinterests"\\\}\],"new\_follow\_up\_questions":\[\\\{"follow\_up\_question":"suggestednewfollow\-upquestion\.Mostofthetimeyoushouldnotsuggestnewfollow\-upquestions\.Butonlyifyoufeelthere’saveryimportantdirectionmissingfromtheproposal,suggestonenewfollow\-upquestionatmost","reasoning":"whythisfollow\-upquestionisimportant,missing,andrelevanttoyourinterests"\\\}\],"user\_response":"naturalresponseasthisuserwouldspeak\(intheformat:selectednumberswithreasoninginparentheses,then’Newfollow\-upquestions:’ifany\)","additional\_context":"anyadditionalpreferencesorclarificationsrelatedtoyourexpertise"\\\}

#### Pause Agent

[⬇](data:text/plain;base64,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)Youareanexpertresearchassistantspecializedinmakingoptimalpausedecisionsduringdeepresearch\.Yourtaskistoanalyzethecurrentresearchstateanddecidewhetherit’sagoodtimetopauseandaskforuserguidanceonwhichresearchbranchestopursue,ortoproceedwiththecurrentresearchplan\.User PromptYouneedtodecidewhethertopauseandaskforuserguidanceorproceedwiththecurrentresearchplan\.\#Context:\*\*OriginalQuery:\*\*\\\{original\_query\\\}\*\*CurrentResearchGoal:\*\*\\\{current\_research\_goal\\\}\*\*UserPersona:\*\*\\\{persona\_text\\\}\*\*UserInterests\(Checklist\):\*\*\\\{checklist\_items\\\}\*\*CurrentSearchDepth:\*\*\\\{current\_depth\\\}/\\\{max\_depth\\\}\#AvailableResearchBranches:\{branch\_summaries\}\#DecisionCriteria:Considerpausing\(PAUSEASK\)when:\-Userinputwouldhelpprioritizewhichdirectiontopursue\-There’suncertaintyaboutwhichdirectionalignsbestwiththeuser’sspecificinterestsAnalyzethesituationandmakeyourdecision\.Yourreasoningshouldbespecifictothecurrentresearchcontext,userpersona,andbranchcharacteristics\.RespondinthefollowingJSONformat:\\\{"type":"object","properties":\\\{"action":\\\{"type":"string","enum":\["PROCEED","PAUSEASK"\],"description":"Decisiontoproceedwithresearchorpausetoaskuserforguidance"\\\},"reasoning":\\\{"type":"string","description":"Detailedexplanationofthedecisionbasedonresearchcontextanduserpersona"\\\}\\\},"required":\["action","reasoning"\]\\\}