Personalized Deep Research Query Refinement with Graph-Scaffolded Evidence Grounding

Paper: arXiv:2608.05876v1 [cs.AI], 6 Aug 2026

---

The paper addresses the challenge of Personalized Deep Research (PDR), where real-world research needs are rarely generic. Even for the same request, users may expect different reports depending on their goals, constraints, preferences, and evaluation criteria. Deep Research Agents (DRAs) have expanded information seeking from answering isolated queries to conducting multi-step investigations and generating comprehensive reports, but user requests serve as research specifications that shape what evidence to seek and how to synthesize it. In personalized deep research, these specifications must additionally reflect user goals, constraints, preferences, and evaluation criteria.

The authors distinguish two approaches to personalization: incorporating user context either within the DRA's internal pipeline or through its input research specification. They focus on the latter—input-level refinement, where a refiner uses the initial query, user profile, persistent memory, and optional user interaction to produce a personalized research specification for an unchanged DRA. This approach is motivated by the fact that pipeline-integrated personalization requires access to internal components and system-specific adaptation, limiting its applicability to proprietary black-box DRAs.

The authors contrast their setting with prior clarification work: "Whereas prior clarifiers primarily recover latent or missing task information through user interaction, PDR query refinement starts from a request that may already be sufficient for generic research and must determine which user-specific factors should alter its framing. A relevant factor may already be supported by the user profile or persistent memory, require direct clarification, or not warrant further acquisition. The refiner must therefore distinguish relevant from distracting context, grounded from weakly supported factors, and worthwhile from unnecessary interaction."

The paper formulates PDR query refinement as three coupled decisions:

1. Framing Relevance — which user-specific factors should shape the final query

2. Evidential Sufficiency — whether the accumulated evidence sufficiently supports each factor

3. Grounding Control — whether to acquire additional evidence or stop and refine the query

As illustrated in the paper's Figure 1, these decisions evolve jointly. Resolving one factor may make another relevant, retrieved memory may remove the need for a question, and partial evidence may instead require direct clarification.

Formally, given an initial user query q and a static user profile p u, personalized query refinement aims to produce a refined query q* that preserves the original intent while making relevant user-specific goals, constraints, preferences, and background explicit. At step t, the agent conditions on q, p u, and an accumulated evidence state Σ t, which records the grounding status, supported value, and evidence source of each framing factor considered so far. The agent selects an action a t ∈ Retrieve, Ask, Stop, where Retrieve searches persistent user memory, Ask elicits evidence from the user, and Stop terminates refinement and generates the refined query. After Retrieve or Ask, the observation updates the evidence state via Σ t+1 = Update(Σ t, a t, o t+1). The policy is π θ(a t q, p u, Σ t), and when it selects Stop at step T, q* = g(q, p u, Σ T).

The proposed framework, G-STEER, is described as a training framework that constructs Graph-Scaffolded Trajectories for learning EvidencE-grounding decisions toward personalized deep research query Refinement. It operates in two stages: first, it organizes user-specific elicitation targets into an Intent Elicitation Graph (IEG) and constructs trajectories that vary target dependencies, evidence availability, and stopping conditions; second, it then optimizes the clarification policy for target coverage and interaction efficiency. At inference, the policy selects Retrieve, Ask, or Stop from the evolving evidence state without access to the elicitation targets or IEG.

Candidate elicitation targets are extracted from the initial query and its personalization rubric items, where rubric items are split only when they contain separable decisions that can independently change downstream planning, and the aspects that would naturally be resolved together remain merged. Each candidate is retained only when its answer can vary across users, materially affect downstream planning, and is not already specified by q. Each retained target is annotated as essential when leaving it unresolved would likely omit a major decision branch, apply inappropriate comparison criteria, or produce materially underspecified recommendations.

Each target e i = (g i, q i seed, a i seed, w i, η i) consists of the target description, seed question–answer pair, rubric-derived importance weight, and essentiality label.

Intent Elicitation Graph (IEG): For each instance, an IEG is constructed "as a directed acyclic graph G = (V, D), where V = E and the initial query acts as an external root. An edge (e i, e j) ∈ D indicates that resolving e i can change whether or how e j should be elicited. The graph therefore captures clarification dependencies rather than topical similarity. A target is marked as late-activatable when it is initially inactive and becomes eligible only after one of its parents has been covered."

Scenario-Controlled Trajectories: Trajectory construction begins with a backbone subgraph containing all essential targets and their ancestors. Additional trajectories include deterministic subsets of optional targets at ratios of 0, 0.5, 1.0. Targets are ordered so no child precedes its parent, and when several targets are eligible, optional targets closer to the essential backbone are prioritized. Each IEG is instantiated into scenarios that vary the target schedule, memory hits, late activation, and stopping point. The trajectory "first retrieves evidence for initially active targets. After each memory or dialogue observation, the corresponding target slots are updated. A late-activatable target remains inactive until an eligible parent slot is covered. Once activated, memory is checked before asking the user. Otherwise, the earliest uncovered target in the order is selected. Each Ask action addresses one target at a time to reduce user burden. At each step, Plan summarizes state changes and explains the next action and Execute produces a retrieval query, clarification question, or refined query for Retrieve, Ask, or Stop. At Stop, the teacher model refines the initial query using the final evidence state."

Policy Warm-up: "The policy is fine-tuned on these trajectories to learn how to use the current evidence, choose between memory and user interaction, activate dependent targets, stop adaptively, and produce the required output for each action. This initialization reduces the burden on reinforcement learning to discover both the interaction protocol and valid clarification behavior."

The policy is refined with trajectory-level Group Relative Policy Optimization (GRPO). Each trajectory receives reward:

R(τ, E) = R align(τ, E) − C acq(τ) − C con(τ)

The alignment reward is R align(τ, E) = α Cov(τ, E) − β C FP(τ, E), where coverage is the normalized sum of important-target weights over one-to-one matched covered slots (using L2-normalized encoder similarities), and the false-positive cost penalizes covered slots that are unmatched or weakly aligned with the elicitation targets.

The evidence acquisition cost is C acq(τ) = λ ask T ask + λ ret T ret, with "λ ask > λ ret because asking the user imposes greater burden than retrieving from memory." Behavioral penalties are C con(τ) = λ rep n rep(τ) + λ multi n multi(τ), penalizing repeated acquisition attempts and Ask actions that address multiple target slots or combine multiple question intents.

Online User State Update: After each clarification trajectory, "reusable evidence is selected from the final covered slots. A slot is retained only when it is directly supported, specific to the user, useful for future clarification or personalization, and not already captured by the current profile. Generic content, assistant advice, weak inferences, temporary details, and unnecessarily sensitive information are excluded. Selected slots are stored as retrieval memories; the profile is updated with a one-sentence summary only when they support a durable, broadly reusable user pattern."

At inference, the offline elicitation targets used for training and evaluation, as well as the IEG used to construct training trajectories, are not provided. The evidence state is represented at runtime as a textual set of slots for user-specific framing factors, where each slot records the factor and its purpose, current grounding status, supported value, and evidence source. The policy initializes Σ 0 by inducing these slots through prompting rather than copying them from offline annotations. When Stop is selected, the policy generates q* from q, p u, and the final evidence state Σ T, with the prompt instructing it to incorporate only slot values supported by the available evidence.

Dataset: The authors evaluate on the processed PDR-Bench, split into 105 training instances (70%) and 45 test instances (30%), totaling 150 instances across 25 users and 30 tasks, with 1,334 total elicitation targets (8.89 average per instance). The test set comprises four groups: seen user/task (13 instances), user unseen (12), task unseen (10), and both unseen (10).

Models: Qwen3-8B serves as the backbone, fine-tuned with LoRA. Qwen3-32B renders SFT trajectories and simulates users, while GPT-5.4 constructs elicitation targets and IEGs. Semantic matching uses sentence-transformers/all-mpnet-base-v2, and memory indexing uses FAISS.

Evaluation metrics: Precision, recall, and F1 against reference elicitation targets, plus weighted coverage (WCov)—the normalized sum of PDR-Bench rubric-derived importance weights over covered targets—and E-WCov, its restriction to essential targets. A target is covered if its maximum cosine similarity to an extracted factor is at least 0.55.

Baselines: Generic prompting methods (Direct Rewrite, Memory-RAG Agent, Clarification Agent, Memory+Clarification Agent), prior clarification methods (Mistral-Interact, CEP-Clarify+Rewrite, IntentRL-style Clarifier), and ablations. The IntentRL-style baseline retains proactive clarification from dependency-structured trajectories but is limited to Ask/Stop and refines the query from dialogue alone, without slot tracking.

G-STEER achieves the best overall performance across all coverage and matching metrics, with particularly strong results under user shift. The full results show G-STEER achieving WCov 0.4253, E-WCov 0.4639, precision 0.4199, recall 0.3971, and F1 0.4082, compared to the strongest baseline (CEP-Clarify+Rewrite) at WCov 0.4060, E-WCov 0.4497, and F1 0.3823. Crucially, G-STEER achieves stronger performance than CEP-Clarify+Rewrite while asking 4.09 user questions on average, compared with 12.04 for CEP-Clarify+Rewrite, roughly one-third as many user questions. The untrained Memory+Clarification baseline never invokes Retrieve despite access to both acquisition actions, showing an Ask bias rather than effective evidence-source routing.

Ablation results show that the untrained Qwen3-32B model performs worse than Qwen3-8B across all metrics, showing that a larger backbone alone does not guarantee better selection of personalization factors. Graph SFT alone achieves the highest F1, and the full model improves WCov by 0.0138 and E-WCov by 0.0227 over Graph SFT while reducing F1 by 0.0166, interpreted as GRPO shifting the policy toward targets assigned greater importance rather than uniformly improving all matching metrics. Critically, randomizing target order while preserving the same target set sharply reduces all metrics, leading to the conclusion that trajectory supervision following IEG dependencies substantially improves factor acquisition.

Comparing E-WCov against the average number of questions and its improvement per question over Direct Rewrite, "G-STEER achieves the highest E-WCov and the largest improvement per question, outperforming IntentRL-style at comparable burden and CEP-Clarify+Rewrite with roughly one-third as many questions, indicating that its gains are not merely due to asking more questions."

Report-level evaluation under two fixed downstream settings—OpenAI Deep Research (o4-mini-deep-research-2025-06-26, a commercial black-box DRA) and OAgents (an open-source agent framework using gpt-5-mini)—shows that G-STEER achieves the highest overall P-score within both downstream settings. With OpenAI Deep Research, G-STEER outperforms CEP-Clarify+Rewrite in personalization (P = 4.35 vs. 4.25) while attaining the second-highest Q-score (5.56), showing that its personalization gains do not substantially compromise report quality. With OAgents, it achieves the highest overall scores for both personalization (5.70) and content quality (5.72). Across unseen settings, G-STEER maintains strong personalization under user and task shift, while report quality remains competitive across downstream DRAs and unseen settings. The authors note that FP, recall, and precision metrics assess target-level framing coverage and selectivity but do not directly verify whether every expressed factor value is entailed by the acquired evidence, and that Using GPT-5.4 for both target construction and factor extraction may bias absolute scores, though all methods share the same evaluation pipeline.

The case study shows how factors specific to the user redirect the searches performed by OAgents. Before refinement, OAgents "searches broadly for guidance on creator platforms and general compliance requirements. After refinement, it focuses on Xiaohongshu carousel formats and policies governing medical and health content, while also seeking references relevant to the user's intended positioning in gentle urban wellness. These changes reflect the user's preferred format, content sensitivity, and brand direction. The added reference branch shows that framing factors shape both search specificity and downstream research dimensions."

The paper concludes: "We introduced G-STEER, a framework for PDR query refinement that routes evidence acquisition between user memory and direct clarification. It uses an Intent Elicitation Graph for dependency-consistent trajectory supervision and target-anchored optimization, improving factor acquisition and prioritizing important factors under interaction costs. Across commercial black-box and open agent settings, G-STEER achieves stronger report personalization with substantially fewer questions than the strongest clarification baseline. The stated limitation is that Our evaluation relies on simulated users and a simple retrieval-based memory store, motivating future work with human interaction and richer memory systems. Overall, effective personalized research requires identifying relevant context, grounding it appropriately, and knowing when to stop."

The paper's stated contributions are: (1) formulating PDR query refinement as three coupled decisions and represent clarification dependencies among user-specific framing factors with an IEG; (2) training "the refiner to choose Retri…

The following concrete improvements can be derived from the G-STEER framework:

1. Add explicit evidence-state tracking for every user-specific framing factor.

Maintain a slot-based state where each factor records its purpose, grounding status, supported value, and evidence source.

The improved AI system can decide, at every step, which user goals/constraints/preferences are already supported, which need retrieval from persistent memory, and which require direct user clarification.

2. Use a three-way grounding action: Retrieve, Ask, Stop.

Instead of always asking the user, route evidence acquisition through memory retrieval first, direct clarification only when memory lacks support, and stop when accumulated evidence is sufficient.

The improved system can reduce user burden dramatically—e.g., asking roughly one-third as many questions as a strong clarification baseline while achieving equal or better personalization.

3. Model clarification dependencies with an Intent Elicitation Graph (IEG).

Treat user-specific factors as nodes in a directed acyclic graph where resolving one factor can activate or change another factor. Enforce that no child factor is elicited before its parent.

The improved system can discover late-activating preferences that would be missed by independent question generation, and can avoid asking irrelevant or premature questions.

4. Train the policy on graph-scaffolded trajectories before reinforcement learning.

Supervise the policy with trajectories that vary target dependencies, memory hits, late activation, and stopping points, then optimize with trajectory-level GRPO.

The improved system learns the interaction protocol and valid clarification behavior from a small dataset, then improves its weighting toward high-importance factors without sacrificing interaction efficiency.

5. Reward target coverage but penalize false positives, acquisition cost, and behavioral violations.

Use a reward that maximizes weighted coverage of important targets, penalizes weakly aligned or unmatched covered slots, charges more for user questions than memory retrievals, and penalizes repeated acquisition attempts and multi-intent questions.

The improved system prioritizes essential personalization factors, avoids over-asking, and produces cleaner, more focused clarification questions.

6. Use a one-question-per-action policy and explicitly track repeated acquisition attempts.

Generate a single clarification question per Ask action and detect when the same factor is queried again.

The improved system minimizes user fatigue and avoids redundant interactions while maintaining high coverage of essential factors.

7. Only promote evidence to long-term memory when it is durable, specific, and reusable.

After each clarification trajectory, update the user profile with a one-sentence summary only for broadly reusable patterns, and store directly supported, user-specific facts as retrieval memories. Exclude generic advice, weak inferences, temporary details, and overly sensitive information.

The improved system becomes more personalized over time without accumulating noisy or irrelevant context, making future refinements faster and more accurate.

8. Perform inference without relying on offline elicitation targets or graphs.

At test time, induce initial evidence slots by prompting, maintain them as a textual set, and generate the refined query only from slot values supported by the available evidence.

The improved system can be applied to new users, unseen tasks, and black-box Deep Research Agents because it requires no access to internal pipelines and no hand-annotated target list at inference.

9. Include downstream report utility as a design objective, not just factor coverage.

Refine queries so that they change the downstream agent’s searches in user-specific ways—e.g., targeting specific platforms, content formats, sensitivities, and brand direction—while preserving report quality.

The improved system can steer a commercial black-box deep-research agent toward genuinely user-relevant evidence and still produce high-quality reports, as shown by high personalization scores with competitive content-quality scores.

10. Explicitly handle user and task shift.

Train with scenarios where the user, task, or both are unseen, and evaluate coverage under those shifts.

The improved system maintains strong personalization when encountering new users or new research tasks, rather than overfitting to memorized user profiles or task templates.