InterDeepResearch: Interactive Deep Research

• InterDeepResearch is an interactive deep research paradigm that enables seamless human–agent collaboration and dynamic navigation of large-scale, heterogeneous information sources.
• It employs hierarchical context management and dynamic memory reduction to mitigate LLM context exhaustion and ensure efficient processing.
• The system features coordinated multi-view interfaces—including chat, graph, and card styles—that provide process transparency, real-time steering, and provenance tracking.

InterDeepResearch is an interactive deep research paradigm and system architecture focused on enabling human–agent collaborative information seeking and extended reasoning over large-scale, heterogeneous knowledge sources. Unlike conventional autonomous “query-to-report” agentic systems, InterDeepResearch explicitly integrates human oversight, dynamic context management, and coordinated visual workflows to support process observability, real-time steerability, and efficient navigation across complex research trajectories. It combines hierarchical context modeling, advanced interaction mechanisms, and Markovian memory curation to mitigate LLM context exhaustion and provenance ambiguity, yielding state-of-the-art research accuracy and collaborative usability.

1. Research Context Architecture and Memory Management

At the core of InterDeepResearch is a dedicated research context management framework, organized hierarchically across three levels (information, actions, and sessions) (Pan et al., 13 Mar 2026):

• Level 1: Research Information Atomic knowledge units ($i_{t+1}$) are accumulated at each step into the information set:

$$\mathcal{I}_{t+1} = \mathcal{I}_t \cup \{i_{t+1}\},$$

where $i_{t+1} \in \{$User Info, Search Result, Source Content, Processed Summary$\}$.

• Level 2: Research Actions Each action $a_t$ is generated by a user or an agent:

$$a_{t+1} = f(\mathcal{A}_t, \mathcal{I}_t), \quad f \in \{\text{Agent}, \text{User}\}$$

Supported action types include User Information actions, Search, Scrape/Source, Processed Info, and Administrative actions.

• Level 3: Research Sessions Sessions are defined as contiguous subsequences of actions bounded by milestone actions ($a_i, a_j \in \mathcal{A}_{\text{milestone}}$), such as User Information or Administrative marks.

Dynamic Context Reduction: To prevent context blowup and “context rot,” the system prunes obsolete or non-milestone information after Processed Information actions and session switches. Specifically, prior Search/Source blocks are replaced with pointers upon summary, and old session data is deleted unless marked as a milestone. This ensures tractable LLM memory usage even in long-horizon research (Pan et al., 13 Mar 2026).

Cross-Action Backtracing and Evidence Provenance: Evidence for any claim or summary is recursively traced upstream in the dependency DAG $G = (V, E)$. Specialized backtrace agents verify the provenance of any statement by walking through ancestral actions and extracting supporting content, terminating at Source Information nodes. This is rendered as a dynamic overlay on the action dependency graph for transparency and validation.

2. System Workflow, Visualization, and Interaction

InterDeepResearch exposes three coordinated interfaces—Chat-Style Flow, Graph-Based Dependency, and Card-Style Information—enabling multiscale sensemaking and process steerability (Pan et al., 13 Mar 2026):

• Chat-Style View: Linear transcript with annotated sessions, live action summaries, and immediate interruption capability (via “Stop” or user message).
• Graph-Based View: Action Dependency Graph organized by both temporal and data dependencies. Expandable/collapsible nodes allow multi-level focus; auto-centering ensures context alignment.
• Card-Style View: Detailed, scrollable cards for all information units (user input, web search, scraped source, synthesized note). Inline citation referencing and “Trace” buttons invoke provenance mechanisms.

Interaction Mechanisms: Clicking any element highlights and links to all its manifestations across views. Hovering over citation superscripts activates provenance highlighting upstream in the dependency graph. Cross-view linkage maintains context coherence during high-branching exploration.

3. Agent Pipeline and Algorithmic Process

The research process follows an iterative Retrieval–Filtering–Synthesis pipeline:

1. PLAN: LLM agent consults current context $(\mathcal{I}_t, \mathcal{A}_t)$ and determines next action type.
2. RETRIEVE: Issues web search queries and collects top-K documents.
3. FILTER: Ranks documents by cosine similarity in embedding space; filters by threshold $\tau$.
4. SCRAPE: Downloads full text from selected URLs and segments into passages.
5. SYNTHESIZE: Generates concise evidence-grounded notes from passages and context via LLM; citations inserted as superscript references.
6. ACTION: Appends synthesized note as a Processed Information action.

• Summarization is analyzed via negative log-likelihood loss:

$\mathcal{L}_s = -\log P(N\,|\,C,P)$

where $N$ is the synthesized note, $C$ the context, and $P$ the source passages.

• Document relevance is scored:

$$r(d) = \cos(f_{\text{query}}(q), f_{\text{doc}}(d))$$

The loop continues until an explicit “Finish” or is preempted by user intervention.

4. Human–Agent Collaboration and Interactivity

Critical to InterDeepResearch is the elevation of the human operator from passive prompter to active cognitive overseer (Ye et al., 21 Jul 2025, Pan et al., 13 Mar 2026). Users steer agent reasoning at multiple control points:

• Query Refinement: Users modulate queries, inject domain-specific terminology, or reweight search priorities before execution.
• Interactive Browsing: URLs can be selected/deselected; requests for deeper web scrapes or alternative passage extraction are actionable mid-loop.
• Live Drafting and Report Editing: During synthesis, users can pause, restructure, request alternative analyses, or provide file-based expert notes.
• Audit and Correction: All system-generated rubrics, reasoning traces, and evaluation loops are transparent and editable; user feedback is automatically integrated as implicit preference signals observed from edits, hesitations, or navigation style.

This oversight prevents error cascades, enables live question pivots, and maintains situational awareness throughout the research trajectory.

5. Evaluation Methodology and Empirical Findings

Quantitative Experiments

• Benchmarks: Xbench-DeepSearch-v1 and Seal-0 (end-to-end, unattended execution).
• Metrics: Precision@K, Recall@K, and F1 against annotated answer sets.

System	Xbench-DeepSearch F1	Seal-0 F1
InterDeepResearch	78.5%	82.1%
Perplexity DeepRes.	75.3%	79.7%
Gemini DeepRes.	76.7%	80.4%

InterDeepResearch matches or surpasses state-of-the-art systems by leveraging full-cycle context management and collaborative workflows (Pan et al., 13 Mar 2026).

User Study Findings

• Participants: $N=15$ (developers, professionals, daily users)
• Key metrics (5-point Likert, mean):
  • Dependency Graph view: μ=4.70
  • Chat-style flow: μ=4.20
  • Card-style view: μ=4.13
  • Cross-view linkage: μ=4.07
  • Backtrace mechanism: μ=4.33
  • Process comprehension: μ=4.80
  • Steering flexibility: μ=4.40
  • Usability/satisfaction: μ=4.53–4.60
  • Report quality: μ=4.47
• Qualitative insights: Transparent agent action and live, provenance-based evidence tracing were considered vital for trust and control. Users adopted diverse navigation strategies enabled by the coordinated multi-view interface.

6. Broader Implications and Future Directions

InterDeepResearch constitutes a shift from monolithic query–output agentic paradigms to a collaborative, oversight-driven research cycle in which intelligence emerges from continuous human–AI interplay (Ye et al., 21 Jul 2025). Its key contributions include:

• Process Transparency: Enables inspectable, stepwise progress rather than black-box execution, mitigating error cascades and supporting real-time debugging.
• Dynamic Question Refinement: Supports flexible goal reprioritization and spontaneous direction changes, preserving an evolving cognitive context.
• Provenance and Verification: Systematic backtracing ensures claim grounding, supporting rigorous scientific workflows and reproducibility.
• Personalization: Implicitly adapts to diverse user behaviors and preferences, suggesting avenues for future adaptive navigation/interaction patterns.

Future enhancements highlighted by user feedback and system analysis involve more granular steering controls, multimodal interaction (voice, drag-and-drop, charting), and extension to team-based, multi-user research threads. A plausible implication is that such architectures will become foundational for transparent, trustworthy AI-augmented research in science, business, and education.

7. Comparison and Relation to Broader Deep Research Trends

InterDeepResearch synthesizes recent trends in agentic deep research (Ye et al., 21 Jul 2025), Markovian state reconstruction (Chen et al., 10 Nov 2025), pipeline management (Zhang et al., 18 Aug 2025), and systematic benchmarking (Li et al., 13 Jan 2026, Pan et al., 13 Mar 2026). Its approach to hierarchical context, interactive workflows, and provenance aligns with emerging best practices in evaluating long-horizon, reasoning-intensive AI agents and addresses fundamental challenges such as context drift, memory saturation, and user alignment. This convergence suggests that interactive, context-pruned, provenance-anchored design principles are likely to define the next generation of human–AI deep research platforms.