Source-linked AI summary
Personalized Auto-Research: Towards a True AI Co-Scientist
Bo Ni, Franck Dernoncourt, Hongjie Chen, Yu Wang, Nesreen K. Ahmed, Zhengzhong Tu, Tyler Derr, Ryan A. Rossi
TL;DR
Current AI co-scientists largely ignore the individual researcher, despite research value and feasibility depending on personal expertise, history, and context. This paper proposes graph-grounded personalization across the research pipeline and argues that researcher-specific evaluation is necessary because generic systems can produce interchangeable outputs and current held-out protocols cannot establish recommendation value.
Problem
AI co-scientists optimize research for goals, evidence, or reviewers rather than the individual researcher who will use the output.
Method
The paper proposes a graph-grounded framework that personalizes retrieval, hypothesis search, experimentation, writing, citation, review, and evaluation for each researcher.
Results
Distinct researchers pursuing the same goal receive different evidence, search trajectories, and output packages optimized for individual feasibility, alignment, and novelty.
Takeaways & Limitations
Personalization is presented as necessary for AI systems to function as genuine co-scientists rather than interchangeable research instruments.
Takeaways & Limitations
Held-out-paper evaluation can show researcher-specific signal but cannot certify that an unpursued recommendation is valuable.
Abstract
from arXiv · showhide
AI co-scientists that generate hypotheses, retrieve related work, design experiments, execute code, and draft full papers are beginning to change how research is carried out. Despite this rapid progress, state-of-the-art systems remain researcher-agnostic: given a research goal, they optimize novelty, validity, or reviewer score while ignoring the individual scientist who will use the output. This overlooks a fundamental fact about research, namely, that what counts as novel, valuable, or feasible depends on the researcher, including their prior work, methodological repertoire, and the collaborators and communities in which they are embedded. In this work, we introduce the problem of personalized auto-research, which conditions every stage of the research process on a representation of the individual researcher. We argue that personalization is not a convenience layer, but rather the fundamental property that allows an AI system to serve as a genuine co-scientist rather than a generic instrument. To address this problem, we propose a general and flexible framework that threads a graph-grounded researcher context through retrieval, hypothesis search, experimentation, writing, and review. The framework consists of three fundamental components: (i) graph-grounded researcher representations, (ii) personalization across the full research pipeline, and (iii) evaluation grounded in the individual. Notably, we highlight a one-size-fits-all failure mode where distinct researchers issuing the same goal receive essentially the same research, erasing the tacit knowledge through which novel ideas arise. Finally, we discuss fundamental open problems and challenges.
1 INTRODUCTION
Personalized auto-research addresses the researcher-agnostic limitation of AI co-scientists by conditioning every research stage on an individual researcher representation. The proposed framework uses graph-grounded context to produce research outputs optimized for each researcher’s feasibility, alignment, and novelty.
- Scientific publication has grown exponentially, making it impossible for individual researchers to fully navigate their fields or integrate adjacent-discipline insights.
- Graph-derived researcher context enables the same goal to yield distinct evidence, search trajectories, and output packages optimized for individual feasibility, alignment, and novelty.
- AI co-scientists automate hypothesis generation, literature synthesis, experiment design and execution, and manuscript drafting, but remain researcher-agnostic.
- Identical outputs erase researchers’ prior failures, methodological preferences, and experimental intuition, despite these differences often driving novel scientific directions.
- Personalized auto-research conditions every research stage on a representation of the individual researcher, distinguishing it from both generic co-scientists and personalized language models.
- The framework formalizes personalized auto-research and personalizes retrieval, hypothesis search, experimentation, writing, citation, and review through a general, flexible framework.
2 BACKGROUND
AI co-scientists now support end-to-end research through agentic search, implementation, experimentation, writing, and review, but they predominantly optimize for tasks, benchmarks, fields, or reviewers rather than individual researchers. Personalized auto-research instead treats personalization as conditioning a sequence of research decisions, extending beyond single-output recommendation or session-level guidance.
- AI Co-Scientists and Auto-Research: AI co-scientists increasingly conduct research end-to-end, combining agentic search, automated implementation, experiment execution, paper generation, and review.The AI Scientist established an autonomous propose–implement–write–review loop, while AI Scientist-v2 added progressive agentic tree search, experiment management, parallel execution, and figure feedback.
- AI Co-Scientists and Auto-Research: Despite these advances, dominant systems optimize for task, benchmark, field, literature, evidence, or reviewer objectives rather than the individual researcher adopting the output.The background characterizes this as a researcher-agnostic objective shared across current systems.
- AI Co-Scientists and Auto-Research: Scientist-in-the-loop systems provide manual, session-level guidance, but this interaction does not constitute persistent personalization.The distinction is between responding to immediate instructions and conditioning research on the individual researcher.
- Personalization: Personalized auto-research conditions a sequence of scientific decisions, including what to retrieve, what to hypothesize, and which experiments to conduct.This scope is broader than personalizing a recommendation or a single generated output.
3 PROBLEM FORMULATION
Personalized auto-research models each research stage as conditioned on a graph-grounded representation of the individual researcher, rather than on the research goal alone. Its outputs should be feasible, aligned with the researcher, and novel relative to both the field and the researcher’s prior work.
- Problem formulation: The research landscape is represented as a heterogeneous graph containing researcher, paper, venue, institution, method, dataset, and topic nodes.Edges encode co-authorship, citation, publication, affiliation, usage, and topic-assignment relationships.
- Problem formulation: Personalized auto-research conditions every stage—from literature retrieval through review—on the researcher context c_u and earlier stage outputs.The personalized stage model is o_i = f_i(g, o_<i | c_u), contrasting with researcher-agnostic models parameterized by the goal and previous outputs alone.
- Problem formulation: The desired output must be feasible under the researcher’s capabilities, resources, and constraints; aligned with their identity, community, and style; and novel.Feasibility and alignment are personalized, while novelty is partly field-level and partly relative to the individual researcher’s prior work.
4 PERSONALIZED AUTO-RESEARCH
The proposed procedure threads graph-grounded, user-specific context and evidence through an agentic auto-research loop, replacing researcher-agnostic search with personalized hypothesis evaluation and execution. It produces auditable research packages containing the artifacts, reviews, provenance, context, and evidence that conditioned each run.
- Personalized Auto-Research: The procedure augments agentic tree-search auto-research with a user, context, personalized evidence set, experiment manager, execution, figure refinement, writing, review, and provenance logging.It replaces the linear 2024 template loop with a loop over partial research states.
- Personalized hypothesis search: Hypotheses are ranked by personalized novelty, relevance, and feasibility rather than a binary global novelty test, allowing the same idea to differ across researchers.The utility is U(h| g,u) = αNov(h, I,c_u) + βRel(h,g,c_u) + γFeas(h, W0,c_u).
- Personalized context and evidence: Graph encoding, context encoding, and retrieval construct evidence relevant to both the research goal and the researcher’s prior work, collaborators, methods, resources, and graph position.The evidence is not simply the literature topically closest to the goal.
- Personalized hypothesis search: User-conditioned trees give researchers with the same goal different selection policies, expansion distributions, utilities, and execution constraints.The search tree is B_u, and candidates are implemented and executed under the user’s constraints.
- Personalized package synthesis: Top-m states are synthesized into reproducible packages containing the hypothesis, personalized score, code, outputs, figures, paper, review, provenance, context, and evidence needed to audit the result.The package is Q_u and preserves the metadata explaining why it was produced for that researcher.
5 PERSONALIZED AUTO-RESEARCH VISION
Personalized auto-research is defined by its relationship to the individual researcher: personalization makes an AI co-scientist a collaborator rather than merely a capable instrument. The proposed agenda centers on researcher representation, personalization across the research pipeline, and evaluation grounded in the individual.
- Personalization distinguishes an AI co-scientist as a collaborator whose outputs are tailored to the researcher, rather than merely a capable instrument.The paper treats personalization as what makes the collaborator metaphor accurate, not as a marginal improvement.
- Researcher Representation: Researcher identity should be represented through graph position and multi-hop neighborhoods, not publications alone, to capture collaborators, adjacent topics, and community structure.This representation can expose structural holes linking the researcher’s region to a nearby, unconnected region.
- Pipeline Personalization: Personalization must span retrieval, hypothesis generation, experiment design, writing, citation, and review rather than stopping at hypothesis generation.The framework separates exploitation of the researcher’s trajectory from exploration that uses the profile to assess feasibility and framing.
- Evaluation Grounded in the Individual: Evaluation should replace next-paper prediction with feasibility alignment and expert-assessed quality grounded in the individual researcher.Held-out papers reflect path-dependent incentives and can penalize recommendations that are better than anything the researcher previously pursued.
6 OPEN CHALLENGES
The paper identifies open challenges as deep tensions in personalized auto-research, including creativity collapse, career-stage dependence, team personalization, and evaluation without ground truth. These challenges require balancing researcher-specific context with collective discovery and developing evaluation beyond historical mimicry.
- Creativity Collapse: Researcher-agnostic systems can collapse collective exploration into a monoculture, redundantly racing globally best ideas while neglecting directions only particular researchers can pursue.Personalization is framed as a decorrelation mechanism for collective discovery, not merely an individual convenience.
- Creativity Collapse: Personalization must use a researcher’s experience to surface counterfactual directions they can develop without reducing recommendations to biographical mimicry.The challenge is to identify experiments a generic system would not propose but the specific researcher can pursue.
- Lifecycle Dependence and Cold Start: The personalization objective changes with career stage: senior researchers need adjacent unexplored territory, while early-career researchers need help establishing an identity.This exceeds classical cold start because the objective function itself changes as a researcher’s career stage changes.
- Team-Personalized Auto-Research: Team-personalized auto-research can represent a team as a subgraph and aggregate feasibility asymmetrically, allowing execution of what any member can execute.Novelty is defined relative to the union of team members’ prior work.
- Evaluation Without Ground Truth: Evaluating recommendations is difficult because a researcher’s value for an unpursued direction is never observed, making historical similarity reward mimicry rather than better recommendations.Held-out evaluation establishes only that the researcher representation carries researcher-specific signal, not that a recommendation is good.
- Evaluation Without Ground Truth: Resolving evaluation requires expert judgment, longitudinal studies, or counterfactual interventions, each of which remains an open research need.These approaches address the distinction between researcher-specific signal and the value of unpursued alternatives.
7 CONCLUSION
The paper introduces personalized auto-research as conditioning the full research process on the individual researcher. It argues that a true AI co-scientist must pursue complementarity by knowing whom it collaborates with, rather than simply producing a stronger universal output.
- Conclusion: Personalized auto-research conditions the full research process on a representation of the individual researcher.The paper frames this as its central problem.
- Conclusion: A system cannot be a true co-scientist if it does not know whom it is collaborating with.This is presented as the paper’s core claim.
- Conclusion: The goal is complementarity rather than similarity, because a stronger universal engine still returns the same high-scoring package to every researcher.The paper compares this with scientists choosing collaborators for expertise they lack.