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
cs.AI, cs.CL
2026-08-15
A position paper: auto-research systems return the same package to every researcher, erasing tacit knowledge; it proposes graph-grounded researcher representations, full-pipeline conditioning, and individual-grounded evaluation.
The AI Scientist lineage can now propose hypotheses, run experiments, and draft full papers, and AI Scientist-v2's progressive agentic tree search has passed workshop-level review. All of it shares one structural blind spot: the systems are researcher-agnostic. Feed in the same goal and the output distribution is identical whether the requester is a first-year PhD student or a senior professor, a graph-mining researcher or a computational biologist. That contradicts how discovery actually works: new directions grow out of a scientist's idiosyncratic mix of past failures, methodological taste, and experimental intuition, and interchangeable output erases exactly the heterogeneity that makes research creative. The authors price the blind spot at three levels. Epistemically, most scientific capability is tacit and absent from the literature, unreachable by any literature-conditioned system. Systemically, many researchers querying one generic engine collapses the field's hypothesis portfolio toward a monoculture, racing the same "best" ideas while counterfactually valuable directions go unexplored. Practically, researchers only verify and adopt directions matched to their expertise and resources.
This is a position-plus-framework paper with no experiments. It poses personalized auto-research: condition every stage of the process (retrieval, hypothesis, experiment design, code execution, writing, citation, review) on a researcher representation, formalized as oi = fi(g, o<i | cu), where cu is the researcher's operationalized context.
Three components:
Algorithm 1 wires all of this onto the SOTA agentic research loop (tree search, experiment manager, code execution, automated review) with optional human critique folded back into the context.
As a position paper its evidence is drawn from cited literature (most impactful research comes from teams; structural holes and good ideas), with no experiments of its own. What is worth reporting is the decomposition of open challenges, each an internal tension rather than an engineering obstacle. Creativity collapse demands rewarding counterfactual complementarity (candidates unlikely under both a universal model and a user-mimicry model yet valuable given accumulated experience); one realization trains a mimicry model of the researcher and optimizes for high utility at low likelihood under both. Lifecycle dependence exceeds the classic cold-start problem because the objective itself changes with career stage: seniors need bridges across structural holes while early-career researchers need to establish an identity. Team personalization aggregates asymmetrically, feasibility as a union, novelty relative to the union of prior work, alignment closer to an intersection, and that asymmetry is precisely why scientists collaborate. And evaluation has no ground truth: the value of a direction to a specific researcher is never observed, because history records only the single trajectory actually followed.
In the AI-for-science infrastructure race this is a cheap and correctly aimed reminder: a stronger universal engine does not solve the problem, it still returns the same high-scoring package to everyone, so personalization is orthogonal to and required on top of current SOTA. For teams working on agent memory and personalized LLMs, the problem list is a ready-made topic pool: researcher representation (multi-hop encoding on the research graph), learning individual-conditioned utilities (fitting the weights from revisions, reviews, and abandoned-versus-published traces), and counterfactual-complementarity reward design are each independently startable subproblems. The monoculture-risk argument also lands for lab directors and funders: if the whole field queries the same three engines, the loss of hypothesis diversity is global.
The paper concedes all of this itself: no system, no experiments, no benchmarks, and the authors say benchmark infrastructure for individual-grounded evaluation is itself a missing major contribution. The heaviest assumption is that a research-graph representation carries enough individual signal; support is indirect (structural-hole literature), and direct evidence must wait for the held-out protocol to run. Privacy is a hard deployment obstacle, since co-authorship graphs, review history, and code repositories are sensitive signals, and the paper dispatches this in one line about HCI and consent. Counterfactual complementarity is elegant but may be unoptimizable: most candidates unlikely under both the mimicry and universal models are simply bad ideas, and the line between filtering bad ideas and keeping weird-but-good ones has no operational criterion. The social-choice problem of aggregating conflicting member preferences into a single team context is flagged by the authors as open.