You ask an agent to search again with a different prompt, yet the same five highly cited papers return. This does not necessarily mean the field has only five relevant studies. The search may be circling the same vocabulary, ranking signals, and source network.
The fix is not “search deeper.” You need to tell the agent how to leave the neighborhood it already knows.
Why repetition happens
First, slightly different prompts may still produce the same query. “AI feedback in higher education” and “generative feedback for university students” share enough concepts to retrieve the same popular papers.
Second, relevance ranking favors well-linked, highly cited, accessible sources. Those are useful anchors, but they can crowd out newer work, regional journals, different disciplines, and studies using unfamiliar terminology.
Third, the apparent duplicates may be versions: preprint, conference paper, repository manuscript, and journal article. An agent may count them separately or repeatedly rediscover the same work.
Finally, follow-up searches often inherit the first answer's vocabulary. Once the agent calls the topic “AI-assisted formative feedback,” every later query may reuse that frame.
Diagnose the loop before changing the prompt
Ask the agent to show:
- exact queries used
- source or search mode
- which results were duplicates or alternate versions
- why each repeated paper ranked highly
- which synonyms, disciplines, and citation paths were not tried
If it cannot provide that information, start a new search and keep your own log. Repetition without a log is impossible to diagnose.
Use search moves that create real diversity
Change one dimension at a time:
- Vocabulary: use terms from relevant papers, controlled vocabularies, and other disciplines.
- Evidence type: search separately for experiments, qualitative work, reviews, theses, or preprints.
- Population and setting: name underrepresented groups, regions, or institutions.
- Time: search recent years separately instead of letting citation rank dominate.
- Citation network: follow references backward and cited-by links forward from different seed papers.
- Adversarial angle: search for null effects, criticism, failure, harms, or alternative explanations.
Tell the agent to exclude the records already reviewed and return only candidates that add a new method, population, mechanism, or conclusion.
Stop collecting when new papers stop changing the map
More unique titles are not always better. Track what each paper adds. If ten new papers repeat the same population, method, and conclusion, your evidence map may have stabilized even though the field contains more papers.
On the other hand, repeated results are not saturation if you have searched only one database, language, or terminology family. Saturation is about new information, not a tired search engine.
Use WisPaper to branch the search
Run distinct queries in WisPaper Search for the missing dimensions rather than repeatedly submitting the broad topic. Save one strong seed from each branch to My Library, then ask Library QA what populations, methods, and outcomes remain absent.
For a longer exploration, Scholar Agent can follow multiple directions. Give it an explicit novelty rule for results: a paper must add something to the evidence map, not merely use a new title.
FAQs
Keep Reading
Why Your AI Paper Summary Misses the Important Details—and What to Ask Next
AI paper summaries often miss sample details, measurements, and uncertainty. Learn what to ask next and how to trace each answer back to the paper.
What Should You Give an AI Research Agent Before It Starts Searching?
Give an AI research agent a usable brief with population, concepts, dates, exclusions, evidence rules, and the output you need.
What Is an AI Scientist—and What Can It Actually Do for a PhD Student?
Learn what today's AI scientist systems can actually do for PhD work, where they help, and which research decisions remain yours.