InfiHelper: Autonomous Research Assistant

• InfiHelper is an AI research assistant that autonomously generates publishable papers using a multi-layered DAG of specialized agents.
• It employs agent-as-tool decomposition, parallelism, and intelligent routing to optimize research workflow efficiency.
• Dual-audit mechanisms ensure high research quality, with empirical results showing significant improvements over baseline systems.

InfiHelper is an AI research assistant system built atop the InfiAgent framework, exemplifying automated end-to-end scientific research using a hierarchical, pyramid-like Directed Acyclic Graph (DAG) of specialized agents. Designed to autonomously generate publishable scientific manuscripts from high-level goals, InfiHelper leverages formalized agent decomposition, dual-audit mechanisms, agent routing, and self-evolutionary restructuring to optimize both quality and adaptability of research workflows (Yu et al., 26 Sep 2025).

1. Hierarchical DAG Architecture and Workflow Orchestration

InfiHelper occupies the top node (level 3) of a deep, multi-layered DAG structured for scientific workflow automation. Given an open-ended target such as "write a publishable paper on topic X," InfiHelper decomposes the research process into distinct phases: literature review, idea generation, experimentation, and paper drafting. Each phase is routed to specialized sub-agents according to the hierarchy:

• Top-level orchestration: Agents such as idea_agent oversee coarse-grained orchestration—selecting, refining, and integrating candidate research directions.
• Intermediate planners: Agents like project_planner_agent and code_builder_agent handle detailed planning, code synthesis, and task sequencing.
• Atomic tool agents: At the base, agents execute atomic actions (e.g., file I/O, running experimental scripts).

A Router module is employed to efficiently assign specific tasks to suitable agents, circumventing exhaustive layer-by-layer search. This pyramid configuration achieves both breadth ($b \approx 3\text{--}5$) and depth ($L=4$), supporting hundreds of concurrent agents, while maintaining a maximum child set per planner ($K_{\max}=5$) to bound local reasoning complexity (Yu et al., 26 Sep 2025).

2. Agent-as-Tool Decomposition and Parallelism

InfiHelper adopts the "agent-as-a-tool" paradigm, underpinning its decomposition formalism:

• Any agent at level $l$ decomposes its assigned task $T^{(l)}$ into subtasks $\{T^{(l+1)}_1,\ldots,T^{(l+1)}_{k_l}\}$, constrained so that

$$\bigcup_{l=0}^L \bigcup_j T_j^{(l)} = T_0; \quad k_l \leq K_{\max} (K_{\max}=5)$$

• The idea_agent, for instance, transforms "Generate research ideas for domain D" into subtasks like "Analyze reference method $i$" for each reference, passing these to idea_generate_agents that may invoke brainstorming or planners below them.
• At the lowest stratum, atomic agents such as code_builder_agent and experimental_execution_agent are treated as callable tools for code generation and experiment execution.

This modular approach enables fine-grained parallelism and re-use across research stages. Specialized functional agents execute independently where query decomposition permits, substantially increasing throughput and flexibility (Yu et al., 26 Sep 2025).

3. Dual-Audit Quality Assurance

To prevent error propagation and ensure research integrity, InfiHelper implements a two-tiered auditing system:

• Execution-Level (Self-Audit): Each agent $A_i$ maintains a quality score:

$$Q_i^{(t+1)} = \alpha Q_i^{(t)} + (1-\alpha) \mathrm{validate}(O_i^{(t)})$$

where $L=4$0 is a smoothing parameter and $L=4$1 indicates local output validity. Agents iterate until $L=4$2 (e.g., $L=4$3) before reporting completion.

• System-Level (Peer-Audit): A Judge Agent at level –1 assesses both self and peer agents’ output artifacts, verifying file existence, code execution, JSON schema adherence, and issues a pass/fail verdict. Cross-examination between peer agents is also performed to further minimize undetected errors.

This dual-audit design amplifies both local accountability and global coherence across the DAG, reducing the likelihood of error cascades in complex research pipelines (Yu et al., 26 Sep 2025).

4. Intelligent Routing and Self-Evolution

Task-agent mapping in InfiHelper is governed by the Router, which selects the highest affinity agent for each subtask by maximizing:

$L=4$4

where $L=4$5 contains metadata (e.g., task type, input size, domain keywords) and $L=4$6 encodes the agent's competencies.

Triggers for self-evolution include repeated Judge Agent failure flags or persistently low $L=4$7 scores ($L=4$8). In response, the DAG undergoes:

• Model-Level Merging/training: Successful model updates are merged and retrained.
• Topology-Level Pruning/Merging: Sub-optimal branches (with average $L=4$9) are pruned; redundant functionalities are merged to higher nodes.
• Agent-Level Expansion: Novel agents are instantiated to address emergent sub-tasks.

This suggests the agent system can adapt dynamically as the underlying scientific domain or research needs evolve (Yu et al., 26 Sep 2025).

5. Evaluation Metrics and Empirical Performance

InfiHelper’s efficacy was established along two primary axes:

Metric	InfiHelper	Baseline (Best)	Relative Improvement
Retrieval success (Lit. Rev)	0.92	–	–
Peer-review avg. score	6.0	4.0 (Sakana-AI)	50%
		4.75 (AI-Researcher)	26.3%

• Literature Retrieval: On 10+ engineering databases (1,000 queries), InfiHelper achieved a success rate of 0.92.
• Peer-Review Quality: Papers authored by InfiHelper received average scores of 6.0 from human reviewers, outperforming Sakana-AI (score 4.0, +50%) and AI-Researcher (score 4.75, +26.3%). Sample papers generated by InfiHelper were recognized at top-tier IEEE conferences (Yu et al., 26 Sep 2025).

6. Limitations and Prospective Extensions

Human oversight remains necessary to refine the initial DAG structures, indicating full autonomy in research generation is not yet realized. Agent orchestration introduces computational overhead, particularly in narrowly scoped tasks such as theorem proving, leading to reduced efficiency in certain domains (notably on MATH benchmarks). Reliance on a single Judge Agent for system-level audits centralizes judgment and may not preclude audit bias.

Proposed future improvements include:

• Integrating domain-expert LLMs at lower DAG levels to boost task-specific accuracy,
• Adopting decentralized audit ledgers to enhance transparency and reproducibility,
• Enabling continuous, user-driven DAG pruning and regrowth for adaptive, user-centric automation (Yu et al., 26 Sep 2025).

7. Significance in Autonomous Research Ecosystems

InfiHelper operationalizes the InfiAgent principles of task decomposition, agent-as-tool abstraction, dual-audit assurance, intelligent routing, and adaptive self-evolution to support near-autonomous scientific manuscript preparation. Its demonstrated ability to produce peer-recognized papers with minimal human input positions it as a referential architecture for next-generation, self-organizing autonomous research assistants, with extensibility to broader scientific and engineering domains (Yu et al., 26 Sep 2025).