FutureX for Medical Research: Automating Literature Review and Hypothesis Generation from Weekly ArxivLens Summaries

Keeping up with the biomedical literature is a bottleneck. ArxivLens weekly roundups help, but even a curated list of 200 papers exceeds human reading capacity. FutureX, the agentic AI coding agent in FIM, can ingest those roundups, summarize them, and produce review drafts and grant proposals. This post shows how to turn a reactive email digest into a continuous agentic research workflow.

The ArxivLens Ingestion Layer#

ArxivLens delivers weekly summaries of arXiv preprints relevant to medicine—bioinformatics, medical imaging, NLP for clinical notes, and drug discovery. FutureX connects to these summaries via the ArxivLens export API or by polling an inbox. It normalizes each entry to a JSON schema: paper_id, title, authors, subjects, abstract, and the ArxivLens-generated digest. This schema is the contract between the external roundup and the agent's internal knowledge graph.

Parsing and Deduplication#

FutureX cross-references entries with previous weeks' state stored in a local SQLite database. It tracks arXiv IDs, flags updates such as v2 or v3, and de-duplicates papers that overlap with existing reviews. The agent uses simple heuristics first—string matching on title and Jaccard overlap on author lists—before falling back to embedding-based semantic similarity. This step ensures the downstream summaries only include genuinely new material.

Multi-Scale Summarization and Trend Detection#

FutureX does not produce one summary per paper. It builds a three-level hierarchy: paper-level summaries, topic-level briefs, and a weekly research synthesis. For each paper, the agent reads the abstract plus any extracted methods text, then writes a structured summary in Markdown: Motivation, Methods, Key Results, and Limitations. This structure mirrors what a clinician or researcher cares about and makes later composition easier. For example, a paper on diffusion-based MRI reconstruction yields a summary where the Key Results line states the quantitative gain over U-Net baselines, and the Limitations line notes the missing runtime benchmark.

Clustering for Emerging Trends#

To identify emerging trends, FutureX embeds the structured summaries using an off-the-shelf sentence transformer, then runs HDBSCAN clustering over the weekly corpus. Clusters that grow in size week over week are flagged as rising topics. The agent also tracks verbs and noun phrases—like diffusion model, foundation model finetuning, and conformal prediction—to catch terminology shifts before they become obvious. This is where agentic AI adds value: not just summarization, but synthesis. The cluster centroids, top terms, and representative paper IDs are persisted as a trend report in a trends/ folder alongside the raw summaries.

Hypothesis Generation as a Side Effect#

Trend detection becomes hypothesis generation when the agent connects clusters across weeks. If cluster A (vision-language models for chest X-rays) and cluster B (uncertainty quantification) both grow, FutureX drafts a hypothesis: Fine-grained uncertainty calibration improves the reliability of vision-language models for detecting incidental findings in chest X-rays. The agent grounds this hypothesis in citations from both clusters, creating an evidence trace that is inspectable by the user. Each hypothesis gets a score derived from cluster growth rate, citation count in the ArxivLens digest, and the novelty of the term overlap.

Drafting Review Papers and Grant Proposals#

The highest-leverage task is writing. FutureX uses the accumulated weekly state to draft narrative review sections. Given a scope such as deep learning in cardiology, the agent selects relevant clusters, orders them by prominence, and writes a background section with chronological citations. It inserts placeholder tables comparing study designs, outcomes, and gaps. A human researcher then edits, but the first-draft effort drops from days to minutes.

Grant Proposal Automation#

For grant proposals, FutureX repurposes the same evidence graph. The Specific Aims page is the hardest part to write. FutureX generates aims by looking at the gap structure identified during trend detection: each gap becomes a candidate aim. The agent then writes the Background and Significance section by summarizing the strongest evidence for that gap, and it populates the Innovation section with the trend data showing why the topic is timely. This is agentic AI for medical research automation at the level of the actual deliverable. FutureX also cross-checks the reference list against the ArxivLens digest to ensure every citation appears in the source corpus, reducing hallucination risk.

Version Control and Scheduling#

FutureX runs as a scheduled job—a weekly Python script, a GitHub Action, or an FIM cron task. Every run produces a new markdown file in a reviews/ repository. Changes are committed with conventional commit messages like feat: add Week 37 synthesis or fix: merge duplicate AlphaFold cluster. This gives you an auditable timeline of how the literature changed. When you need to draft a review, you can branch from a specific week state rather than starting from scratch. The agent also maintains a CHANGELOG.md that records which hypotheses were added, refined, or invalidated by newer papers.

Prompt Architecture and Guardrails#

FutureX uses a multi-stage prompt stack rather than a single monolithic instruction. Stage one extracts structured data from ArxivLens JSON. Stage two summarizes individual papers. Stage three clusters and identifies trends. Stage four composes documents. Each stage writes intermediate files, so you can inspect or override any step. Guardrails are implemented as schema validators: if a generated hypothesis does not reference at least two paper IDs from distinct weeks, the agent rejects it and re-runs with a narrower context. This modular design keeps the workflow trustworthy and debuggable.

A Concrete Walkthrough#

Let's tie this together with an example. Suppose the latest ArxivLens summary contains 40 papers on protein structure prediction and oncology. FutureX ingests them, clusters five papers around AlphaFold variants on mutated tumor proteins. It also finds three earlier weeks where this cluster was smaller. The agent creates a synthesis paragraph, then proposes a hypothesis: Conformational ensembles from AlphaFold models can predict driver mutation effects in EGFR more accurately than single-structure scoring. For a targeted grant, FutureX assembles a background section citing 12 papers, a specific aim to benchmark this approach, and a methods sketch using reference implementations from the cluster.

The researcher reviews, adjusts the claim, and exports to Overleaf. The total manual work is less than an hour; the alternative—manually reading 40 abstracts—takes a day. The same workflow scales to thousands of papers per week without extra human effort, which is why medical research automation is the most natural fit for agentic AI.

Conclusion#

ArxivLens gives you coverage; FutureX gives you comprehension. By combining structured ingestion, multi-scale science summarization, trend clustering, and generative drafting, medical researchers can shift from reactive reading to a proactive research program. The architecture is simple enough to run on a laptop: a cron job, a SQLite database, and a sequence of prompts inside FutureX. What changes is the speed at which a literature review becomes a testable hypothesis. For teams building on FIM, this workflow is a starting point for larger agentic systems—ones that not only summarize science but participate in the cycle of hypothesis and validation.