acceptodds
Under review as a conference paper at ICLR 2027

When Decision Theories Empower AI Agents: Learning and Orchestrating Theory–Agent Alignment for Personalized Decision-Making

Abstract

Large language models (LLMs) hold great promise as personalized decision companions. However, using LLM agents for decision modeling also introduces new challenges: while LLMs offer richer and more flexible modeling capabilities, they often lack theoretical grounding, exhibit unstable reasoning behavior, and provide limited interpretability. This raises two central questions: how can AI agents be aligned with decision theories to support trustworthy reasoning, and how can the most appropriate theory-guided agents be selected or orchestrated for different tasks and personas? We address these questions through the lens of the wisdom of crowds: rather than searching for a single universal decision theory, we treat diverse decision theories as complementary behavioral experts. Building on this idea, we propose Learning to Agentic Orchestration (L2AO), a framework that orchestrates theory-guided AI agents for trustworthy and personalized decision modeling. L2AO consists of two key components: Theory-guided Agent Learning (TAL), which constructs and trains decision agents under different behavioral theories, and Consensus-First Expert Adaptation (CFEA), which coordinates learned agents through a democratic AI mechanism that prioritizes cross-theory consensus while adapting to task- and persona-specific differences. Experiments on five real-world survey-based decision datasets show that L2AO achieves state-of-the-art performance compared with utility-based, machine-learning, and LLM-based baselines, improving average accuracy by 15.0% over the strongest baseline. Beyond prediction, L2AO further reveals potential systematic biases that can arise when a single decision theory is embedded into an LLM agent to model diverse human decisions. The source code will be fully released upon paper acceptance.

Primary area:
neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)
Keywords:
Large Language Models, Neurosymbolic AI, Decision Theory, Human Decision-Making, Discrete Choice Modeling

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_bAGWOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

Although large language models (LLMs) offer flexible modeling capabilities for personalized decision-making, they lack a solid theoretical foundation, exhibit behavioral instability, and suffer from insufficient interpretability. This paper introduces the "Learning to Agentic Orchestration" (L2AO) framework. It employs "Theory-guided Agent Learning" (TAL) to train agents that make decisions based on various theories, and utilizes "Consensus-First Expert Adaptation" (CFEA) to coordinate the decision outputs of multiple agents, thereby balancing both the commonalities and the distinct differences among them. Experiments conducted on five decision-making datasets, derived from real-world survey data, demonstrate that L2AO achieves state-of-the-art performance, outperforming utility-based, machine learning-based, and existing LLM-based baseline methods. Furthermore, the paper reveals that relying on a single decision theory to guide agent behavior can introduce potential systematic biases.

Critical strengths
  1. L2AO integrates various decision theories with LLM agents, combining theory-specific numerical utility gradients with textual gradients to facilitate agent learning. Furthermore, it employs an ensemble of multiple agents to achieve additional performance gains.
  2. The paper conducts a comprehensive comparison of the proposed method against existing approaches across five decision-making datasets based on real-world surveys, demonstrating a clear advantage for L2AO.
Critical weaknesses
  1. There are issues with the citation information and formatting. When citing the paper "The rise and potential of large language model based agents: A survey," the authors included eight names that do not appear in the original source. Furthermore, ICLR 2027 requires an author-year citation format, whereas this paper uses a numbered citation format.
  2. In L2AO, the results derived from decision theory serve merely as inputs for the LLM, which then makes a final decision by synthesizing these theoretical results with textual knowledge and profiles. However, the paper does not analyze how the LLM resolves conflicts between theoretical results and textual knowledge, nor does it examine the extent to which the LLM relies on the theoretical results. Consequently, it remains difficult to determine the degree to which the theory actually guides the final decision-making process.
Questions for authors

Q1 Could the authors please correct the citation information and verify the citation format?

When citing the paper "The rise and potential of large language model based agents: A survey," the authors included eight non-existent names in the author list: Wei Xu, Junjie Ye, Guiyang Hou, Yan Liu, Bin Gui, Haoxiang Fen, Yongzu Cano, and Zhengando Wen. The original paper can be found here: https://arxiv.org/abs/2309.07864v3. Furthermore, ICLR 2027 requires an author-year citation format, yet this manuscript uses a numbered citation style. Adhering to the specified formatting guidelines is a fundamental requirement for submission. Please ensure these are corrected.

Q2 Could the authors analyze how an LLM makes decisions when there is a conflict between theoretical results and textual knowledge?

It is important to examine whether an LLM can appropriately weigh theoretical results against textual knowledge, and to what extent theoretical results actually guide the LLM’s decision-making process. This relates to the core assumption of this paper, namely, that decision theory can help LLMs make better decisions.

5 comments

We sincerely thank the reviewer for their time and thoughtful comments.

W1&Q1 Citation and formatting corrections.

We thank the reviewer for identifying this error and sincerely apologize for the inaccurate attribution.

We carefully reviewed the original paper and manually imported its citation metadata from Google Scholar. However, during a subsequent round of AI-assisted language polishing and formatting, the author list was inadvertently altered to include nonexistent authors. Because the reference contained a long author list and the incorrect names appeared in the middle, we unfortunately overlooked these changes during our final manual verification before submission.

The error was limited to the author list. The cited paper itself, including its title, publication information, and DOI, was correctly identified. We take full responsibility for this oversight and have corrected the reference metadata in the revised manuscript. We sincerely apologize for this error and appreciate the reviewer for bringing it to our attention.

W2&Q2 How does decision theory guide LLM predictions?

Decision theories guide both the learning process and final predictions through explicit behavioral structures. We clarify this through the optimization mechanism and an additional analysis of symbolic–LLM prediction consistency.

(a) Theory guide LLM in both the forward process and backward optimization.

In the forward process, each agent computes theory-conditioned choice probabilities from its learned utility function. These probabilities serve as explicit behavioral priors for LLM reasoning, alongside contextual knowledge and theory-specific assumptions. In the backward process, prediction errors guide updates to both symbolic utilities and contextual knowledge. Theory-specific constraints govern these revisions, while numerical optimization and validation-based evaluation regulate utility updates. This ensures that each agent retains its theory-specific choice formulation throughout iterative optimization.

The optimized agents preserve distinct theoretical structures across demographic groups. For example, on SWM, Group 2 (age < 39, income < 50k) learns linear time–cost utilities under MNL, with a car-availability coefficient of 2.42. MXL instead introduces nonlinear cost transformations and trip-purpose heterogeneity (SD = 0.41). In Group 9 (age > 65, income 50–100k), MXL learns demographic interactions and greater trip-purpose heterogeneity (SD = 3.80). PT retains reference-dependent valuation (α = 0.54, β = 0.44, λ = 1.60), with opposite travel-pass effects for Train (+2.75) and Swissmetro (−2.84). PCL learns logarithmic fares and mode-specific attributes. Theory-specific utility structures remain distinguishable after optimization.

(b) Interaction between symbolic predictions and LLM reasoning.

LLM decisions primarily follow theory-guided predictions while retaining flexibility for contextual adjustments. We examine cases where symbolic utility predictions conflict with contextual predictions across all four theory-guided agents in following table. Such conflicts occur in 66.2% of cases. The final LLM follows the theory in 71.0% of these cases, compared with 19.8% following the context and 9.2% selecting another alternative. Theory-following rates exceed 66% for all four theories.

Table: Conflict resolution between symbolic and contextual predictions. Conflict rate denotes the proportion of cases where the two predictions differ. Follow Theory, Follow Context, and Other are calculated among conflict cases.

TheoryConflictFollow TheoryFollow ContextOther
MNL64.2%76.8%16.5%6.8%
MXL68.4%66.8%22.4%10.8%
PT68.0%73.0%16.5%10.5%
PCL64.1%67.6%23.9%8.5%
Overall66.2%71.0%19.8%9.2%

We further compare symbolic and final LLM prediction accuracy in the following table. Across five datasets, symbolic predictions achieve 57.4% accuracy, which increases to 63.3% after LLM reasoning. When symbolic predictions are correct, the LLM preserves the correct decision in 91.3% of cases. When symbolic predictions are incorrect, the LLM corrects 25.6% of these errors. Beneficial corrections (10.9%) substantially exceed harmful changes (5.0%), resulting in a net accuracy gain of 5.9 percentage points.

Table: Symbolic–LLM prediction consistency. CC/CW/WC/WW denote correctness combinations; net gain is LLM accuracy minus symbolic accuracy (percentage points).

TaskCCCWWCWWNet gain
ALC36.4%6.3%18.1%39.2%+11.8
JOB55.0%4.1%11.7%29.2%+7.6
MAJ50.6%6.4%10.9%32.1%+4.5
SWM54.9%6.0%10.7%28.4%+4.7
VAC64.9%2.3%2.9%29.9%+0.6
Overall52.4%5.0%10.9%31.7%+5.9

These findings reveal how symbolic and contextual reasoning complement each other in L2AO. Theory-derived predictions provide a behavioral foundation that is largely preserved by the LLM, while contextual reasoning allows the model to revise symbolic predictions when additional information supports a different choice. As a result, Theory guidance informs final predictions without rigidly determining every choice, providing measurable theoretical grounding while retaining the flexibility of LLM reasoning.

These analyses are now included in Section 4.3 of the revised manuscript.

We would like to kindly follow up on our rebuttal and the revised manuscript. We have addressed both concerns regarding citation accuracy and the role of decision theory in LLM decision-making.

We would greatly appreciate your feedback on whether these revisions have adequately addressed your concerns. If any questions remain, we would be happy to clarify or discuss them further. Thank you again for your time and valuable feedback.

Reviewer_3Kv5OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes Learning to Agentic Orchestration (L2AO) for predicting observed human choices from individual attributes and context. It treats multiple different behavioural decision theories as complementary experts. Theory-guided Agent Learning (TAL) constructs agents conditioned on MNL, MXL, prospect theory, and paired combinatorial logit. Each agent jointly learns a symbolic choice model and a textual reasoning component . Consensus-First Expert Adaptation (CFEA) aggregates predictions using consensus when agents agree and, when agents disagree, uses group-theory alignment and selects the assumption with the profile that best matches the group proxy.

Evaluation covers five survey datasets: SWM (Swiss metro travel mode choice), VAC (COVID 19 vaccine uptake), JOB (job choice during COVID 19), and MAJ/ALC (NSDUH marijuana and alcohol use recency). Each has about 900–1,000 samples, three classes, and nine age × income groups. Table 1 reports mean weighted F1 of 0.681 and accuracy of 0.692, compared with 0.607 and 0.602 for DualMind across all 5 datasets. Ablations, alternative aggregation rules, and selected split/group-boundary sensitivity analyses support the usefulness of the pipeline.

Critical strengths
  1. The paper presents a novel idea with consistent gains over the closest prior method under a matched LLM setup. Using several classical choice models as competing structured priors for an LLM predictor, and treating their disagreement as a signal, is a sensible extension of DualMind/ATHENA-style hybrids. L2AO beats DualMind on all five datasets with the same forward and backward model.
  1. The paper also provides ablations and evidence validating its pipeline:
  • removing TAL costs about 0.1 F1
  • removing any single theory lowers performance
  • L2AO's scores are stable across re-splits
Critical weaknesses

The paper does not show that the gain comes from theory diversity rather than plain ensembling, and the evaluation is too small and too weakly baselined to support the comparisons.

Most of the gain over DualMind is already achieved by plain majority voting. Each single-theory agent in L2AO is essentially DualMind with a different choice map plugged in. The reported 11.1–15.4% gap over single agents implies they score roughly 0.59–0.61 F1, which is about where DualMind sits (0.607). Plain hard voting over the four agents already reaches 0.656 (Table 6), so simple voting accounts for about two-thirds of L2AO’s improvement over DualMind. The rest of CFEA adds comparatively little, improving on hard voting by 0.025 F1, on validation weighting by 0.011, and on MoE-soft by only 0.001.

The experiments do not isolate the contribution of theory diversity. The paper never compares against an equal-size ensemble of agents built on the same theory, for example four MNL-TAL agents from independent runs. Without it, the paper’s central thesis remains untested. The thesis that orchestrating different theories is what drives the improvement, as opposed to the generic variance reduction any ensemble provides.

The statistical power is low. Test sets contain only 205–283 cases per dataset, or roughly 30 per demographic group. Table 1 reports no variance, and the baselines appear to be single runs. The component ablations in Table 5 differ from the full model by only 0.003–0.012 F1, yet the paper describes them as “consistent” drops. Table 9 also shows how noisy single-split numbers are: re-splitting moves PT’s F1 on SWM from 0.464 to 0.638.

Questions for authors
  1. What do you get from 4 independently trained single-theory agents (e.g., 4×MNL, or 4 DualMind runs) passed through the same hard-vote and CFEA pipelines? This experiment would help to distinguish the contribution of ensembling from theory orchestration.

  2. Please report the standard deviation over the 3 runs for every Table 1 entry, multiple runs for the LLM baselines, and paired significance tests for L2AO vs DualMind and for CFEA vs hard voting and validation weighting.

  3. What are the per-group train/validation/test sizes? How were τ=0.36\tau=0.36, λ=0.4\lambda=0.4, the hand-set zaz_a vectors and MAX_EDITS chosen?

  4. For each theory agent, how often does the final LLM prediction disagree with the argmax of its own πx,a\pi_{x,a}? How different are the learned utility forms across theories after the textual updates? This is needed to support the claim that agents remain consistent with their theoretical foundation.

5 comments

We thank the reviewer for raising significant questions.

W1 Methodological contribution beyond DualMind and generic ensembling.

L2AO integrates theoretical diversity into the reasoning and learning processes of AI agents. Different decision theories make distinct assumptions about how individuals evaluate alternatives, construct utilities, and make choices. This diversity provides complementary behavioral perspectives, with different theories better suited to different decision tasks and population groups. While DualMind establishes a foundation for integrating a single decision theory into an AI agent, L2AO advances this paradigm by enabling agents to learn under multiple theoretical assumptions and adaptively orchestrating their complementary strengths. Rather than relying on a predefined theory or simply aggregating predictions, L2AO learns theory-specific decision mechanisms and identifies which theoretical perspectives are most appropriate for different decision contexts.

CFEA further improves prediction by coordinating theory-guided agents, particularly when their predictions disagree. Hard voting across theory-guided agents already achieves an F1 of 0.656, compared with 0.607 for DualMind, demonstrating the benefits of theoretical diversity. CFEA further improves F1 to 0.681 by combining consensus-first aggregation with group–theory alignment. Among 238 non-consensus cases, 134 are recoverable because at least one agent makes the correct prediction. CFEA correctly resolves 76 of these 134 cases (56.72%), outperforming stacking (52.99%), MoE-soft (44.78%), and hard voting (28.36%).

W2&Q1 Theory diversity beyond repeated ensembling.

Heterogeneous theories provide broader cross-task adaptability than repeated single-theory ensembles. To distinguish the benefits of theory diversity from generic ensembling, we compare ensembles of four independently trained agents under the same theory (MNL, PT, PCL, or MXL) with an ensemble of four agents guided by different theories, across JOB, MAJ, and VAC. All configurations use identical training and aggregation settings, differing only in their theoretical assumptions. These additional results are included in Appendix D.3 of the revised manuscript.

Table: Same-theory vs. heterogeneous ensembles. Weighted F1 under identical aggregation mechanisms.

TaskAggregationMNL×4PT×4PCL×4MXL×4Four theories
MAJHard vote0.6310.6020.6050.6730.683
MAJCFEA0.6560.6160.6100.6910.719
JOBHard vote0.6830.6290.6810.7480.760
JOBCFEA0.7040.6470.6960.7610.774
VACHard vote0.6310.6440.6310.6380.655
VACCFEA0.6280.6780.6540.6350.669
MeanHard vote0.6480.6250.6390.6860.699
MeanCFEA0.6630.6470.6530.6960.721

Across three datasets, heterogeneous theories achieve the highest average F1 under both hard voting (0.699 vs. 0.686) and CFEA (0.721 vs. 0.696), compared with the strongest same-theory ensemble. While PT×4 slightly outperforms the heterogeneous ensemble on VAC (0.678 vs. 0.669), this illustrates that individual theories can have advantages in specific decision contexts. Such task-dependent differences reinforce our motivation: no single theory is expected to dominate universally, whereas heterogeneous theories offer broader cross-task adaptability.

Q3 Parameter settings and optimization protocol.

We clarify the parameter choices and provide complete per-group data splits. The consensus threshold τ = 0.36 is fixed across datasets to require meaningful cross-theory agreement (Appendix A.7). The theory-derived profiles za are manually initialized from the behavioral assumptions of MNL, MXL, PT, and PCL; their exact vectors and rationales are provided in Appendix A.8.3. We set λ = 0.4 to retain these theory-based priors while assigning greater weight (60%) to validation-derived empirical profiles (Appendix A.8.4). These settings are fixed at the beginning.

TAL uses 16 optimization iterations with validation-F1 checkpointing and rollback (Appendix A.6). We set MAX_EDITS=3 to limit each iteration to at most three knowledge revisions, preventing large updates that could destabilize the learned contextual representation. We additionally provide the exact train/validation/test counts for all 45 demographic groups in Appendix B.4, using the same splits across L2AO and all baselines.

Q4 Theory-specific structure and prediction consistency.

Optimized agents retain distinct theory-specific utility structures, with adaptations reflecting both behavioral assumptions and demographic differences. We illustrate this using two SWM groups: Group 2 (age < 39, income < 50k) and Group 9 (age > 65, income 50–100k).

In Group 2, MNL learns linear time–cost utilities with a strong car-availability coefficient (2.42), while MXL introduces nonlinear cost transformations and modest random-coefficient heterogeneity (trip-purpose SD = 0.41). PT instead adapts reference-dependent valuation and loss aversion, whereas PCL retains its initial systematic utilities.

In Group 9, MXL learns demographic interactions and greater trip-purpose heterogeneity (SD = 3.80). PT assigns opposite travel-pass effects to Train (+2.75) and Swissmetro (−2.84), with α = 0.54, β = 0.44, and λ = 1.60. PCL differentiates alternatives through logarithmic fares and mode-specific attributes.

Theory-specific differences also extend to contextual knowledge. For Group 2, MNL, MXL, and PCL express cost–income information as threshold rules, whereas PT frames it as gains and losses relative to Train. Recursive optimization preserves distinct behavioral representations.

We further examine how symbolic predictions relate to final LLM decisions. For each theory-guided agent, we compare its symbolic utility prediction with the contextual prediction and examine how the final LLM resolves disagreements. Across all four theories, symbolic and contextual predictions conflict in 66.2% of cases. Among these cases, the final LLM follows the theory in 71.0% of cases, compared with 19.8% following the context. This pattern is consistent across theories, with theory-following rates ranging from 66.8% to 76.8%.

Table: Conflict resolution between symbolic and contextual predictions. Conflict rate denotes the proportion of cases where the two predictions differ. Follow Theory, Follow Context, and Other are calculated among conflict cases.

TheoryConflictFollow TheoryFollow ContextOther
MNL64.2%76.8%16.5%6.8%
MXL68.4%66.8%22.4%10.8%
PT68.0%73.0%16.5%10.5%
PCL64.1%67.6%23.9%8.5%
Overall66.2%71.0%19.8%9.2%

We also examine whether disagreements improve or reduce prediction accuracy. The LLM preserves 91.3% of correct symbolic predictions while correcting 25.6% of symbolic errors. Beneficial corrections (10.9%) exceed harmful changes (5.0%), improving accuracy from 57.4% (symbolic) to 63.3% (LLM). These findings demonstrate that theory-guided symbolic reasoning provides a strong foundation for LLM predictions, while contextual reasoning offers complementary improvements.

Both analyses are now included in Section 4.3 of the revised manuscript.

W3&Q2 Statistical robustness and run-to-run variability.

Thank you for your helpful comments. We have added error bars and further clarified our experimental settings and implementation details.

We conducted three independent full-pipeline runs and now report mean ± standard deviation across all five datasets. Under identical experimental settings and fixed data splits, L2AO achieves an average weighted F1 of 0.681 across five tasks. The detailed results are summarized below. The standard deviations range from 1.2% to 4.7% of the corresponding mean F1 scores, remaining below 3% for four of the five datasets.

Table: Performance stability of L2AO across three independent runs. Weighted F1 is reported as mean ± standard deviation (SD).

DatasetWeighted F1 (Mean ± SD)Relative SD (%)
ALC0.552 ± 0.0101.8
JOB0.774 ± 0.0212.7
MAJ0.719 ± 0.0344.7
SWM0.694 ± 0.0081.2
VAC0.669 ± 0.0101.5
Average0.682 ± 0.0172.4

We further report run-level variability for DualMind and LLM-V. The original manuscript already reports DualMind's mean performance over three runs; we now provide the corresponding standard deviations. For LLM-V, we conducted three additional runs and report the mean and standard deviation. All results are summarized below.

Table: Performance variability of LLM-based baselines. Weighted F1 is reported as mean ± SD (%).

DatasetDualMind F1LLM-V F1
ALC46.90 ± 1.3227.40 ± 1.08
JOB64.50 ± 1.8544.90 ± 1.87
MAJ63.30 ± 1.3423.96 ± 3.45
SWM67.70 ± 1.5244.41 ± 2.60
VAC61.00 ± 0.7539.26 ± 5.32
Average60.6835.99

Paired significance tests support the performance differences against DualMind and hard voting. We use 10,000 paired randomizations for weighted F1 and exact McNemar tests for accuracy, comparing predictions on matched test examples.

Table: Paired significance tests. Two-sided p-values from the supplemental matched-prediction runs.

ComparisonMetricTestTwo-sided p
L2AO vs. DualMindWeighted F1Paired randomization (10,000)< 0.0001
L2AO vs. DualMindAccuracyExact McNemar1.2 × 10⁻¹¹
L2AO vs. hard votingWeighted F1Paired randomization (10,000)< 0.0001
L2AO vs. hard votingAccuracyExact McNemar8.7 × 10⁻⁵

Our evaluation scale is consistent with established practices in LLM research. For example, TextGrad [1] evaluates comparably sized benchmarks, including GPQA Diamond (198 questions) and MMLU subsets (92–112 questions). Collecting individual-level behavioral observations with demographic information introduces additional challenges. We evaluate L2AO across five distinct decision tasks and conduct repeated full-pipeline experiments to strengthen the empirical evidence. Developing larger and more diverse behavioral benchmarks remains a valuable direction for the research community.

The component ablations further demonstrate the complementary contributions of CFEA components. As shown in paper, removing any individual component reduces the average F1 by 0.003–0.012. While each component contributes incrementally, their combined effect enables CFEA to achieve its strongest performance. These findings reinforce our central motivation: integrating complementary theoretical perspectives into the reasoning and learning processes of LLM agents to better model heterogeneous human decisions.

We further clarify the data-split sensitivity results in Table 9. The PT results correspond to the standalone Prospect Theory baseline, rather than our PT-conditioned LLM agent. Although PT exhibits substantial sensitivity to different data splits on SWM (0.464–0.638), L2AO maintains more stable performance (0.676–0.694). Across the three splits, L2AO achieves 0.685±0.009 on SWM, compared with 0.540 ± 0.089 for PT. A similar pattern is observed on VAC, where L2AO achieves 0.657 ± 0.025, compared with 0.595 ± 0.042 for PT. These results demonstrate that L2AO maintains its performance advantage across independently regenerated data splits.

References

  • [1] Mert Yuksekgonul, Federico Bianchi, Joseph Boen, Sheng Liu, Pan Lu, Zhi Huang, Carlos Guestrin, and James Zou. Optimizing generative ai by backpropagating language model feedback. Nature, 639(8055):609–616, 2025.
Reviewer_QuBaOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The authors propose L2AO, a framework for training agents for decision modeling. They demonstrate their techniques on 5 survey datasets and compare them to various baselines. Their main contributions are (1) TAL – a method for agent learning which utilizes symbolic modeling (a learned parametric model) and a learned textual prompt, which are both components of the agent that produce a decision prediction; (2) CFEA – a way to coordinate these agents and to take a decision when they agree or disagree.

Their method improves over all baselines and gives us a new interpretability tool that partially allows us to measure the correlation of different decision models with groups and benchmarks.

Critical strengths

(1) Using symbolic choice modeling is a crucial part of choice models when you want the results to be calibrated with real data.

(2) I loved the real-world datasets and showing that it can help us better model decision making.

(3) The authors used a variety of baselines.

(4) The paper is well written and explained carefully.

(5) They experiment on 5 datasets, provide baselines, and did 3 runs for each task.

(6) The paper is reproducible: the authors say they will publish the source code (though it could be shared via an anonymous repo in the meantime). They also provide many details in the appendix.

Critical weaknesses

(1) Figure 2 is a bit dense, and the flow of the method is not so clear.

(2) I am aware of high API costs, but GPT-4o-mini is a relatively weak and old model - demonstrating the results with a more recent reasoning/frontier model will amplify the value of your method. If you would stick with a similar capability model, I would choose an open source version.

(3) They didn’t provide error bars for their results.

Questions for authors

(1) Adding benchmarks and baselines from well-established, competitive domains such as economics and behavioral games would strengthen your results.

(2) Why have MNL, MXL, PT, and PCL been chosen over other models?

4 comments

We thank the reviewer for the constructive comments.

W1 Clarity of Figure 2.

We thank the reviewer for this helpful suggestion. We simplify the presentation of Figure 2 to better illustrate the overall workflow and key interactions between components in the revised manuscript.

W2 Evaluation with alternative LLMs.

We additionally evaluated L2AO with Llama 4 Scout, Gemini 4 Argon and Gemma 4-26B as alternative forward models. As shown in the table below, L2AO demonstrates consistent performance across different model backbones, with all configurations outperforming every machine-learning and utility-based baseline. We have incorporated these results into the main comparison table and added a cross-LLM analysis in Section 4.2 of the revised manuscript.

Table: Cross-LLM evaluation. Weighted F1 across five datasets.

DatasetMLPPCLDualMindGPT-4o-miniLlama-4-ScoutGemini-4-ArgonGemma-4-26B
ALC0.4180.4000.4690.5520.5370.5260.527
JOB0.4320.5620.6450.7740.7150.7360.725
MAJ0.5770.4650.6330.7180.6690.6580.666
SWM0.6170.5170.6770.6940.7190.6990.669
VAC0.5970.6050.6100.6690.6700.7430.714
Mean0.5280.5100.6070.6810.6620.6720.660

These results highlight the value of L2AO’s framework design beyond the capability of the LLM backbones. The original GPT-4o-mini configuration achieves the highest average F1 (0.681), compared with 0.662 for Llama 4 Scout, 0.672 for Gemini 4 Argon and 0.660 for Gemma 4 26B. This suggests that stronger or newer forward models do not necessarily yield better decision predictions within the same framework. Through theory-guided learning and adaptive orchestration, L2AO enables a relatively lightweight model to achieve competitive performance alongside more recent LLMs. Meanwhile, all three configurations outperform the strongest utility-based and machine-learning baselines on average, further supporting the effectiveness and flexibility of our approach.

W3 Statistical reliability and error bars.

Thank you for this suggestion. We have added results from repeated experiments, including the corresponding error bars, to Table 1. The performance stability of the complete TAL–CFEA pipeline across five datasets is summarized below. The relative standard deviations range from 1.2% to 4.7%. These results demonstrate the stability of L2AO across repeated runs. We additionally report the mean ± standard deviation for both weighted F1 and accuracy in the Appendix E.

Table: Performance stability of L2AO across repeated runs.

DatasetWeighted F1 (Mean ± SD)Relative SD (%)
ALC0.552 ± 0.0101.8
JOB0.774 ± 0.0212.7
MAJ0.719 ± 0.0344.7
SWM0.694 ± 0.0081.2
VAC0.669 ± 0.0101.5
Average0.682 ± 0.0172.4

Q1 Additional economics and behavioral-game benchmarks.

Thank you for the constructive suggestion. We have expanded our baseline comparisons as requested, and the additional results further demonstrate the consistent superiority of L2AO.

We have already included six commonly used baselines in our original manuscript. In our manuscript, we have already inlcuded six established choice-model baselines (MNL, MXL, PCL, PT, NL, and RRM). All those baselines are widely adopted as the common baseline for human decision-making and behavior modeling [1][2]. Our model show clear improvements by 33.5% over the strongest baseline across tasks.

We additionally evaluated APM as a new baseline. In response to the reviewer’s suggestion, we additionally evaluated the Attribute-based Processing Model (APM), an established approach to modeling attribute processing in behavioral choice [3]. Across five datasets, L2AO achieves an average weighted F1 of 0.681, compared with 0.491 for APM and 0.510 for PCL, the strongest original choice-model baseline. The additional results are summarized below.

Table: Additional APM baseline compared with L2AO. Weighted F1 and accuracy across five datasets.

TaskAPM F1APM AccL2AO F1L2AO Acc
ALC0.3850.3820.5520.590
JOB0.4930.5250.7740.762
MAJ0.4720.4300.7180.762
SWM0.5000.5300.6940.695
VAC0.6070.6150.6690.649
Mean0.4910.4960.6810.692

Five real-world survey-based benchmarks have been evaluated. The five datasets cover transportation (SWM), employment (JOB), vaccination (VAC), and substance-use decisions (MAJ and ALC), drawing on widely used, publicly available survey datasets. L2AO consistently outperforms competing methods across all five benchmarks.

We are extending our evaluation to additional benchmarks. We are currently working on additional behavioral decision-making benchmarks, including CPC18 and choices13k. Each benchmark requires training four theory-guided agents (MNL, MXL, PT, and PCL) across multiple population groups and LLM backbones, together with competing machine-learning, utility-based, and LLM-based baselines. A complete evaluation is expected to require more than two weeks of experimental work and approximately 1.2–1.8 billion LLM tokens for theory-guided agent training alone, excluding additional backbones and baselines. We are actively conducting these experiments.

Q2 Rationale for selecting MNL, MXL, PT, and PCL.

We thank the reviewer for these constructive suggestions.

The four theories form a behaviorally diverse expert pool for studying theory-guided learning and orchestration. MNL, MXL, PT, and PCL are well-established theories widely used in economics and behavioral choice modeling. Each captures distinct behavioral assumptions and decision mechanisms: MNL provides a foundational random-utility framework; MXL captures preference heterogeneity; PT models reference dependence and loss aversion; and PCL captures pairwise substitution patterns. These complementary perspectives motivate our central research question: How can AI agents learn from diverse decision theories and adaptively orchestrate them to model heterogeneous human decisions across tasks and populations?

L2AO provides an open and extensible framework for integrating decision theories. Our framework supports two key capabilities: (1) integrating individual decision theories into theory-conditioned AI agents for decision modeling, and (2) adaptively orchestrating complementary theory-guided agents across diverse decision tasks and contexts. New decision theories can be readily incorporated as additional agents within the existing framework, enriching the expert pool and expanding its ability to capture heterogeneous decision mechanisms. We envision this extensibility as a foundation for community contributions, enabling researchers to introduce new theories and extend L2AO to broader behavioral settings and application domains.

References

  • [1] Gerardo Berbeglia, Agustín Garassino, and Gustavo Vulcano. A comparative empirical study of discrete choice models in retail operations. Management Science, 68(6):4005–4023, 2022.
  • [2] Shenhao Wang, Baichuan Mo, Yunhan Zheng, Stephane Hess, and Jinhua Zhao. Comparing hundreds of machine learning and discrete choice models for travel demand modeling: An empirical benchmark. Transportation Research Part B: Methodological, 190:103061, 2024.
  • [3] James R Bettman, Mary Frances Luce, and John W Payne. Constructive consumer choice processes. Journal of consumer research, 25(3):187–217, 1998.

Thank you for the thorough response. Your modifications and additional experiments are highly appreciated and will help strengthen the paper. However, the absence of frontier OpenAI’s and Anthropic’s models from the evaluation stands out. Could you explain why they were not included?

We sincerely thank the reviewer for this valuable suggestion. Due to computational resource constraints, we were unable to include these models in our initial evaluation. We have now additionally evaluated GPT-6.1 Sol and Claude Sonnet 5.5, replacing only the forward model while keeping other settings unchanged. Despite dataset-specific variations, both models achieve comparable overall performance (average F1: 0.681 and 0.666, respectively), suggesting that L2AO is relatively robust to LLM backbone selection and generalizes across different model families.

DatasetGPT-6.1 Sol F1AccClaude Sonnet 5.5 F1Acc
ALC0.5330.5270.5160.513
JOB0.7280.7250.6910.697
MAJ0.6950.7190.6650.671
SWM0.7120.7100.7220.722
VAC0.7370.7380.7370.738
Average0.6810.6840.6660.668

We would greatly appreciate any further feedback and welcome additional discussion.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.