AI for Literature Review: 6 Jobs AI Agents Do for Researchers

Most of a literature review is finding and reading: searching several databases, screening what comes back, pulling the same fields out of every paper, checking that each citation says what you claim, and catching what was published since you last looked. That is where AI for literature review helps most. AI agents for literature review can do much of that work, provided they search the papers, preprints, registries and grant records themselves and return every finding with its source.

This guide follows six literature review jobs, each for a different researcher. We ran every search on November 9, 2026 with Valyu's arXiv, PubMed, medRxiv, ClinicalTrials.gov and NIH RePORTER datasets and its Contents API, then checked each finding against the paper, PubMed, the registry or NIH RePORTER.

If you're an agent: Set up Valyu and make your first search.

If you're a human: Create an account and get an API key.

MCP onboarding: Connect your agent to https://mcp.valyu.ai/mcp and follow the MCP setup guide for authentication.

What AI agents for literature review do

An AI agent for literature review is a language model that can search papers, preprints, trial registries and grant databases, read what it finds and return cited, structured answers. It speeds up the searching, screening and extraction in a review. It doesn't replace a registered protocol, a librarian-designed search or two independent reviewers in a formal systematic review.

Figure 1. Six literature review jobs we ran on November 9, 2026, with the sources used, what came back and what to do next.

1. Scoping search: map a fast-moving field

The problem: a PhD student in computer science is writing the related-work chapter on evaluating hallucination in retrieval-augmented generation (RAG) and needs the recent papers.

What we ran: one search of arXiv, limited to papers from 2025 onwards.

from valyu import Valyu
 
valyu = Valyu()  # reads VALYU_API_KEY from the environment
 
response = valyu.search(
    "evaluating hallucination and faithfulness in retrieval-augmented generation",
    included_sources=["valyu/valyu-arxiv"],
    start_date="2025-01-01",
    max_num_results=10,
)
papers = {result.url: result.title for result in response.results}  # one entry per paper
for url, title in papers.items():
    print(title, url)

What came back: 10 passages from 7 papers published between March and November 2025, including a faithfulness leaderboard for RAG (arXiv 2505.04847), hallucination detection with semantic entropy (2505.07528), an evaluation of hallucinations in agent memory systems (2511.03506) and an agentic evaluation framework for domain-specific RAG (2511.04502).

What we checked at the source: the arXiv records for the papers. Titles and dates match.

The catch: search returns passages, not papers, so the same paper can appear more than once. Group results by URL before you count or cite them.

The decision: use the seven papers as seeds, and follow their references and citing papers to fill the chapter.

2. Screening and data extraction: one table, every trial

The problem: a systematic review team assessing GLP-1 receptor agonists for alcohol use disorder needs the randomised trials and the same fields from each one.

What we ran: a PubMed search for randomised trials, then the Contents API on three trial papers with a JSON schema, so every paper returns the same fields.

schema = {
    "type": "object",
    "properties": {
        "study_design": {"type": "string"},
        "participants_randomized": {"type": "integer"},
        "population": {"type": "string"},
        "intervention_and_dose": {"type": "string"},
        "duration_weeks": {"type": "number"},
        "primary_outcome": {"type": "string"},
        "primary_outcome_result": {"type": "string"},
        "supporting_quote": {"type": "string"},
    },
    "required": ["study_design", "participants_randomized", "primary_outcome", "primary_outcome_result", "supporting_quote"],
}
 
data = valyu.contents(
    urls=[
        "https://pubmed.ncbi.nlm.nih.gov/PMC9675448",
        "https://pubmed.ncbi.nlm.nih.gov/PMC11822619",
        "https://pubmed.ncbi.nlm.nih.gov/PMC10721313",
    ],
    summary=schema,
    response_length="medium",
)
for result in data.results:
    print(result.url, result.summary)

What came back: one row per paper, each with a verbatim supporting quote:

What we checked at the source: the full text of each paper. The numbers and quotes match.

The catch: the dulaglutide row is right on the numbers and wrong on the framing. The paper is a predefined secondary analysis of a smoking-cessation trial. Alcohol intake is the primary outcome of that analysis, not of the trial, and only 151 of the 255 participants drank alcohol at baseline and were analysed. An extraction table needs a field for "analysis type" and a reviewer who reads the methods.

The decision: keep the dulaglutide study as supporting evidence from a non-AUD population, and add "parent trial" and "participants analysed" to the schema before extracting the rest.

3. Citation check: does the paper say what the manuscript claims

The problem: a peer reviewer reads the sentence "Once-weekly exenatide reduced heavy drinking days in adults with alcohol use disorder", cited to the 2022 exenatide trial, and wants to know whether it is accurate.

What we ran: the Contents API on the cited paper, with the claim in the summary instruction.

check = valyu.contents(
    urls=["https://pubmed.ncbi.nlm.nih.gov/PMC9675448"],
    summary=(
        "A manuscript cites this paper for the claim: 'Once-weekly exenatide reduced heavy drinking days "
        "in adults with alcohol use disorder.' Answer Supported, Partly supported or Not supported, quote the "
        "sentences that decide it, and state any subgroup or condition the claim leaves out."
    ),
)
print(check.results[0].summary)

What came back: "Partly supported", with two quotes. Exenatide "did not significantly reduce the number of heavy drinking days compared with placebo". An exploratory subgroup analysis found that in patients with a BMI above 30 it reduced heavy drinking days by 23.6 percentage points.

What we checked at the source: the paper. The exploratory subgroup had 30 patients. In patients with a BMI below 25, heavy drinking days increased with exenatide compared with placebo.

The catch: run the check on the full text. Our first attempt used the DOI link, which returned only the publisher's landing page, and the agent answered "Not supported" because it couldn't see the results. A PubMed Central or open-access link gives the agent the whole paper.

The decision: ask the authors to rewrite the sentence: exenatide did not reduce heavy drinking days overall; an exploratory analysis suggested a benefit in patients with obesity.

4. Living review update: what's new since the last search

The problem: a guideline group last searched the evidence on GLP-1 drugs and alcohol at the end of 2025 and needs everything since.

What we ran: date-limited searches of PubMed and medRxiv (we ran them as two calls; one call with both sources works too), a registry search, and a combined PubMed and web search.

new_papers = valyu.search(
    "GLP-1 receptor agonists alcohol consumption or alcohol use disorder",
    included_sources=["valyu/valyu-pubmed", "valyu/valyu-medrxiv"],
    start_date="2026-01-01",
)
trials = valyu.search(
    "GLP-1 receptor agonist semaglutide trial for alcohol use disorder",
    included_sources=["valyu/valyu-clinical-trials"],
)
journals = valyu.search(
    "semaglutide versus placebo in alcohol use disorder randomised trial 2026",
    included_sources=["valyu/valyu-pubmed", "web"],
    max_num_results=10,
)

What came back: the date-limited searches returned two 2026 systematic reviews, an October 2026 medRxiv meta-analysis on tobacco and alcohol dependence, an observational study from the NIH All of Us cohort and a rodent study of tirzepatide. The registry search returned three completed Phase 2 semaglutide trials and a recruiting VA Phase 3 trial with 622 participants planned. The search with web added two randomised trials published in 2026. In The Lancet in May, 108 patients with alcohol use disorder and obesity were treated for 26 weeks, and semaglutide reduced heavy drinking days more than placebo. In the American Journal of Psychiatry in July, 50 treatment-seeking adults took oral semaglutide or placebo for 8 weeks. It did not meet its primary outcome, laboratory alcohol craving, but reduced heavy drinking days, a key secondary outcome.

What we checked at the source: ClinicalTrials.gov and PubMed. The Lancet paper reports the Copenhagen trial NCT05895643, completed in July 2025, and PubMed links it to that NCT number. The American Journal of Psychiatry paper matches the University of Colorado trial NCT05892432 (oral semaglutide, 50 participants), but its PubMed record carries no NCT number, so a search by trial ID misses it. The third completed Phase 2 trial, NCT05891587 in Oklahoma, had no publication we could find. The VA trial, NCT07218354, has a primary completion date of March 2029.

The catch: Valyu's PubMed dataset covers open-access papers and is updated monthly, so our PubMed-only searches missed both new trials, the most important results of the year. Adding web found them. And no single lookup is complete: searching PubMed by NCT number found one trial and missed the other. For a living review, search papers, preprints, the registry and the web, and match each completed trial to its publication.

The decision: add both trials. The Lancet trial tests the subgroup finding from the 2022 exenatide trial in patients with obesity. Set a check on the Oklahoma trial and the VA Phase 3 for the next update.

5. Grant background: who is already funded to do this

The problem: a grant writer planning an application on GLP-1 drugs for alcohol use disorder needs to show the work it builds on and how it differs from what NIH already funds.

What we ran: a search of NIH RePORTER.

grants = valyu.search(
    "GLP-1 alcohol",
    included_sources=["valyu/valyu-nih-reporter"],
)

What came back: 25 projects with title, organisation, investigators and summary. Two were directly on GLP-1 and alcohol: an R01 at UTHealth Houston, "Treating Alcohol Use Disorders in Primary Care with Semaglutide", and an F32 fellowship at SUNY Binghamton on adolescent ethanol exposure and GLP-1 receptor function. A third tested GLP-1 drugs for opioid use disorder at Mayo Clinic. Most of the rest were alcohol studies without GLP-1.

What we checked at the source: NIH RePORTER. The UTHealth R01 (1R01AA033667-01) is funded by NIAAA with $842,191 in fiscal 2026 and runs from September 2026 to August 2031. It uses a Bayesian adaptive enrichment design.

The catch: grant search behaves like keyword search, so short queries work better and results need screening. A longer, more specific question returned only intramural NIH projects.

The decision: position the application against the primary-care semaglutide R01, for example by population or outcome, and cite the funded projects in the significance section.

6. Preprint to publication: did the numbers change

The problem: a journal club picked a January 2026 medRxiv preprint on GLP-1 prescriptions and alcohol consumption in the NIH All of Us cohort, and wants to know whether it has been published and whether the results changed.

What we ran: the PubMed search in job 4 returned two records for the study, the preprint and a July 2026 journal version. We extracted the main result from both.

versions = valyu.contents(
    urls=[
        "https://pubmed.ncbi.nlm.nih.gov/PMC12870566",  # preprint
        "https://pubmed.ncbi.nlm.nih.gov/PMC13347274",  # journal version
    ],
    summary={
        "type": "object",
        "properties": {
            "participants_with_two_or_more_prescriptions": {"type": "integer"},
            "main_result_active_prescription": {"type": "string"},
            "supporting_quote": {"type": "string"},
        },
    },
)

What came back: the same main result in both versions: participants with an active GLP-1 prescription had modestly lower AUDIT-C alcohol scores, an incidence rate ratio of 0.95 (95% CI 0.91 to 0.99). The journal version appeared in Alcohol, Clinical and Experimental Research on July 9, 2026.

What we checked at the source: both full texts. Both analyse 15,447 participants with at least two GLP-1 prescriptions, out of 393,596 All of Us participants with health records.

The catch: our first extraction asked for "participants analysed" and got 393,596 from one version and 15,447 from the other. A vague field name gives inconsistent answers, so name the exact denominator you want.

The decision: cite the journal version, and note that it is an observational association from a single survey, not evidence of an effect.

What the six jobs cost

Searching arXiv, PubMed and medRxiv uses 0.05 credits per result, half the price of web search, and a Contents call with a summary or schema uses 0.6 credits per paper. Clinical trial records use 0.8 credits each, which made the registry search the most expensive step. At pay-as-you-go rates of 1,600 credits for $10, the 15.1 credits cost about $0.09.

Go further with DeepResearch: an evidence table in one task

The jobs above each answer one question. A review's actual output is an evidence table and a summary of findings. DeepResearch builds these. It plans the searches across papers, preprints and registries, reads the studies and returns a cited report with files such as an xlsx evidence table.

task = valyu.deepresearch.create(
    query="Build an evidence table of randomised trials of GLP-1 receptor agonists for alcohol use disorder or alcohol consumption.",
    mode="standard",
    research_strategy="Search PubMed, medRxiv, ClinicalTrials.gov and the web. Include every randomised trial, published or registered. For each, record whether the alcohol outcome is the trial's primary outcome or a secondary analysis. Search PubMed by NCT number for completed trials.",
    report_format="A summary of findings, then an evidence table: trial, NCT number, population, intervention, duration, participants analysed, primary outcome, result with 95% CI, analysis type, publication status. Every row cited.",
    search={"included_sources": ["valyu/valyu-pubmed", "valyu/valyu-medrxiv", "valyu/valyu-clinical-trials", "web"]},
    deliverables=["An xlsx evidence table with one row per trial and a source link for every value"],
)
 
result = valyu.deepresearch.wait(task.deepresearch_id)
print(result.output)  # cited report in markdown
for file in result.deliverables:
    print(file.type, file.url)  # signed download link

The research strategy carries the lessons from the six jobs: search the web as well as open-access PubMed, label secondary analyses, and look up completed trials by NCT number.

Start from a ready-made workflow

Valyu's curated DeepResearch workflows include several built for evidence work, each with set inputs and file deliverables:

To run one from code, pass its slug as workflow_id:

task = valyu.deepresearch.create(
    workflow_id="ls-clinical-trial-tracker",
    workflow_params={"disease": "alcohol use disorder"},
)
result = valyu.deepresearch.wait(task.deepresearch_id)

For more worked examples, see the literature review use case and running a workflow in the docs.

Where AI agents stop in a literature review

The jobs above show what an agent can do. Four limits shape how to use one:

• Coverage decides recall. Valyu's PubMed dataset holds open-access papers, so paywalled journals such as The Lancet sit outside it, and the newest papers can take a month to appear. Pair it with web search and the trial registry, and for a formal systematic review run the full database searches your protocol specifies. For licensed full text, Valyu also offers Wiley Health and Life Sciences content: more than one million peer-reviewed articles from over 900 journals since 2019, with full text, figures and references, available to eligible startups through the Wiley x Valyu HLS Startup Program.
• Extraction needs a second reader. In our run, every number was right, but one study type was wrong and one field gave two different denominators. Treat AI extraction as one reviewer, not two.
• Full text beats landing pages. Checks on a DOI link that resolves to a publisher page can fail for lack of text. Use PubMed Central, arXiv or medRxiv links where they exist.
• Judgement stays with the reviewers. Risk-of-bias assessment, deciding what counts as evidence and writing the conclusions remain human work.

How to set up a literature review agent

If you don't want to write code as shown above, there is another option to connect Valyu to Claude or ChatGPT as a custom connector at https://mcp.valyu.ai/mcp. Our step-by-step guide to running a literature review in Claude or ChatGPT with verified citations walks through a full example. To build your own agent, see how to build a biomedical research agent with PubMed and ClinicalTrials.gov and the best academic search APIs for AI agents.

Three habits made the jobs above reliable:

• Search more than one source. Papers, preprints, the registry and the web each caught something the others missed.
• Ask for a verbatim quote with every extracted value. It makes each row quick to check against the paper.
• Name exact fields. "Participants with at least two prescriptions" gets the same answer from every paper; "participants analysed" doesn't.

FAQ

What is AI for literature review?

AI for literature review is the use of language models and AI agents to search, screen, extract and check research papers. Agents search databases such as PubMed and arXiv, pull the same fields from every study into a table, check citations against the cited paper and find new evidence since the last search, with a source for every finding.

What are AI agents for literature review?

AI agents for literature review are language models that search papers, preprints, trial registries and grant databases, read the full text and return cited, structured answers. They handle the searching, screening and extraction in a review, while researchers set the question, judge study quality and write the conclusions.

Can AI do a systematic review?

AI can speed up most stages of a systematic review, but not replace the method. A PRISMA systematic review needs a registered protocol, reproducible database searches, documented screening and two independent reviewers. In a November 2026 test, AI extraction got every number right but mislabelled one secondary analysis, so it works best as one of the two reviewers.

What is the best AI for literature review?

AI literature review tools fall into two groups. Research assistants such as Elicit, Consensus and Semantic Scholar search papers inside their own apps. Search APIs such as Valyu let you build your own literature review agent, or connect Claude and ChatGPT to PubMed, arXiv, bioRxiv, medRxiv, ClinicalTrials.gov and NIH grants, with a source for every result.

Can AI extract data from research papers?

Yes. Valyu's Contents API takes a JSON schema and returns the same fields from every paper, with a verbatim supporting quote. In a November 2026 test on three GLP-1 trials, every extracted number matched the paper, but one study was labelled as a primary trial when it was a secondary analysis, so a reviewer should check the methods.

How do I check whether a citation supports a claim?

Give an AI agent the full text of the cited paper and the exact claim, and ask for Supported, Partly supported or Not supported with the deciding quotes. In a November 2026 test, this showed that a claim about exenatide reducing heavy drinking overstated the trial: the effect appeared only in an exploratory subgroup of 30 patients with obesity.

What is a living literature review?

A living literature review is updated whenever new evidence appears instead of once. AI agents help by rerunning date-limited searches of papers, preprints and trial registries. In a November 2026 update on GLP-1 drugs and alcohol, adding web search found two 2026 randomised trials, in The Lancet and the American Journal of Psychiatry, that an open-access PubMed search missed.

Which databases should an AI literature review agent search?

Search papers (PubMed, arXiv), preprints (bioRxiv, medRxiv), trial registries (ClinicalTrials.gov) and the web for paywalled journals and news. For grant background, add NIH RePORTER. Valyu searches all of these through one API, and eligible startups can add the full text of more than one million Wiley health and life sciences articles through the Wiley x Valyu HLS Startup Program.

Is AI-generated literature review accurate?

It is as accurate as the sources it reads and the checks you run. In a November 2026 test, every finding matched its source, but three needed care: duplicate passages from one paper, a mislabelled secondary analysis and a vague field that returned two different denominators. Check key values against the paper before they reach a manuscript.

How much does an AI literature review cost?

In a November 2026 test, six literature review jobs with Valyu used 10 API calls and 15.1 credits, about $0.09 at pay-as-you-go rates. arXiv, PubMed and medRxiv results cost 0.05 credits each. A standard DeepResearch evidence table uses 100 credits, about $0.63, and free credits are included every month.

Start with one job

Pick the job that takes your group the most time, such as screening and extraction or keeping a review up to date, and run it on a question you have already answered by hand. Compare the agent's table with yours.

Create a Valyu account for free credits every month, up to 2,000 with a work email. Then connect it to Claude or ChatGPT at https://mcp.valyu.ai/mcp, or browse the data sources. For research beyond papers, see AI for Drug Discovery and AI for Life Sciences.