acceptodds
Under review as a conference paper at ICLR 2027

Developing Niche Persona for Personalized Decision-making through Symbolic Learning

Abstract

Ecological niches illustrate the close relationship between the information organisms perceive and the behavioral strategies they employ. A similar interdependence arises in modeling human decisions across heterogeneous populations, which requires determining both which information is relevant and how that information should be used. The usefulness of a feature set depends on the decision model, while the model’s effectiveness depends on the information it receives. This paper captures this interdependence through the concept of decision niches. We propose Niche Persona from Observation (NPO), a theory-guided LLM-agent framework that jointly conducts feature selection and human decision modeling. NPO defines a decision niche as the alignment between three elements: the population being modeled, the information exposed to the model, and the behavioral theory used to structure reasoning. For each group, theory-guided decision agents learn choices under candidate behavioral assumptions, while selection agents refine the visible feature space using validation feedback and error patterns. We evaluate NPO on three real-world datasets covering job choice, marijuana behavior, and alcohol behavior. Across all LLM backbones, NPO achieves average relative F1 gains of 9.38%–10.24% over the strongest baseline for each dataset, while using up to 43.75% fewer features than the strongest LLM-based baseline. Beyond prediction, NPO uncovers task- and group-specific decision niches, such as the stronger role of structured trade-offs in job choice and distinct theory–feature alignments across health-related behaviors. With NPO, personalized decision agents achieve stronger prediction with less information, select group-specific behavioral theories, and lay the foundation for future decision companions. The source code will be fully released upon acceptance.

Primary area:
neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)
Keywords:
Decision-making, LLMs, Information Alignment

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_U9iXOpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes Niche Persona from Observation (NPO), an LLM-based framework for personalized human decision modeling. Motivated by the interdependence between feature selection and behavioral theory selection, NPO introduces a co-evolutionary process in which decision agents are optimized under different behavioral theories, while a selection agent iteratively refines the feature set based on validation feedback. The framework identifies group-specific feature–theory combinations to improve decision prediction. Experiments on three real-world datasets involving job choice, alcohol use, and marijuana use demonstrate improvements in predictive performance and feature efficiency over existing baselines.

Critical strengths

The paper presents an interesting perspective on personalized decision modeling by considering the interdependence between feature selection and behavioral theory selection. The idea of identifying group-specific feature–theory combinations is intuitive and well motivated. I also appreciate the attempt to incorporate established behavioral theories into LLM-based decision agents, which provides a more structured approach to human behavior prediction. The experimental results across three real-world datasets and multiple LLM backbones are encouraging, showing improvements in both predictive performance and feature efficiency.

Critical weaknesses

The paper's motivation that feature selection and behavioral theory selection should be considered jointly is reasonable. Although Tables 2 and 3 demonstrate the benefits of feature selection and theory competition, they do not sufficiently establish whether NPO's specific co-evolutionary design is more effective than simpler joint optimization strategies, such as conventional alternating optimization or direct LLM-based joint search. In particular, since the search space and validation objective are explicitly defined, a natural alternative would be to use a general-purpose LLM agent to directly search for effective feature–theory combinations. Given that NPO already employs LLMs to guide feature selection, such an agentic search approach seems to be a relevant baseline. A comparison under the same evaluation budget would help clarify whether NPO's structured co-evolutionary process provides meaningful benefits over a more straightforward joint search strategy.

Questions for authors

Justification of the Framework:

The motivation for jointly considering feature selection and behavioral theory selection is reasonable, but what justifies the relatively complex design of NPO over simpler joint optimization approaches? For example, why not include a conventional alternating optimization baseline that iteratively updates the feature subset under a fixed behavioral theory, and then selects the best-performing theory under the updated feature set? Such a comparison would help determine whether NPO's improvements arise from the general idea of alternating feature–theory optimization or from its specific LLM-guided search strategy.

Comparison with General Agentic Search:

Since the search space and validation objective are explicitly defined, could a general-purpose LLM agent directly search for effective feature–theory combinations? Why is such an approach not included as a baseline?

6 comments
AuthorsOpenReview ↗

Response to Reviewer U9iX

We sincerely appreciate the reviewer's insightful comments regarding comparing NPO with simpler search strategies. We evaluated LLM-based alternating optimization and direct joint search with general-purpose agents below.

W1 Effectiveness of framework and general agent search

We agree that the co-evolution of theories and features in NPO introduces additional complexity and we would like to address this concern from the following aspects:

Conventional search strategy. PySR and gplearn provide relevant comparisons with conventional symbolic search: they search over both the variables appearing in an expression and its functional structure. We also included selection strategies such as recursive feature elimination and forward selection, which also jointly conduct feature selection and model evaluation. These baselines therefore address whether a simpler numerical joint search can provide competitive predictive models.

NPO connects alternating optimization and agentic search. We agree that those baselines still differ from the alternating optimization method. However, it is difficult to define an established numerical alternating optimization baseline for this setting. The main challenge is that candidate theories differ in how they specify choice probabilities and estimate parameters, so such a baseline would need to combine theory-specific fitting procedures with a common validation criterion.

This motivates NPO’s design. An LLM uses validation feedback to propose feature updates, and each candidate theory retains its own fitting procedure. We address the general-purpose agentic alternative separately through Direct LLM Joint Search, in which the agent proposes feature–theory pairs directly.

Direct comparisons. To address the challenge discussed above, we evaluated an LLM-driven alternating optimization approach (LLM-AO) as a baseline in Q1. We use GPT-4o-mini as a common selector for theory-feature combinations. In Q2, we further evaluate general-purpose LLM agents that directly propose feature–theory pairs. We will add a section describing both LLM-AO and direct search with general-purpose agents in the revised version.

Q1 Justification of the framework

Following the suggestion to evaluate a more straightforward alternating optimization approach, we implemented LLM Alternating Optimization (LLM-AO) as a baseline for comparison with NPO’s structured co-evolutionary procedure. The results are presented below and we would like to provide more details and further justification in Response Part 2.

MethodALCMAJJOB
LLM-AO (GPT-4o-mini)0.473 / 0.4920.563 / 0.5900.590 / 0.613
NPO (GPT-4o-mini)0.496 / 0.5130.673 / 0.7110.655 / 0.635

Q2 Comparison with general agentic search

We agree that general-purpose LLM agents provide a relevant baseline. To ensure a fair comparison, we packaged the background information and model-fitting process of the five theories into a skill, which we provided to the agents. We chose two representative agents: Codex CLI (version 0.167.0) and Claude Code CLI (version 2.1.311). For Codex, we evaluated GPT-6-Astra and GPT-6-Luna. For Claude Code, we evaluated Claude-Opus5.5 and Claude-Haiku5.5. We set the reasoning effort to medium for all models and used the same search space, validation objective, and other settings as in NPO. All agents searched directly for feature–theory combinations using validation feedback. The detailed results are presented below:

MethodALCMAJJOB
Vanilla LLM (F)0.284 / 0.3380.088 / 0.1910.224 / 0.230
TextGrad (F)0.274 / 0.3120.106 / 0.1680.389 / 0.426
Vanilla LLM (S)0.315 / 0.3750.137 / 0.2110.308 / 0.287
TextGrad (S)0.318 / 0.4250.156 / 0.2110.374 / 0.357
Codex CLI+ GPT-6-Luna0.411 / 0.4000.542 / 0.5040.531 / 0.520
Codex CLI+ GPT-6-Astra0.446 / 0.4330.537 / 0.5000.544 / 0.541
Claude Code+Haiku 5.50.419 / 0.4040.562 / 0.5270.567 / 0.557
Claude Code+Opus 5.50.437 / 0.4290.555 / 0.5200.530 / 0.525
NPO+GPT-4o-mini0.496 / 0.5130.673 / 0.7110.655 / 0.635

Direct agentic search with Codex and Claude Code achieves substantial improvements over LLM baselines such as vanilla LLM and TextGrad. However, despite using more capable backbone models, these general-purpose agents still underperform NPO with GPT-4o-mini across all three datasets, which shows the value of the co-evolution framework of NPO.

AuthorsOpenReview ↗

Response to Reviewer U9iX (Part2)

Q1 Justification of the framework

The main challenge of comparing NPO with conventional numerical alternating optimization lies in the difference of fitting procedures and parameter settings between theory candidates. Our baselines provide two types of approximations: PySR and gplearn adopt iterative evolutionary search based on the results of symbolic regression, while forward selection and recursive feature elimination alternately optimize the model and the feature set. We also set up the LLM-AO as the alternating optimization baseline using GPT-4o-mini. LLM-AO iteratively updates the feature subset under a fixed behavioral theory, and then selects the best-performing theory under the updated feature set.

One major motivation of NPO is to bridge the gap between these well-acknowledged theories and conventional alternating optimization. By leveraging the generalizability of LLMs, NPO can incorporate updated features into each behavioral model in a form appropriate to its underlying theory. It can also draw on feedback from the candidate theories to propose new features for behavioral prediction. Beyond predictive performance, the feature–theory alignment analysis in Figure 3(a) illustrates how prior research can help interpret the features selected by NPO [1,2].

[1] Fergusson, David M., and Joseph M. Boden. "Cannabis use and later life outcomes." Addiction 103.6 (2008): 969-976.

[2] Yamaguchi, Kazuo, and Denise B. Kandel. "On the resolution of role incompatibility: A life event history analysis of family roles and marijuana use." American journal of Sociology 90.6 (1985): 1284-1325.

Thanks for the response from the authors. I appreciate the additional experiments on the coding agents and this is my major concern given the coding agent is really powerful nowadays. I am still wondering if the authors can provide more details of how they conduct the agent experiments. How many iterations can an agent try in your experiments (feedback - submission loop)? What the agent receives as the feedback? How many times have you tried the general coding agent? Since the Codex and Claude Code achieves substantial improvements, what's their limit performance if they were given enough trial budgets (keep iteration until no better solution found)? I am interested in this because if codex and claude can be as good as NPO or even better given enough budget, why we need a relative complex method like NPO?

Reviewer_tr5WOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes NPO, an LLM-based framework for jointly selecting features and theory-guided prediction models. It partitions respondents into age–income groups and learns a separate predictor for each group. Within each group, an LLM proposes feature additions based on validation feedback and prediction errors, while candidate predictors use five behavioral theories to structure their utility functions and textual rules. The final predictor is selected using validation performance. Experiments on three real-world survey datasets covering employment status, alcohol use, and marijuana use show improved classification performance over conventional machine learning and LLM baselines, often using fewer features.

Critical strengths

The strongest point is the evaluation across three real-world datasets and four LLM backbones. NPO consistently improves over the conventional feature-selection baselines, and the fixed-theory comparisons provide evidence that allowing theory selection is useful.

Critical weaknesses
  • The benefit of group-specific modeling is not established. The method splits each dataset into nine age–income groups and learns a separate predictor for each, leaving only a few dozen training examples per group on average. Although this supports explicit group-level descriptions, it prevents sharing training data across groups and may weaken the conventional ML baselines. A pooled model using age and income as input features is an important missing comparison: it could capture group differences while benefiting from all available training data. Without this comparison, it is unclear whether separate “decision niches” improve prediction or unnecessarily fragment the data.
  • The computational overhead is substantial relative to the scale of the tasks, and the performance–cost tradeoff is insufficiently evaluated. With GPT-4o-mini, each dataset requires roughly 40,000–48,000 LLM calls despite containing only a few hundred training examples. It therefore remains unclear whether the predictive gains justify this complexity over simpler alternatives.
Questions for authors
  1. What is the methodological novelty beyond LLM-guided feature selection combined with validation-based model selection? How does NPO differ from a standard joint search over feature subsets and candidate models?

  2. How does NPO compare with a pooled model trained on all respondents, using age and income as input features? This would clarify whether separate group-level models are beneficial despite their smaller training sets.

6 comments
AuthorsOpenReview ↗

Response to Reviewer tr5W

We thank the reviewer for suggestions about the pooled baselines, the cost–performance trade-off, and what NPO contributes beyond joint search. We would like to provide our answer and clarifications as below.

W1 Pooled ML baselines evaluation

The reported results of group-specific ML baselines are intended to match NPO’s group-level modeling design and ensure a fair comparison with further evaluation using RFE as upstream strategy to search group-specific features. We agree that pooled ML models are important comparisons. We evaluated pooled LR, RF, and XGBoost on all three datasets. Each model was trained on the union of the canonical group training splits, with age and income included as input features. Results are presented below in Q2.

W2 Computational cost

We agree that the performance–cost trade-off should be highlighted. NPO indeed introduces additional overhead over LR, RF, and XGBoost because of the API costs of LLMs. We would like to address this concern from the following perspectives:

Agentic decision modeling. NPO provides a direction for theory-grounded LLM agents by maintaining the selected behavioral theory and feature set as transparent, inspectable checkpoints. We collect the costs of two LLM frameworks, Athena and DualMind, under the same settings using the full feature set as above. This comparison can further illustrate the cost-performance trade-off:

DatasetMethodCost (USD)Weighted F1
JOBAthena$1.350.499
JOBDualMind$2.130.612
JOBNPO$5.770.655
MAJAthena$6.580.569
MAJDualMind$9.770.619
MAJNPO$9.020.673
ALCAthena$6.720.396
ALCDualMind$8.860.483
ALCNPO$7.890.496

On the high-dimensional MAJ and ALC tasks, NPO’s outer feature selection loop avoids repeatedly processing the full feature set, allowing it to update five candidate theories at a lower cost than full-feature DualMind. Furthermore, NPO achieves the highest weighted F1 on all three tasks. We will add this trade-off discussion to the revised version.

Local Deployment. The competitive performance achieved by NPO using Qwen3-32B suggests that the corresponding agent components can use a locally deployed opensource model, reducing reliance on paid external APIs and offering a route to lower operating costs when suitable local computing resources are available.

Q1 Methodological novelty and scotietal benefits

We agree that jointly selecting feature subsets and candidate models is not, by itself, novel in the domain of AI research. NPO contributes a theory-guided update procedure and an explicit representation of the resulting group-specific decision model.

Theory-grounded reasoning. Transparency is a common concern in LLM-based frameworks because a prediction alone reveals little about how the LLM used the input features to compare alternatives [1]. NPO uses established behavioral choice theories to specify an explicit structure linking selected features to each option. As presented in Table 3, the selection LLM further synthesizes validation feedback across these candidates, which allows the agent to draw on several behavioral perspectives.

Personalized decision agents. NPO discovers group-specific decision niches and retains the corresponding information such as feature sets. It provides a direction towards group-level trustworthiness and diversity. In Figure 3(a), RRM accounts for 32 of the 54 final selected features on MAJ. During feature selection, the young, high-income group added marital status and several employment variables. These choices are consistent with longitudinal research linking cannabis use to employment outcomes and family roles in young adulthood [2,3]. The resulting niche gives agent designers a concrete basis for choosing which aspects of a target group’s circumstances to represent.

Q2 Comparison with pooled models

We agree that the group division may weaken the performance of ML baselines. We evaluated the performance of pooled ML baselines with age and income included as features, and provide the results below with weighted F1. NPO still achieves competitive performance on three tasks.

MethodALCMAJJOB
Pooled LR0.4690.5380.648
Pooled RF0.4610.5280.649
Pooled XGB0.4740.5890.632
NPO (GPT-4o-mini)0.4960.6730.655

[1] Shen, Xu, et al. "Faithcot-bench: Benchmarking instance-level faithfulness of chain-of-thought reasoning." International Conference on Learning Representations. Vol. 2026. 2026.

[2] Fergusson, David M., and Joseph M. Boden. "Cannabis use and later life outcomes." Addiction 103.6 (2008): 969-976.

[3] Yamaguchi, Kazuo, and Denise B. Kandel. "On the resolution of role incompatibility: A life event history analysis of family roles and marijuana use." American journal of Sociology 90.6 (1985): 1284-1325.

Thank you for the detailed response and the additional experiments. After careful consideration, I decide to keep my score.

  • The pooled baselines do not address my concern about modelling each group separately. Pooled classical models nearly match NPO on two of the three datasets.
  • The cost table resolves my concern regarding cost reporting, however it compares only against other LLM frameworks. Given that classical models reach similar performance at almost no cost, the volume of LLM calls is not straightfoward to justify.
AuthorsOpenReview ↗

Response to the Followup of Reviewer tr5W (Part1)

Thanks for your reply, and we appreciate your comments, but we still want to do our best and try to seek another potential increase. We would like to address your questions and concerns based on the following answer. First, we would like to clarify our goal, motivation, and the contribution of NPO to LLM-based decision modeling, especially how NPO differs from existing, lower-cost approaches. Second, we sincerely hope, in light of these clarifications, you may reconsider the first point you proposed. Evaluation against pooled baselines is currently a common practice. But we are confident that niche persona can have clear better performance gains and insights that benefits to specified groups. And we provide our thoughts as follows:

Critical challenges of personalized decision modeling. Current approaches face different challenges in incorporating contextual information, specifying behavioral models, and validating decision rules:

  • Machine learning models make it difficult to turn contextual information into an explicit account of how people choose. Textual context must first be encoded as model inputs, and high predictive accuracy can coexist with less reliable estimates of behavioral indicators [1].
  • Classic utility-based theories require analysts to specify a behavioral structure in advance. For each application, analysts must decide which variables to include and how they enter the utility functions, a process that can require substantial modeling effort [2].
  • Recent LLM frameworks leave open which decision rule governs their predictions. These frameworks can revise their reasoning over multiple rounds [3], and identifying an explicit rule that can be tested against observed choices remains a question.

Motivation of NPO. These challenges motivate NPO: a framework that uses observations from a target group to identify both the information relevant to its decisions and a behavioral theory that specifies how that information enters the comparison of alternatives. Compared to classic theories, the decision niche is flexible because both the selected features and the behavioral theory can vary across groups. Compared to ML methods, it is more interpretable because the fitted theory explicitly specifies how those features enter the comparison of alternatives. Compared to LLM frameworks that generate choice predictions directly, NPO constrains the LLM’s reasoning with candidate behavioral theories and gives inspectable checkpoints.

Figure 3 illustrates how NPO adapts its search across tasks and groups. For instance, the MAJ run can provide cases of group-level predictive gains. NPO with DeepSeek selected random regret minimization (RRM) as the final theory for all three groups and achieved higher test weighted F1 than each pooled ML baseline. The gains over the strongest pooled baseline were 0.126 for the young, high-income group, 0.111 for the middle-aged, low-income group, and 0.058 for the middle-aged, middle-income group. The young, high-income group’s niche included additional details about working in the previous week and having a job or business, whereas both middle-aged groups retained a more compact set covering general employment status and marital status.

MAJ groupFinal theoryNPOPooled LRPooled RFPooled XGBGain over best pooled
Young, high incomeRRM0.5810.2780.4550.432+0.126
Middle-aged, low incomeRRM0.5260.4150.3810.359+0.111
Middle-aged, middle incomeRRM0.6410.3630.4340.583+0.058

Innovation and contribution of NPO. The innovation of NPO mainly lies in the co-evolution of theories, features, and decision reasoning. It demonstrates how an LLM can use a limited set of observations from a target group to construct an explicit decision niche containing a group-level theory-feature configuration. This capability bridges well-established human behavioral theories with generalizable LLM-driven reasoning. It is particularly valuable for personalized decision modeling when collecting additional information about a group requires additional effort and expense [4].

[1] Martín-Baos, José Ángel, et al. "A prediction and behavioural analysis of machine learning methods for modelling travel mode choice." Transportation research part C: emerging technologies 156 (2023): 104318.

[2] Ortelli, Nicola, et al. "Assisted specification of discrete choice models." Journal of Choice Modelling 39 (2021): 100285.

[3] Zhao, Yibo, et al. "Personalized Decision Modeling: Utility Optimization or Textualized-Symbolic Reasoning." Advances in Neural Information Processing Systems 38 (2025).

[4] Sahlqvist, Shannon, et al. "Effect of questionnaire length, personalisation and reminder type on response rate to a complex postal survey: randomised controlled trial." BMC Medical Research Methodology 11 (2011): 62.

Reviewer_TJe3OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes the Niche Persona from Observation (NPO) framework. Inspired by the concept of ecological niche differentiation, NPO investigates how to jointly identify relevant information features and appropriate behavioral decision theories for different population groups, thereby improving the ability of large language models to predict human behavior in a personalized manner.

The main contribution lies in introducing the concept of a Decision Niche, defined by the alignment among specific population groups, information features, and behavioral theories. NPO employs behavioral theories to guide LLM-based decision agents and iteratively refines feature selection through validation feedback and error analysis, enabling the mutual adaptation of information selection and decision modeling. NPO integrates feature selection, behavioral decision theories, and population heterogeneity modeling into a unified LLM-based multi-agent optimization process.

Experimental results demonstrate that NPO achieves improved behavioral prediction performance while using fewer explicit input features. Its primary innovation lies in the joint adaptation of information features and decision models guided by behavioral theories, rather than the development of a fundamentally new decision theory.

The study provides empirical evidence that jointly adapting information features and behavioral models to heterogeneous populations can improve both predictive performance and feature efficiency.

Critical strengths

The paper addresses an important and thought-provoking research problem. The authors recognize the interdependence between information selection and decision-making mechanisms, thereby moving beyond the conventional approach of treating feature selection and decision modeling as separate processes. The concept of Decision Niche provides a valuable conceptual framework for modeling the behavior of heterogeneous populations. NPO integrates behavioral theory-guided LLM reasoning, feature selection, validation feedback, and group-specific model selection into an iterative optimization framework, achieving competitive experimental results.

Critical weaknesses

The main weakness lies in the lack of theoretical rigor and formal mathematical foundations. The concept of Decision Niche still requires a more rigorous mathematical definition, explicit measures of information–theory coupling, and a theoretical characterization of stability. Moreover, the proposed Co-evolution mechanism is primarily realized through iterative feature augmentation and model competition, leaving considerable room for further strengthening its theoretical foundations and depth.

Questions for authors

Suggestions for future work: I recommend that the authors evaluate the framework’s ability to recover underlying behavioral theories and relevant features using synthetic data with known ground-truth behavioral mechanisms. Authors could investigate the stability of decision niches across different samples, population groups, and distribution shifts.

2 comments
AuthorsOpenReview ↗

Response to Reviewer TJe3

We sincerely appreciate the reviewer for encouraging us to clarify the theoretical foundations of decision niches and examine their stability across runs.

W1 Mathematical definition and theoretical foundations

We thank the reviewer for raising concerns regarding mathematical definition and theoretical foundations, and would like to clarify from the following perspectives:

Mathematical definition. Equation (3) in Section 3.2 defines the group-level decision niche state as ηt=(St,mt)\eta_t=(S_t,m_t). This state represents the feature–theory configuration at each iteration. This follows a standard state-space formulation of stochastic search, in which successive configurations form a trajectory through a discrete configuration space. Previous LLM-based frameworks also use evaluation feedback to guide subsequent proposals [1,2].

Theory-information coupling. The theory-information coupling can be analyzed from the perspectives of both theory and features. In Figure 3(a), RRM accounts for 32 of the 54 final selected features on MAJ. During feature selection, the young, high-income group added several employment variables. This feature pattern is consistent with evidence that work demands and social roles can shape cannabis use and decisions to reduce it [3,4].

Stability analysis. We compare the final niches from two runs. Feature similarities are averaged across nine groups. A group counts toward top-2 alignment if either run’s selected theory ranks first or second in the other run. The results show partial repeatability in the process of decision niche discovery. But we acknowledge that these observations do not establish stable recovery for MAJ. One possible explanation for the limited theory agreement on MAJ is that respondents within the same demographic group may follow different decision strategies, making it difficult for a single behavioral theory to characterize the group. We will add this discussion to the revised version.

DatasetTheory Agreement RateTop-2 Theory Alignment RateFeature Dice SimilarityAdded-Feature Overlap Coefficient
JOB4/98/976.0%80.6%
ALC6/98/979.7%80.6%
MAJ2/96/980.1%50.0%

Further theoretical foundations. We agree that the co-evolution perspective leaves room for deeper theoretical development. Our current formulation provides an instantiation within discrete choice tasks, where information is represented by selected features. Extending this perspective to broader tasks and decision mechanisms calls for a more general formulation of information–theory representation. We will discuss the scope of the current formulation and possible theoretical extensions in the revised version.

Q1 Future work regarding underlying theories and features

We thank the reviewer for this constructive suggestion. We agree that controlled synthetic data with known generating mechanisms would complement our predictive evaluation by testing theory and feature recovery. However, designing a behaviorally meaningful synthetic dataset needs a dedicated study, given recent evidence that LLM-simulated economic preferences agree only modestly with individual human responses [5]. One approach would be to sample choices from a prespecified discrete-choice model with a known set of relevant features and use an LLM to generate alternative descriptions that preserve the options and attribute values. We will outline this direction in the revision as future work.

[1] Kim, Doyoung, Jaehyeok Doo, and Minjoon Seo. "TSLM: Tree-Structured Language Modeling for Divergent Thinking." arXiv preprint arXiv:2601.22688 (2026).

[2] Gong, Junfeng, et al. "From large to small: Transferring cuda optimization expertise via reasoning graph." International Conference on Learning Representations. Vol. 2026. 2026.

[3] Fergusson, David M., and Joseph M. Boden. "Cannabis use and later life outcomes." Addiction 103.6 (2008): 969-976.

[4] Yamaguchi, Kazuo, and Denise B. Kandel. "On the resolution of role incompatibility: A life event history analysis of family roles and marijuana use." American journal of Sociology 90.6 (1985): 1284-1325.

[5] Chen, Fadong, et al. "Simulated responses from large language models are weak proxies for human economic preferences across 74 countries." Nature Human Behaviour (2026): 1-14.

AuthorsOpenReview ↗

Followup to Reviewer TJe3

Dear Reviewer TJe3,

Thank you for your valuable comments on the theoretical foundations of our work and the recovery of underlying behavioral patterns. These suggestions have helped us better clarify the scope of NPO and have provided constructive insights for future research on LLMs and personalized decision-making. We sincerely welcome any further discussion.

Best regards,

The Authors of Submission 11777

Comments on the paper

1 comment
AuthorsOpenReview ↗

Revision Summary

We thank the reviewers for their constructive feedback. In response, we made the following changes to the manuscript. For ease of reference, the additions are marked in blue in the revised PDF.

LocationReviewer concernRevision
Section 1tr5W-Q1, TJe3-W1Updated the summary of findings to include the new search and cost comparisons and clarify how NPO represents group-specific decision niches.
Section 4.6, Table 4U9iX-W1, U9iX-Q1, U9iX-Q2Added comparisons with LLM-AO and direct feature–theory search using Codex CLI and Claude Code.
Section 4.6, Table 5tr5W-W2Compared API cost and weighted F1 with full-feature Athena and DualMind.
Appendix B.4, Table 8TJe3-W1Compared theory agreement, top-2 alignment, and feature overlap across two runs.
Section 5, Appendix CTJe3-W1Discussed the scope of the discrete-choice formulation and extensions to broader decision settings.
Section 5TJe3-Q1Discussed synthetic benchmarks with known choice-generating mechanisms for testing theory and feature recovery.

We sincerely appreciate the reviewers’ suggestions, which helped us clarify the scope of the method and strengthen its empirical evaluation. And we are looking forward to further discussion.

Best regards,

Authors of Submission 11777

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.