Agentic Research Ideation System (ARIS)

• ARIS is a multi-layered platform that supports autonomous scientific ideation with iterative agent loops and memory-augmented reasoning.
• It integrates specialized agent roles, strict policy enforcement, and transparent execution to ensure reproducibility and accountability.
• Its architecture features layered design, multi-agent coordination, and diverse memory models for scalable and verifiable research ideation.

An Agentic Research Ideation System (ARIS) is a multi-layered architecture designed to autonomously support and accelerate scientific ideation through iterative agent loops, memory-augmented reasoning, typed and governed tool interfaces, and multi-agent coordination. ARIS separates cognitive reasoning from execution, formalizes memory retrieval and update policies, incorporates entropic and novelty-diversity metrics directly into ideation workflows, and enforces enterprise-grade governance to ensure transparency, traceability, and reproducibility. As a composable platform, ARIS integrates human-in-the-loop controls, rigorous evaluation metrics, and explicit design rules that enable the automated exploration, implementation, and validation of research ideas while preserving intellectual ownership and providing fine-grained transparency over agentic reasoning (Alenezi, 11 Feb 2026, Audran-Reiss et al., 19 Nov 2025, B et al., 1 Jan 2026, Kim et al., 14 Apr 2026, Cheng et al., 26 Jan 2026, Zhou et al., 20 Oct 2025, Liu et al., 17 Jan 2026, Zimmer et al., 16 Mar 2026).

1. Layered Reference Architecture

ARIS is structured in a five-layer architecture, with cross-cutting concerns for governance and observability:

• Layer 0: Human Interface (UI/Chat/API)
• Layer 1: Perception & Context Builder
• Layer 2: Planning (Cognitive Kernel)
• Layer 3: Execution Engine (Typed Tools)
• Layer 4: Memory Stores (Short-Term & Long-Term)

The agent loop proceeds as follows: user input is processed into structured observations and context, a plan step (thought or tool call) is proposed, the execution engine dispatches validated tool invocations, and multi-tiered memory is updated on every step. The state at time $t$ is $s_t = (b_t, g, h)$ where $b_t$ is a belief vector, $g$ is the user goal, and $h$ is the episode history. The observation, context, planning, execution, belief update, and memory write functions are formally defined by the following equations:

$o_t = \mathit{Perceive}(r_{t-1}, s_{t-1})$

$$c_t = \mathit{BuildContext}(s_{t-1}, o_t, P_{\mathrm{policies}})$$

$$p_t = \mathit{PlanStep}(\mathrm{LLM}, c_t), \quad p_t \in \{\texttt{Thought}, \texttt{ToolCall}\}$$

$$z_t = \begin{cases} \mathit{ExecuteTool}(p_t), & \text{if } p_t\text{ is ToolCall} \ \text{null}, & \text{else} \end{cases}$$

$b_t = \mathit{UpdateBelief}(b_{t-1}, o_t, z_t)$

$s_t = (b_t, g, h)$0

Typed tool interfaces are enforced using schema-validated registries (e.g., SearchTool, SummarizationTool, NotebookTool), ensuring that the planning layer emits only serializable PlanStep objects, never invoking tools directly. Execution is strictly mediated by RBAC-governed, sandboxed microservices, supporting robust policy enforcement and auditability (Alenezi, 11 Feb 2026).

2. Memory-Augmented and Diversity-Driven Ideation

ARIS implements a multi-tiered memory model involving:

• Short-Term Memory: in-RAM context, LRU-evicted, capacity $s_t = (b_t, g, h)$1
• Episodic Memory: chronological events $s_t = (b_t, g, h)$2
• Semantic Memory: persistent vector store $s_t = (b_t, g, h)$3
• Profile Memory: user/system preferences $s_t = (b_t, g, h)$4

Retrieval is governed by embedding cosine similarity, with regular updates for novel content and periodic compaction by clustering. Ideation diversity is formally defined via Shannon entropy over the distribution of high-level model architectures $s_t = (b_t, g, h)$5 and by tracking distinct counts over multiple candidate drafts $s_t = (b_t, g, h)$6. Diversity and novelty promotion mechanisms include:

• Entropy-based auxiliary rewards in system prompts.
• Structured sibling memory to prevent mode collapse.
• Explicit diversity clauses and prompt-adaptive complexity cues.
• Online metrics $s_t = (b_t, g, h)$7, $s_t = (b_t, g, h)$8 regularly logged per task and trajectory.

Experiments show that higher ideation diversity yields superior agent performance (e.g., for AIRA$s_t = (b_t, g, h)$9: baseline median rate 52.8%, ablated/low-diversity 44.4%; effect size $b_t$0), with causal linkages demonstrated via controlled ablation (Audran-Reiss et al., 19 Nov 2025).

3. Multi-Agent Coordination and Distributed Workflows

ARIS enables instantiation of teams of specialized agents using standard topologies:

• Hub-and-Spoke (orchestrator with workers)
• Ring (token-pass consensus)
• Fully-Connected (gossip protocol)

Each agent may specialize by function (problem framing, literature harvesting, hypothesis generation, statistical evaluation, etc.) as in the MIDAS framework’s ideator, evaluator, and curator roles (B et al., 1 Jan 2026). Communication is orchestrated via persistent “Vaults” (Problem, Idea, Literature, Concept) and all intermediate artifacts are serialized for transparent handoff between agents. Ring and gossip models employ consensus protocols (e.g., Raft), leader election, heartbeat monitoring, and rate limiting to mitigate class-specific failures.

Task allocation and evidence flow for claims or hypothesis breakpoints are handled by planner, librarian, reasoner, and producer agent chains, as exemplified by the Research IDE design ($b_t$1 comprising $b_t$2) (Cheng et al., 26 Jan 2026).

4. Governance, Transparency, and Reproducibility

ARIS is hardened for enterprise and academic deployment:

• Full RBAC on tools, memory, and data.
• Policy-as-code (e.g., Rego/OPA), schema validation at ingress, and contract versioning.
• All actions, tool invocations, and outcomes are logged in immutable, append-only audit trails (JSON schema includes timestamp, agent, planstep, tool versions, cost accounting, policy decisions).
• Reproducibility requirements: LLM model pinning, tool contract versions, semantic memory snapshots, container image digests, and full metric tracking (step latency, token usage, memory cache statistics).
• Proof-carrying PlanSteps include provenance pointers and confidence scores.
• Timeouts, circuit-breakers, and strict cost budgets per request are policing gateways.

OpenAPI and Protobuf protocols govern typed contracts for all tool endpoints, supporting verifiability and safe composition (Alenezi, 11 Feb 2026). Human-in-the-loop workflows can be tightly coupled, with explicit approval, revision, and iterative feedback on generated artifacts (e.g., IRB protocols, cohort definitions, ML model reports) (Kim et al., 14 Apr 2026).

5. Human Agency, Ownership, and Reflexive Control

Three agentic roles—Ideator, Writer, Evaluator—are modularized with a retrieval-augmented LLM pipeline and exposed via graded user-control spectra (Low, Medium, Intensive). Distinct feature availability matrices govern interaction at each level: from full automation (Low; keywords and selection only) to maximum reflexivity (Intensive; full prompt/edit/search, revision tracking, and evaluation customization) (Liu et al., 17 Jan 2026). Experimentation shows that creativity support does not increase monotonically with control level, and participant effort shifts from ideation to verification as more supervision is allowed. Ownership attribution (Human, AI, or Co-Created) is tracked as a function of contribution and decision effort.

Critical interaction design principles include explicit revision tracking, visible stepwise agent logs, and hooks for meta-cognitive reflection, e.g., prompts such as “Why is this idea novel?” All evidence and chain-of-thought reasoning is surfaced for user inspection, never auto-rewriting the researcher’s prose (Liu et al., 17 Jan 2026, Cheng et al., 26 Jan 2026).

6. Domain Adaptation and Application Examples

ARIS is domain-agnostic, with minimal adaptations required for new fields:

• Biomedical applications: specialized entity-relation templates, PubMed/crossref adapters, revised knowledge graph schemas (Zhou et al., 20 Oct 2025).
• Clinical research: privacy-preserving, coding-free interaction using MCP, strict separation between data locations and orchestration, federated result synthesis, and compliance with evidence-based frameworks like TRIPOD+AI (Kim et al., 14 Apr 2026).
• Mathematics and ML: autonomous CLI coding agents within sandboxed containers, orchestrated via commandment-rich prompt files, and automated report, experiment, and code management at scale; all actions subject to symbolic verification and rigorous step logging (Zimmer et al., 16 Mar 2026).

Case studies demonstrate ARIS can efficiently execute literature-curated ideation, experiment design, implementation, evaluation, and meta-analysis across scientific disciplines while ensuring every step is logged, versioned, and verifiable.

7. Methodological Guardrails and Evaluation Metrics

At the core of ARIS are explicit, falsifiable operational rules—prompted as “commandments” or formalized as assessment rubrics in multi-stage agent chains:

• Never fabricate citations; each must be verified with title, author, venue, and DOI (Zimmer et al., 16 Mar 2026).
• One variable per experiment, with tiered evaluation and mandatory debugging before moving on.
• All claims must be verified via dedicated scripts, with results recorded as verified/partially/unverified.
• Agentic outputs scored by unified metrics, e.g., $b_t$3 and semantic novelty via $b_t$4 for literature alignment (B et al., 1 Jan 2026).
• Ideation diversity tracked online ($b_t$5, $b_t$6), and performance measured via normalized scores, percentile ranks, and gold/bronze/silver medal rates (MLE-bench).
• Logs, revision histories, internal agent decisions, and external groundings are uniformly exportable for audit and further analysis.

The ARIS blueprint specifies best practices that combine cognitive-execution decoupling, multi-agent collaboration, memory-augmented reasoning, diversity facilitation, contract governance, explicit methodology, and reflexive human engagement—realizing a composable, scalable, and auditable agentic research ideation platform (Alenezi, 11 Feb 2026, Audran-Reiss et al., 19 Nov 2025, B et al., 1 Jan 2026, Kim et al., 14 Apr 2026, Cheng et al., 26 Jan 2026, Zhou et al., 20 Oct 2025, Liu et al., 17 Jan 2026, Zimmer et al., 16 Mar 2026).