Explainable AI is increasingly important to scientific discovery. However, existing methods largely ignore that explanation quality is not universal: experts differ in how they assess evidence, prioritize mechanisms, and construct explanatory narratives. We introduce perspective-conditioned explanations, a framework for adapting explanation generation to epistemic variation in expert judgment. Using knowledge graph reasoning paths in drug discovery, we show that preferences organize into coherent epistemic perspectives that can be captured by agentic personas, representations of how experts evaluate explanations. Persona-aligned rewards then guide reinforcement learning-based explanation generation without large-scale expert supervision. Expert user studies show that perspective-conditioned explanations are preferred over general-purpose explanations and improve perceived relevance and validity. Moreover, they match or exceed state-of-the-art predictive performance and reduce expert feedback time by two orders of magnitude. Together, these findings demonstrate that explanation quality is perspective-dependent and that modeling this variation enables scalable and human-aligned explanation generation for scientific discovery.
Figures & tables
Figure 1: General-purpose and perspective explanation examples for Fenofibrate treats Coronary Artery Disease . (a) The general-purpose explanation (generated by REx) includes peripheral paths such as side effects and loosely related drug classes. (b) The perspective explanation, conditioned on a persona that prioritizes mechanistic directness, selects paths that focus on the core therapeutic mechanism and relevant biological targets.
Figure 2: Overview of the perspective explainability approach. Phase I creates agentic personas by embedding expert qualitative feedback, clustering responses to identify shared evaluative patterns, and synthesizing structured persona narratives via LLM summarization. Phase II integrates these personas into the reinforcement learning loop: as the RL agent generates candidate explanation paths over the knowledge graph, each persona scores them according to its epistemic perspective, producing a persona-conditioned reward signal that guides the agent toward explanations aligned with distinct expert reasoning strategies.
Figure 3: Persona discovery: quantitative evidence. a,b , Persona creation ratings by system and task (means ± SD on 1–5 scale), showing high variance reflecting interpretive diversity. c , Mean pairwise Spearman correlations ( r ) between expert ratings and between expert and persona ratings across all drug repurposing and drug-target interaction tasks; all p<.001 . d , Clustering of expert feedback embeddings projected via PCA, comparing K-Means, Agglomerative, and HDBSCAN algorithms. K-Means and Agglomerative converge on an identical two-cluster solution, while HDBSCAN does not resolve cluster structure.
Figure 4: Agentic persona profiles and assignment. a , Narrative profile and evidence summary for Elena , who prioritizes mechanistic rigor. b , Narrative profile and evidence summary for Leo , who prioritizes focused clarity. c , Three-question profiling instrument used to assign participants to personas for the persona evaluation study based on discriminative characteristics identified in expert feedback. For each question, the first response option (purple) aligns with Elena’s preferences, while the second option (green) aligns with Leo’s preferences. Response order was randomized to control for ordering effects.
Figure 5: Perspective explanation evaluation. a , Average ratings across evaluation criteria for general-purpose (REx) and persona-conditioned (Leo, Elena) explanations per hypothesis, with REx ratings reported as the shared baseline. b , Spearman correlation between persona ratings and their constituent participants’ ratings for drug repurposing in the persona creation study. c , Spearman correlation between persona ratings and their assigned participants’ ratings for drug repurposing in the persona evaluation study. d , Participant preferences between perspective and general-purpose explanations. e , Estimated feedback collection time under different supervision conditions; persona-conditioned training reduces expert time by two orders of magnitude. f , Link prediction performance on the drug repurposing task; perspective explanations exceed state-of-the-art methods across all metrics. Asterisks in b and c denote significance: ∗p<.05 , ∗∗p<.01 , ∗∗∗p<.001 .
Figure 6: Example explanations for the hypothesis Etidronic acid treats Paget’s disease of bone : a , general-purpose (REx); b , the Elena perspective (mechanistic rigour); c , the Leo perspective (focused clarity).
Appendix figures & tables6 assets
Supplementary material from the paper’s appendix.
Appendix
Study (n = 11)
Age Range
25–34: 9; 35–49: 2
Primary Background
Life and Health Sciences: 4 CS/AI-heavy w/ Biomedical Experience: 5 Hybrid Computational Biology: 2
Domain Knowledge
Knowledge Graphs: 4 Experts, 2 Competent, 3 Novice, 2 No Knowledge AI Systems: 1 Expert, 7 Competent, 1 Novice, 2 No Knowledge Molecular Biology: 2 Experts, 8 Competent, 1 No Knowledge
Appendix
Table 1: Persona creation study demographics, expertise, and domain knowledge ( n=11 ).
Figure 7: Clustering validation. a , Candidate solutions ( k=2 – 5 ): Silhouette Score, Inertia, Davies–Bouldin Index, Calinski–Harabasz Index, Gaussian mixture model selection (BIC and AIC), and bootstrap cluster stability (Jaccard index, 200 resamples). Orange rings indicate the best k for each metric. b , Subcluster analysis within the majority cluster (Elena, n=13 ): K-Means split ( k=2 , silhouette = 0.13), HDBSCAN (3 micro-clusters, 6 noise points), and per-sample silhouette coefficients for the base k=2 solution.
Category
Distribution
Age Range
20–24: 2; 25–29: 12; 30–34: 4; 40–44: 3; 45–54: 1
Persona Assigned
Leo: 12; Elena: 10
Primary Background
Life and Health Sciences: 15 (68.2%) CS/AI-heavy with Biomedical Experience: 3 (13.6%) Hybrid Computational Biology: 4 (18.2%)
Main Expertise Area
Biochemistry: 2
Animal Cell Technology: 1
Microbiology: 2
Data Analytics: 1
Computer Science: 2
Biochemistry and Biotechnology: 1
Appendix
Table 2: Persona evaluation study participant demographics, background categories, main expertise areas, and domain knowledge ( n=22 ).
Contrast
β
SE
95% CI
p
Relevance
+0.468
0.082
[+0.31,+0.63]
<.001
Completeness †
+0.245
0.129
[−0.01,+0.50]
.057
Validity
+0.450
0.085
[+0.28,+0.62]
<.001
Overall
+0.388
0.059
[+0.27,+0.50]
<.001
Elena users
+0.570
0.081
[+0.41,+0.73]
<.001
Leo users
+0.236
0.084
[+0.07,+0.40]
.005
Appendix
Table 3: Linear mixed-effects model of explanation ratings ( n=22 ; 1,320 observations) with crossed random intercepts for participant and hypothesis, fit by REML. Coefficients give the perspective-minus-general difference on the 1–5 scale.
Comparison
Pairs
Persona paths/pair
REx paths/pair
Path length (hops)
Pairs sharing no path
Elena vs REx
46
1.83
2.11
3.0
65%
Leo vs REx
44
1.66
2.16
3.0
50%
Appendix
Table 4: Path properties of general-purpose (REx) and persona-conditioned explanations on the drug repurposing task.
Question
Elena
Leo
Q1: Mechanistic Detail
“A detailed step-by-step explanation” Supporting statements: - “combine therapeutic, genetic, and adverse-effect details for depth” - “combine multiple mechanistic or genetic details to reinforce direct drug-disease links”
“A streamlined explanation with fewer steps” Supporting statements: - “offer multiple paths for completeness while limiting their number to avoid excessive complexity” - “concise explanations that link drugs by shared treatment lines”
Q2: Explanation Complexity
“I do not need high-level concepts, I am happy with seeing only a detailed explanation” Supporting statements: - “explicitly state direct, clinically established drug-disease links” - “avoid generic class or side-effect links, deliver a clear therapeutic rationale”
“I prefer to see high-level concepts that frame why things belong together” Supporting statements: - “use ontological expansion to clarify connections between claims” - “leverage ontology for non-obvious or indirect relations”
Q3: Decision-making Context
“Explanations should be based on specific relations such as ‘inhibits’ or ‘binds’ ” Supporting statements: - “prefer ‘binds’ to ‘causes’ for drug-target interactions” - “eliminate vague predicates like ‘causes,’ ‘includes,’ or ‘participates’ ”
“Explanations can also include less specific relations such as ‘treats’ or ‘causes’ ” Supporting statements: - “relations…already covered by ‘causes’ or ‘treats’ ” (accepting them) - “ontological expansion specifically to indirect relations”
Appendix
Table 5: Mapping of profiling questions to persona characteristics based on expert feedback analysis.
Modern artificial intelligence excels at prediction but cannot explain. From large language models to AI-for-science systems, today's machines answer what by recombining patterns already present in the human literature, yet they cannot reason out why a phenomenon must arise from underlying principles even though explanation, not prediction, lies at the heart of scientific discovery. Here we ask whether the structure of scientific explanation can be operationalized to guide how a machine generates hypotheses. We introduce DN-Hypo-Pipeline, a hypothesis-generation framework that adopts a layered, explanation-theoretic scaffold: Hempel's deductive-nomological (DN) model supplies the output form and deductive validity of a hypothesis, Salmon's causal-process account supplies an organizing constraint on where to search for the governing laws, and Armstrong's view of laws as relations between universals supplies the bridge from a phenomenon's constituent processes to the laws that may be associated with it. Rather than searching the space of what has been written, the framework searches the space of what principles govern a phenomenon: given an explanandum, it abstracts the universals instantiated in the phenomenon's formation process, retrieves the laws relating those universals, and deductively reconstructs a new, testable explanation. Evaluated in data-science modeling and judged by both LLMs and human experts, hypotheses generated through this principled reasoning significantly outperform those from direct prompting. Crucially, we translated the two highest-scoring hypotheses into novel algorithms one that reduces the Transformer's theoretical complexity with only minimal performance loss, and another that achieves competitive accuracy with substantially fewer parameters.
Lei Lin, Ronghao Wang, Chunbao Zhou +2
Computer Network Information Center, Chinese Academy of Sciences, China
Do expert personas improve language model performance? The Wharton Generative AI Lab reports that they do not, broadcasting to millions via social media the recommendation that practitioners abandon a technique recommended by Anthropic, Google, and OpenAI. We demonstrate that this null finding was structurally predictable. Five core mechanisms precluded detection before data collection began: baseline contamination elevating the starting point to near-ceiling, system prompt hierarchy subordinating experimental manipulation, impossible expert specifications collapsing to generic competence, format constraints suppressing reasoning processes, and provider exclusion limiting generalizability. Controlled trials correcting these limitations reveal what the original design obscured. To test this, we selected the GPQA Diamond hardest questions to prevent baseline pattern matching, forcing reliance on genuine expert reasoning. On items with valid key answers, expert personas achieve ceiling accuracy. They eliminated all baseline errors through confidence amplification. Furthermore, forensic examination of model divergence identified that half of the hardest GPQA items contain chemically or logically indefensible answers. The model's CoT revealed reasoning away from impossible answers, yielding penalization for accurate chemistry. These findings recontextualize the original null results. Methodologically sound persona research faces measurement constraints imposed by benchmark validity limitations. Answering the persona question requires evaluation infrastructure the field does not yet possess.
Drake Mullens, Stella Shen
Tarleton State University · University of Texas at Tyler
Persona prompting can steer LLM generation towards a domain-specific tone and pattern. This behavior enables use cases in multi-agent systems where diverse interactions are crucial and human-centered tasks require high-level human alignment. Prior works provide mixed opinions on their utility: some report performance gains when using expert personas for certain domains and their contribution to data diversity in synthetic data creation, while others find near-zero or negative impact on general utility. To fully leverage the benefits of the LLM persona and avoid its harmfulness, a more comprehensive investigation of the mechanism is crucial. In this work, we study how model optimization, task type, prompt length, and placement can impact expert persona effectiveness across instruction-tuned and reasoning LLMs, and provide insight into conditions under which expert personas fail and succeed. Based on our findings, we developed a pipeline to fully leverage the benefits of an expert persona, named PRISM (Persona Routing via Intent-based Self-Modeling), which self-distills an intent-conditioned expert persona into a gated LoRA adapter through a bootstrapping process that requires no external data, models, or knowledge. PRISM enhances human preference and safety alignment on generative tasks while maintaining accuracy on discriminative tasks across all models, with minimal memory and computing overhead.
Zizhao Hu, Mohammad Rostami, Jesse Thomason
University of Southern California Los Angeles, California, USA