TrustResearcher: Automated Research Ideation

• TrustResearcher is a multi-agent system that automates literature-based research ideation through structured knowledge curation and transparent intermediate states.
• The system organizes research into four stages—curation, diversified idea generation, selection, and expert review—ensuring traceability and evidentiary alignment.
• It addresses information overload by balancing the generation of novel hypotheses with rigorous grounding and user-controlled audit trails.

TrustResearcher is a multi-agent demo system for automating literature-based research ideation in a way that is explicitly knowledge-grounded and transparent. It is designed around a recurring tension in automated scientific ideation: generating hypotheses that are diverse and genuinely novel without sacrificing alignment to evidence, provenance, or user control. Its pipeline is organized into four stages—Structured Knowledge Curation, Diversified Idea Generation, Multi-stage Idea Selection, and Expert Panel Review & Synthesis—and exposes intermediate reasoning states, execution logs, structured artifacts, and tunable agent parameters so that researchers can inspect, audit, and steer the process rather than treating ideation as a black box (Zhou et al., 20 Oct 2025).

1. Research problem and design orientation

TrustResearcher is motivated by the claim that research ideation is increasingly constrained by information overload and bounded human attention. Earlier LLM-based ideation pipelines are described as suffering from three recurring limitations: black-box behavior, weak grounding to literature, and limited transparency and control. In this framing, expansive prompting can increase exploration while weakening evidentiary anchoring, whereas stricter prompting can improve grounding at the cost of novelty; the system is presented as an attempt to manage this novelty–grounding trade-off without collapsing either side of it (Zhou et al., 20 Oct 2025).

The system’s central design commitment is that ideation should remain inspectable. Rather than presenting only final hypotheses, it persists intermediate artifacts such as topic decomposition results, knowledge-graph entities and relations, Graph-of-Thought paths, internal and external selection scores, and reviewer outputs. This design implies a conception of research assistance in which traceability is treated as part of the substantive quality of ideation, not merely as an interface convenience.

A second defining feature is domain-agnosticism. The pipeline is stated to be instantiable in any scientific field so long as literature sources exist. This does not mean that all domains are treated identically; rather, prompts, query formulations, and entity types are adaptable while the orchestration pattern remains stable.

2. Four-stage architecture

TrustResearcher orchestrates specialized agents across four stages that mirror a stylized research workflow: retrieving and structuring literature, generating diversified candidate ideas, selecting among them, and conducting a reviewer-style synthesis. The pipeline begins with a seed topic and target number of ideas, and each stage emits structured outputs that are reused downstream (Zhou et al., 20 Oct 2025).

Stage	Primary function	Main outputs
A. Structured Knowledge Curation	Retrieve, deduplicate, and structure literature into a lightweight KG	Curated paper set, KG entities/relations, structured context
B. Diversified Idea Generation	Explore KG-grounded reasoning paths and over-generate hypotheses	Candidate proposals with facets, reasons, scores, supporting references
C. Multi-stage Idea Selection	Score, merge, and externally filter ideas for distinctness and grounding	Pruned, consolidated, evidence-aligned ideas with similarity metadata
D. Expert Panel Review & Synthesis	Critique and aggregate ideas using reviewer-style agents	Reviewer/novelty JSON, meta-review, unified scores, ranked portfolio

At a systems level, this organization separates evidence acquisition from ideation, and ideation from adjudication. That separation is significant because it makes it possible to inspect where a candidate hypothesis entered the pipeline, which KG traces informed it, how redundancy was resolved, and why it survived later filtering. Compared with single-pass prompting, the architecture introduces explicit checkpoints in which both diversity and evidentiary alignment can be assessed.

3. Knowledge curation and graph-grounded ideation

Stage A uses the Semantic Scholar API as its primary source. An LLM first decomposes the seed topic into domain terms, methods, and application facets, then generates multiple query combinations ranging from keywords to higher-order tuples. Retrieval is asynchronous and rate-limited, and retrieved papers are merged after semantic deduplication. The default run retrieves 50 papers, scored by a relevance–citation mixture

$$s(\text{paper}) = 0.7 \cdot s_{\text{rel}} + 0.3 \cdot s_{\text{cite}}.$$

Metadata ingestion includes authors, keywords, methods, time, and venue, and the resulting KG uses typed relations such as uses, improves, compares, and based_on (Zhou et al., 20 Oct 2025).

The KG is expanded in four controlled phases: LLM-based entity extraction; mini-batch enrichment from metadata; degree-based expansion over the top-$k$ high-degree nodes with default $k=10$; and hybrid sampling with 60% high-degree and 40% random nodes. This controlled expansion is intended to preserve coherence while still surfacing latent links. Because downstream reasoning is constrained to KG nodes and edges, later ideation remains anchored to the curated literature rather than wandering into unconstrained associative generation.

Stage B then performs diversified idea generation. A planner agent analyzes the top-20 high-degree KG entities and summarizes the top-5 semantically relevant papers to extract research cues. Gap analysis identifies 3–5 open limitations, and each candidate idea is decomposed into three facets: Problem Statement, Proposed Methodology, and Experimental Validation. Graph-of-Thought exploration is implemented through asynchronous DFS with branching factor $b = 3$ and depth $d = 5$, using the path score

$$s(\text{path}) = 0.6 \cdot Q_{\text{nodes}}(\text{path}) + 0.2 \cdot D_{\text{edges}}(\text{path}) + 0.2 \cdot L(\text{path}).$$

Candidate generation is intentionally over-complete. The over-generation factor is $\alpha = 10$ times the requested number of ideas. Three strategies are run in parallel: base variants extending faceted plans, GoT variants reformulating high-scoring KG paths, and cross-pollination hybrids across top-ranked ideas via KG cross-connections. Redundancy is pruned using string and semantic matching, while iterative self-critique rounds refine methodological and validation detail. Every idea records its KG entities, planning traces, and top-3 supporting papers.

4. Selection, review, and transparency mechanisms

Stage C performs a two-level filtering process. Internal selection scores candidates using weighted criteria: novelty $0.30$, feasibility $0.25$, clarity $0.20$, and impact $k$0. Redundancy is handled by computing pairwise Jaccard similarity over idea facet sets,

$k$1

and merging pairs with Jaccard similarity greater than $k$2 until convergence. Over-generated ideas are reduced from $k$3 to $k$4 the requested count before external filtering (Zhou et al., 20 Oct 2025).

External selection compares candidates against retrieved literature using BGE-M3 embeddings over combined title, abstract, and methods fields, with cosine similarity

$k$5

Similarity thresholds are defined as low if below $k$6, mid for $k$7–$k$8, and high above $k$9. Candidates whose maximum similarity is below $k=10$0 are retained as sufficiently distinct, and the system logs top overlapping papers to preserve traceability.

Stage D introduces two asynchronous reviewer agents: a Reviewer agent, scoring feasibility, technical soundness, clarity, impact, and implementation complexity, and a Novelty agent, scoring originality and distinctiveness relative to prior work across technical, problem, application, theoretical, and empirical dimensions. Both operate with rubric-driven 1–5 scoring and return valid JSON containing strengths, weaknesses, and recommended revisions. A unified score is then formed by weighted aggregation, with stage defaults giving reviewer criteria weights $k=10$1. Ideas with unified scores above $k=10$2 are classified as high-quality, approximately corresponding to “weak accept”; if more ideas exceed this threshold than the target portfolio size, they are retained rather than discarded.

Transparency is not confined to the final review stage. The system exposes topic decomposition lists, retrieval batches, KG entities and relations, GoT paths and scores, ideation drafts and revisions, internal and external selection logs, reviewer and novelty JSON, and the final portfolio. Execution logs are streamed to a status monitor, and all artifacts are saved as structured JSON. Hyperparameters such as retrieval batch size, KG expansion settings, GoT depth and branching, over-generation factor, merge thresholds, similarity thresholds, and reviewer weights are surfaced for user control. Optional human-in-the-loop supervision is available at planning, refinement, selection, and review stages.

5. k-truss breaking case study

The paper demonstrates TrustResearcher on the graph-mining problem of k-truss breaking. Let $k=10$3 be a simple undirected graph, and let $k=10$4 denote the number of triangles in a subgraph $k=10$5 that contain edge $k=10$6. The $k=10$7-truss $k=10$8 is the maximal subgraph $k=10$9 such that every edge in $b = 3$0 participates in at least $b = 3$1 triangles within $b = 3$2. The k-truss breaking problem is formulated as finding the smallest edge set $b = 3$3 such that removing $b = 3$4 eliminates all $b = 3$5-truss subgraphs:

$b = 3$6

The problem is described as combinatorially complex with cascading effects, but the paper does not assert a formal complexity result such as NP-hardness (Zhou et al., 20 Oct 2025).

In this demonstration, Stage A retrieves work on k-truss decomposition, scalable algorithms, parallel index construction, and dynamic maintenance, then encodes these materials in the KG. Stage B produces several distinct directions, including scalable localized algorithms that bypass global recomputation of trussness, epidemic-containment strategies on temporal contact networks, and learning-based edge-importance prediction for deletion prioritization.

A representative hypothesis, detailed in the appendix, is a “Scalable Approximation Framework for Real-Time k-Truss Disruption in Massive Dynamic Graphs.” Its proposed components include incremental k-truss maintenance via sketch-based triangle estimation, hierarchical graph partitioning, locality-sensitive edge prioritization, adaptive deletion budgeting, distributed integration through Apache Flink or Spark Streaming, and approximation bounds under probabilistic graph models such as Chung–Lu and SBM. The proposed experimental setup includes R-MAT, Barabási–Albert, and SBM graphs up to $b = 3$7 edges, streaming Twitter and citation-network experiments, robustness and adversarial noise tests, ablations isolating incremental versus partitioning components, and distributed stress tests.

The reviewer stage scored the scalable algorithm idea at $b = 3$8 for novelty and $b = 3$9 for clarity, while noting weaknesses related to extreme-scale scalability. Across the portfolio, five high-quality candidates were retained with an average overall score of approximately $d = 5$0. The paper explicitly presents these outputs as hypotheses and critiques rather than executed experiments.

6. Trust, limitations, and broader significance

TrustResearcher’s notion of trust is procedural rather than anthropomorphic. The system attempts to make ideation trustworthy by grounding hypotheses in a curated literature set and KG, attaching supporting papers and reasoning traces to each proposal, exposing intermediate states as JSON, and preserving why ideas were retained, merged, or filtered. The paper also identifies risks that remain: hallucinations, uneven domain coverage, prompt sensitivity, and overconfident novelty claims. Its safeguards are correspondingly procedural—literature grounding, external similarity checks, reviewer-style critique, and exposed logs—rather than guarantees of correctness (Zhou et al., 20 Oct 2025).

The system’s limitations are explicit. It is a demo ideation system and does not execute experiments. Formal guarantees of correctness are not part of its design. Evaluation centers on transparency and portfolio quality rather than benchmark comparison against other ideation systems. Scalability and cost are nontrivial: the case study reportedly used approximately 200K tokens in a 15-minute session, and very large KGs or deep GoT sampling require careful tuning. Grounding quality may degrade in domains with sparse literature or limited API access.

These limitations align with broader findings on AI use in research. In think-aloud observations of early-stage scholarly workflows, researchers described trust in AI as fragile and context-dependent, with provenance opacity and confident tone creating accountability burdens that users compensate for through manual verification and selective reliance (Gautam et al., 25 Apr 2026). Interviews with novice researchers similarly identified idea leakage, memorization skepticism, and distrust of unverifiable deletion or “no training” controls as central obstacles to trustworthy LLM-assisted research workflows, motivating recommendations such as institution-level sandboxed isolation and verifiable data-deletion audits (Zhang et al., 2 Jun 2026). At the level of agent evaluation, AARRI-Bench argues that trust in researcher-like agents depends on professionalism, thoroughness, nuanced reasoning, field sensitivity, research ethics, and scientific judgment; its best-performing configuration reaches only $d = 5$1 success on tasks designed to be easy for human researchers but error-prone for agents, underscoring the continuing importance of transparent, auditable workflows rather than opaque autonomy (Wang et al., 5 Jun 2026).

In that wider landscape, TrustResearcher occupies a specific position. It is neither a benchmark for researcher-like behavior nor an autonomous experiment-executing agent. It is a structured ideation environment whose main contribution lies in making literature-based hypothesis generation inspectable, parameterized, and evidence-linked. Its significance therefore lies less in replacing scholarly judgment than in reorganizing ideation so that judgment can be exercised with clearer access to provenance, critique, and intermediate state.