acceptodds
Under review as a conference paper at ICLR 2027

GreedyBM: Greedy Utility Discovery for Automated Modeling of Human Strategic Behavior

12 views

Abstract

Predicting human strategic behavior requires models that capture systematic departures from individual payoff maximization. Random utility models address this need by separating systematic utility from stochastic influences, but specifying the systematic utility typically demands substantial domain expertise and extensive manual refinement. We introduce GreedyBM, a method that constructs systematic utilities greedily, adding one term at a time, with each term encoding an interpretable behavioral tendency. At each stage, GreedyBM prompts a large language model (LLM) to propose candidate terms based on the current systematic utility and on previously evaluated candidates, and then adds the highest-scoring one. We derive a computable lower bound on the training negative log-likelihood (NLL) reduction achieved in a single stage, together with a multi-stage bound that relates candidate quality to the attainable training NLL. Across three strategic scenarios (the Ultimatum Game, repeated rock–paper–scissors, and continuous double auctions), GreedyBM achieves a lower mean test NLL than both expert-designed models and models discovered by existing LLM-based methods, while consuming substantially fewer tokens during search.

Primary area:
applications to neuroscience & cognitive science
Keywords:
human decision-making, automated model discovery, large language model

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

The paper introduces GreedyBM, a method that uses an LLM to build models of human strategic choices. Across three decision-making tasks, it predicts choices more accurately than existing methods while using substantially fewer LLM tokens. The resulting models also capture recognizable behavioral patterns, making them easier to interpret.

I'm not confident and an expert in this area but had a detail look at the paper and made some comments from my opinion.

Critical strengths
  1. The authors have provided comprehensive results to show the capabilities of their method, such as experiments for weights and training NLLs.
  2. A good literature review has been done.
Critical weaknesses
  1. There is not clear definition of human strategic behavior and in which applications this is applicable.
  2. Just LLM-based baselines have been compared while other non-LLM baselines not working?
  3. The authors compared the model with expert-designed models while no information about these models included in main body (related work).
  4. definition for I^{train} on page 3 is not clear.
  5. Figure 1 is not reflected by the text in section 3. what is "Selected Term 𝑓𝑡" (in Fig. 1)? it's not in the text.
Questions for authors
  1. why Predicting human strategic behavior is essential? what makes this area interesting? Using examples may be helpful.
  2. Expert-designed models are from "Automated human strategic behavior modeling via large language models" 2026? This is not cited in the main body on page 6. Have you talked about these models in related work section?
  3. In Figure 1, 𝑞𝑉𝑇 (𝑦 ∣ 𝑥) in prediction step, in which equation is comupted?
  4. What is the measurement on y axis in Figs. 5-6? what is V1-V5? V is utility but it's confusing what is V1-V5 (seems to be parameter) and why V should be a paremeter? based on what the Equation on Page 9 shows, V1 is computed from other values but it's not clear for V2-5.
4 comments
AuthorsOpenReview ↗

Response to Reviewer NUVP: Motivation and problem clarification

Thank you for your valuable suggestions. Page numbers refer to the revised manuscript unless stated otherwise.

W1: Definition of human strategic behavior and applicable settings

There is not clear definition of human strategic behavior and in which applications this is applicable.

Human strategic behavior concerns choices in situations where the actions of two or more participants jointly determine their outcomes and payoffs. A bidder's outcome, for example, depends on both their own bid and competing bids. We model how a person chooses an action given the strategic context available at the time of choice (Section 3.1, p. 3). This formulation applies to auctions, bargaining, and market trading. We have added an explicit definition and an online bargaining example at the start of the Introduction (Section 1, p. 1).

Our experiments examine bargaining offers in the Ultimatum Game (UG), next actions in repeated rock-paper-scissors (RPS), and buyer bids in continuous double auctions (CDA). Section 4 (p. 6) describes the strategic context and predicted action for each task.

W2: Non-LLM baselines

Just LLM-based baselines have been compared while other non-LLM baselines not working?

The comparison in Table 1 (p. 7) includes nine non-LLM expert-designed behavioral models:

  • UG: quantal response equilibrium (QRE), inequity aversion (IA), and equity, reciprocity and competition (ERC).
  • RPS: subgame perfect Nash equilibrium (SPNE), regret matching (RM), fictitious play (FP), and win-stay, lose-shift (WSLS).
  • CDA: belief-based bidding (BB) and time-dependent bidding (TDB).

To make this distinction easier to identify, the revised Table 1 explicitly labels this group as “Expert-designed models (non-LLM).”

W3: Information about the expert-designed models in the main text

The authors compared the model with expert-designed models while no information about these models included in main body (related work).

These baselines express established hypotheses about preferences and learning. For example, IA and ERC capture how relative payoffs affect bargaining choices. RM and FP model adaptation to previous play. BB and TDB represent belief-based and time-dependent bidding. The “Human choice modeling” paragraph in Related Work (Section 2, p. 2) outlines these preference and learning hypotheses and connects them to the expert-designed baselines evaluated in Section 4. Table 1 reports their predictive performance, and Appendix B.2 (starting on p. 19) provides the individual specifications and original references.

W4: Definition of Itrain\mathcal I^{\mathrm{train}}

definition for ItrainI^{train} on page 3 is not clear.

Itrain\mathcal I^{\mathrm{train}} is the set of indices of the training records. Thus Dtrain=(xi,yi):i∈Itrain\mathcal D^{\mathrm{train}}=\\{(x_i,y_i):i\in\mathcal I^{\mathrm{train}}\\} and N=∣Itrain∣N=|\mathcal I^{\mathrm{train}}|. For example, if records 1, 3, 6, and 8 are used for training, then Itrain={1,3,6,8}\mathcal I^{\mathrm{train}}=\{1,3,6,8\} and N=4N=4. We have added an explicit definition immediately before the training-NLL objective in Section 3.1 (p. 3).

W5: Figure 1 and the selected term ftf_t

Figure 1 is not reflected by the text in section 3. what is "Selected Term ftf_t" (in Fig.1)? it's not in the text.

The selected term ftf_t is the highest-scoring candidate from the stage-tt candidate pool Pt\mathcal P_t. The “Select & Refit” paragraph in Section 3.3 (p. 5) describes this selection step, followed by the joint weight refitting in Eq. (8). Combining the selected terms with their refitted weights yields the updated systematic utility in Eq. (3). To make the connection between the diagram and these steps easier to follow, we have revised the Figure 1 caption (p. 4) to identify ftf_t explicitly and include the corresponding equation references.

Q1: Why predicting human strategic behavior matters

why Predicting human strategic behaviour is essential? what makes this area interesting? Using examples may be helpful.

On an online bargaining platform, buyers choose offers and sellers accept, reject, or counter them. Predicting these choices helps the platform assess how minimum offer thresholds or limits on bargaining rounds affect agreement rates and transaction prices. Human decisions are shaped by fairness concerns and past interactions, often departing from predictions based solely on individual payoff maximization. Predictive models capture these preferences and history dependence from interaction data. We have added this online bargaining example and its use in evaluating platform rules to the opening of the Introduction (Section 1, p. 1).

AuthorsOpenReview ↗

Response to Reviewer NUVP: Sources, figures, and manuscript changes

Q2: Source and citation of the expert-designed models

Expert-designed models are from "Automated human strategic behavior modeling via large language models" 2026? This is not cited in the main body on page 6. Have you talked about these models in related work section?

Yes, we follow the expert-designed baseline set compiled by Xie et al. (2026), “Automated human strategic behavior modeling via large language models,” which introduces AutoBM. The evaluation paragraph in Section 4 (p. 7) cites this paper as AutoBM (Xie et al., 2026). The underlying behavioral models themselves originate in the economics and game-theory literature, for example, inequity aversion from Fehr and Schmidt (1999), ERC from Bolton and Ockenfels (2000), and regret matching from Hart and Mas-Colell (2000). Related Work discusses these model families (p. 2), and Appendix B.2 gives the exact specifications and primary references used in our experiments.

Q3: Equation for the prediction step in Figure 1

In Figure 1, qVT(y∣x)q_{V_T}(y\mid x) in prediction step, in which equation is computed?

The prediction is computed by substituting the final systematic utility VTV_T from Eq. (3) into the logit choice rule in Eq. (1):

qVT(y∣x)=exp⁡(VT(x,y))∑a∈Yexp⁡(VT(x,a)).q_{V_T}(y\mid x)=\frac{\exp(V_T(x,y))}{\sum_{a\in\mathcal Y}\exp(V_T(x,a))}.

The revised Figure 1 caption (p. 4) now points directly to Eq. (1) for this prediction step.

Q4: Measurements in Figures 5–6 and the meaning of V1V_1–V5V_5

What is the measurement on y axis in Figs.5-6?

In Figure 5 (p. 10), the left panel measures the difference between repeat rates after wins and after losses (in percentage points), while the right panel shows the probability of choosing the unplayed action after playing two different actions (in percent). In Figure 6 (p. 10), the two panels show the probabilities of repeat bids (∣y−h∣<0.01|y-h|<0.01) and raised bids (y−h>0.01y-h>0.01), both expressed as percentages, where hh denotes the player's previous bid. Dashed lines indicate empirically observed values. We have updated the Figure 6 caption to state the percentage units and stage-model notation explicitly.

what is V1-V5? V is utility but it's confusing what is V1-V5 (seems to be parameter) and why V should be a parameter?

Along the horizontal axes, V1V_1–V5V_5 label the systematic utility functions obtained after stages one through five. They do not denote fitted scalar parameters. Specifically, Eq. (3) defines each VtV_t as a function of the strategic context and action, Vt(x,y)=∑s=1tλs(t)fs(x,y)V_t(x,y)=\sum_{s=1}^{t}\lambda_s^{(t)}f_s(x,y). Here, the weights λs(t)\lambda_s^{(t)} are the fitted scalar parameters, and the superscript (t)(t) reflects that previously selected terms can receive updated weights during joint refitting at later stages. Pairing VtV_t with the logit rule in Eq. (1) defines the stage-tt behavioral model.

based on what the Equation on Page 9 shows, V1 is computed from other values but it's not clear for V2-5.

For CDA, the displayed equation on p. 10 (p. 9 of the original submission) writes out the stage-1 systematic utility V1V_1 explicitly. Stages two through five each add one selected term and jointly refit all weights. Combining the term expressions in Appendix B.3 (pp. 23–24) with the stage-wise weights in Table 11 (p. 24) via Eq. (3) gives the explicit expressions for V1V_1 through V5V_5.

Ethics flag

Flag For Ethics Review: Yes

Our evaluation reuses previously released behavioral datasets from laboratory or classroom studies, as noted in the Ethics Statement (p. 11), and no new human-subject data were collected for this study. The models take as input recorded action choices together with task-level game or market variables, such as resource endowments, action histories, and bid/ask records (Section 4, p. 6). All primary data sources (Lin et al. (2020), Komai et al. (2022), and Ikica et al. (2023)) are cited in Section 4 and Appendix B.2. We would be glad to provide further clarification on any specific ethical concern.

Changes made in the revised manuscript

In response to your suggestions, we have:

  • Defined human strategic behavior and added an online bargaining example explaining its practical use in the opening of the Introduction (Section 1, p. 1).
  • Added an explicit definition of the training-index set Itrain\mathcal I^{\mathrm{train}} in Section 3.1 (p. 3).
  • Identified the selected term ftf_t and added equation references for joint refitting and prediction in the Figure 1 caption (p. 4).
  • Explicitly labeled the expert-designed baselines as non-LLM models in Table 1 (p. 7).
  • Identified VtV_t as the stage-tt model and specified the percentage units in the Figure 6 caption (p. 10).

Thank you again for your constructive feedback. We welcome any further questions.

Thanks for your clear response to my comments and questions. I've found these answers satisfactory and increase my score to 3.

We are pleased that our responses have addressed your concerns. Thank you for your time and for the discussion.

Reviewer_kZyxOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary
  • Specifying the systematic utility of random utility models for human strategic behavior normally requires extensive expert effort. The paper aims to automate this with LLMs while keeping models interpretable and the search efficient.
  • The authors propose GreedyBM which builds the utility greedily, one term per stage. An LLM proposes candidate terms, each candidate is scored by the gradient of the training NLL without refitting, and the best is added before all weights are refit.
  • Theorem 1 shows the squared selection score lower-bounds the single-stage decrease in training NLL. Theorem 2 bounds the gap to any weight-bounded reference model by 2C2/AT2C^2/A_T, where ATA_T accumulates the squared candidate-quality ratios ρt\rho_t.
  • On the Ultimatum Game, repeated RPS and continuous double auctions, GreedyBM achieves lower mean test NLL than AutoBM, BoxLM, FunSearch and expert-designed models. It uses 87.2–90.9% fewer search tokens than AutoBM, and stagewise analyses relate the selected terms to known behavioral patterns.
Critical strengths
  1. Principled problem decomposition. Fixing the additive RUM structure narrows each LLM proposal to a single term. This turns open-ended model search into gradient-boosting-style greedy approximation with an approximate oracle. It matters because it yields a transparent selection rule and an explicit dependence of fit on candidate quality ρtρ_t, which prior LLM-based discovery methods lack.
  2. Consistent gains at much lower search cost. Mean test NLL is lower than all baselines on all three tasks, and the advantage persists at matched term counts (Table 3). Token use is roughly 8–11× lower than AutoBM, and the clearest win is in the Ultimatum Game (5.121 vs. 5.369). This matters because it shows the decomposition improves not onlz efficiency but predictive fit as well.
  3. Moment matching enables behavioral diagnostics. Proposition 1 ensures that each selected term’s observed and predicted averages match after refitting. This makes stagewise behavioral checks (Figs. 5, 6, 9) principled rather than anecdotal.
Critical weaknesses
  1. Empirical margins and baselines are thin. In Table 1, GreedyBM and AutoBM are often overlapping with /pm2∗std/pm 2*std confidence bounds. Figure 4 only reports a single run instead of the 5-average in Table 1.
  2. ρt\rho_t can not be computed in practice. It is defined relative to sup⁡f∈FScore\sup_{f \in F} Score, where FF is the open-ended space of allowed expressions. The quantity driving Theorem 2 therefore cannot be estimated, and the bound cannot be checked on any run.
Questions for authors
  1. Can you report, per stage, the guaranteed decrease 12[ScoreVt−1(ft)]2\tfrac{1}{2}[\mathrm{Score}_{V_{t-1}}(f_t)]^2 against the realized decrease L^(Vt−1)−L^(Vt)\hat{L}(V_{t-1}) - \hat{L}(V_t) after refitting? Can you give any empirical proxy for ρt\rho_t?
  2. How does test NLL behave for T>5T > 5, and would validation-based stopping change the conclusions?
  3. Are results sensitive to the chosen LLM in this case DeepSeek-v4? Is it possible that some algorithm is better with one LLM while others with another one?
  4. The final Ultimatum Game term has outer weight λ5(5)≈1025.86\lambda_5^{(5)} \approx 1025.86 on tanh⁡(g5)\tanh(g_5), where g5=∑cwcexp⁡{−τ(∣y−c∣+10−12)p}g_5 = \sum_{c} w_c \exp\{-\tau (|y - c| + 10^{-12})^{p}\} with τ≈21.2\tau \approx 21.2 and p≈0.13p \approx 0.13. By our calculation, the kernels decay to roughly 10−410^{-4} within about one grid spacing (1/499≈0.0021/499 \approx 0.002), so the term appears to act as spikes on the grid points nearest the focal shares c∈{1/2,1/3,2/3,1/4,3/4}c \in \{1/2, 1/3, 2/3, 1/4, 3/4\}. 4a. Is this reading correct? 4b. How sensitive are the term and its test-NLL contribution to the grid resolution (e.g., ∣Y∣=250|\mathcal{Y}| = 250 or 10001000)? 4c. How would you interpret this term behaviorally?
6 comments
AuthorsOpenReview ↗

Response to Reviewer kZyx: Empirical comparisons and candidate quality

We sincerely thank you for your careful review and valuable suggestions. Your questions prompted us to strengthen the empirical comparisons and evaluate the fitting guarantees numerically. We also appreciate your detailed calculation for the final Ultimatum Game term.

W1. Empirical margins and baselines

Empirical margins and baselines are thin. In Table 1, GreedyBM and AutoBM are often overlapping with /pm2 * std confidence bounds. Figure 4 only reports a single run instead of the 5-average in Table 1.

To assess the statistical significance of the test-NLL differences between GreedyBM and AutoBM, we conducted two-sided Welch's t-tests and permutation tests using five runs per method for each task, applying Holm correction across the three tasks. As shown below, both tests yield Holm-adjusted p-values below 0.05 for UG and RPS, confirming statistically significant NLL improvements in these two tasks.

TaskNLL difference: GreedyBM-AutoBM95% Welch CIWelch pHolm pPermutation pHolm permutation p
UG-0.248[-0.411, -0.0854]0.01020.02700.01590.0476
RPS-0.00941[-0.0157, -0.00308]0.009000.02700.02380.0476
CDA-0.0520[-0.273, 0.169]0.6000.6000.5790.579

Table 1 in our paper also compares against BoxLM, FunSearch, and nine expert-designed models across the three tasks. GreedyBM achieves lower mean test NLL than every baseline in each task.

We also emphasize that a major advantage of GreedyBM over AutoBM lies in its substantially lower token consumption during model search. Across the three tasks, Table 3 of the revised manuscript reports 87.2% to 90.9% fewer tokens overall. Notably, on CDA, GreedyBM achieves test NLL comparable to AutoBM while using about 13% of AutoBM's search tokens.

We have also expanded Figure 4 from one run per method to five runs per method across all three tasks. Table 4 of the revised manuscript reports the matched-token comparison, i.e., it compares the validation NLL achieved by GreedyBM and AutoBM under the same token budget.

TaskToken budgetGreedyBMAutoBM
UG345,6935.173 ± 0.1385.395 ± 0.066
RPS640,6431.0754 ± 0.00331.0907 ± 0.0020
CDA708,7324.801 ± 0.0825.209 ± 0.181

Values are validation NLL of the best retained model at the common terminal token budget in Figure 4, reported as means ±\pm sample SD over five runs per method. GreedyBM achieves lower mean validation NLL at these budgets in all three tasks.

W2 and Q1. Empirical lower and upper bounds on ρt\rho_t

ρt\rho_t can not be computed in practice. It is defined relative to sup⁡f∈FScore\sup_{f\in F}\mathrm{Score}, where FF is the open-ended space of allowed expressions.

Can you give any empirical proxy for ρt\rho_t?

The candidate quality ρt\rho_t in Theorem 2 is not known before search, but we can estimate it using candidate terms collected across several runs. We obtain lower and upper bounds by rescoring these terms at the current model and combining their scores with a computable upper bound on the best possible score.

Let sts_t be the selected term's score. For the upper bound, we evaluate a finite reference set of 2,420 terms, including the selected term, at the current model and take their maximum score MtM_t. This finite-set maximum cannot exceed the best score over the full term space, so st/Mts_t/M_t upper-bounds ρt\rho_t. For the lower bound, bounded term outputs let us bound every possible score using the absolute training residuals. Calling this computable score bound UtU_t, we obtain the lower bound st/Uts_t/U_t. Thus,

ρ‾t=stUt≤ρt≤stMt=ρ‾t.\underline{\rho}_t=\frac{s_t}{U_t}\leq\rho_t\leq\frac{s_t}{M_t}=\overline{\rho}_t.

Across five runs and five stages per task, the resulting bounds on mean ρt\rho_t are:

TaskMean ρ̲ₜMean ρ̅ₜ
UG0.2110.587
RPS0.03660.401
CDA0.1180.380

Each entry averages the 25 corresponding stagewise bounds. These intervals provide an empirical estimate of average candidate quality while accounting for the unknown full-space optimum. Appendix A.7.1, Eqs. (22) and (23) and Table 5 (p. 17), gives the bounds and the calculation of UtU_t.

AuthorsOpenReview ↗

Response to Reviewer kZyx: Fitting guarantees

Q1. Guaranteed and realized training-NLL decreases at each stage

Can you report, per stage, the guaranteed decrease 12[ScoreVt−1(ft)]2\tfrac{1}{2}[\mathrm{Score}_{V_{t-1}}(f_t)]^2 against the realized decrease L^(Vt−1)−L^(Vt)\hat{L}(V_{t-1})-\hat{L}(V_t) after refitting?

Yes. At each stage, we compute the guaranteed minimum decrease Gt=st2/2G_t=s_t^2/2 from the selected term's score before weight refitting. The tables compare this guarantee with the realized decrease Dt=L^train(Vt−1)−L^train(Vt)D_t=\widehat L_{\mathrm{train}}(V_{t-1})-\widehat L_{\mathrm{train}}(V_t) after adding the term and refitting weights. Theorem 1's prescribed update adjusts the new term's weight. Joint refitting optimizes all weights and can lower training NLL further.

Ultimatum Game (UG)

StageGuaranteed minimum training-NLL decreaseRealized training-NLL decrease after refitting
10.3150.454
20.1260.172
30.03930.0866
40.01390.0919
50.01700.0753

RPS

StageGuaranteed minimum training-NLL decreaseRealized training-NLL decrease after refitting
10.005050.0161
20.001310.00402
30.0008140.00267
40.0009450.00669
50.0003380.00271

CDA

StageGuaranteed minimum training-NLL decreaseRealized training-NLL decrease after refitting
10.07060.534
20.05730.339
30.05210.150
40.007000.0651
50.008440.125

Each row reports the mean over five runs for that task and stage. For each task, the mean realized decrease exceeds the guaranteed minimum at all five stages. The stagewise results are in Appendix A.7.2, Table 7 (p. 18).

W2. Estimating and checking the multistage guarantee

The quantity driving Theorem 2 therefore cannot be estimated, and the bound cannot be checked on any run.

We evaluated the refined multistage bound of Theorem 2 (Appendix A.3, Eq. (19), p. 15) separately for all 15 five-stage runs using the candidate-quality bounds above. In every run, the observed cumulative training-NLL decrease exceeded its guaranteed minimum.

The estimated intervals, averaged over five runs per task, are 0.00484 to 0.0285 for UG, 4.35 × 10⁻⁶ to 4.33 × 10⁻⁴ for RPS, and 0.00121 to 0.0118 for CDA. The observed cumulative decreases average 0.880, 0.0322, and 1.212, respectively. The lower endpoints give the guaranteed minima from the conservative candidate-quality bounds, and the upper endpoints use the finite-reference estimates.

The analysis combines worst-case estimates of curvature, the reference direction, and the best score over the term space. The multistage estimates and observed decreases are reported in Appendix A.7.1, Table 6 (p. 18).

AuthorsOpenReview ↗

Response to Reviewer kZyx: Stage budget and LLM backends

Q2. Test performance beyond five stages and validation-based selection

How does test NLL behave for T>5, and would validation-based stopping change the conclusions?

Extending from T=5 to T=10 lowers test NLL in all 15 runs:

TaskTest NLL at T=5Test NLL at T=10Paired decrease (T=5 minus T=10)
UG5.151 ± 0.1174.194 ± 0.5080.957 ± 0.490
RPS1.0771 ± 0.01111.0649 ± 0.00990.0122 ± 0.0090
CDA4.797 ± 0.2094.484 ± 0.3250.313 ± 0.199

The table compares stages five and ten of the same runs, following Appendix D.2, Table 15. Values are means ± sample SD over five runs per task.

To assess validation-based selection, we choose the stage with the lowest validation NLL among stages 1 to 10, then report that stage's test NLL:

TaskLowest validation NLLValidation-selected TTest NLL at the selected T
UG4.164 ± 0.49410 in all five runs4.194 ± 0.508
RPS1.0691 ± 0.010210 in four runs, 9 in one1.0649 ± 0.0099
CDA4.493 ± 0.31610 in all five runs4.484 ± 0.325

Every validation-selected model achieves lower test NLL than the stage-five model from the same extended run. The conclusions therefore hold under validation-based selection. Appendix D.2, Table 15 (p. 32), reports the paired test-NLL changes.

Q3. Sensitivity to the LLM backend

Are results sensitive to the chosen LLM in this case DeepSeek-v4? Is it possible that some algorithm is better with one LLM while others with another one?

GreedyBM achieves lower mean test NLL in five of the six task and backend comparisons with Qwen3.8-Flash-Next (FP8) and GPT-6-Luna (medium reasoning effort), including RPS and CDA under both backends. The mean NLLs in the remaining Luna-UG comparison are close, at 5.22 for GreedyBM and 5.18 for AutoBM.

The table reports means ± sample SD over five runs per strategic scenario. We use the same experimental setup as in the DeepSeek experiments, changing only the LLM backend. Further details are provided in Appendix D.3.

BackendMethodUG NLL ↓RPS NLL ↓RPS Acc. (%) ↑CDA NLL ↓CDA MSE ↓
QwenGreedyBM4.56 ± 0.5741.067 ± 0.012144.1 ± 1.654.65 ± 0.1560.0335 ± 0.00230
QwenAutoBM5.16 ± 0.07541.089 ± 0.0043341.4 ± 2.514.75 ± 0.1080.0361 ± 0.00623
LunaGreedyBM5.22 ± 0.1721.071 ± 0.0033343.4 ± 0.9434.81 ± 0.2540.0356 ± 0.00646
LunaAutoBM5.18 ± 0.1931.074 ± 0.012841.5 ± 2.844.86 ± 0.1660.0307 ± 0.00248

The statistical comparison of predictive performance is reported in Appendix D.3, Table 17.

GreedyBM also reduces search-token consumption by 77.7% to 90.0% across the two backends. The following table reports token usage in thousands as means ± SD:

BackendMethodUGRPSCDA
QwenGreedyBM650 ± 60.01,240 ± 89.3925 ± 66.9
QwenAutoBM3,880 ± 1656,280 ± 3027,040 ± 569
LunaGreedyBM429 ± 75.4999 ± 194819 ± 262
LunaAutoBM4,310 ± 1324,480 ± 1286,440 ± 189

We have added the additional-backend performance, statistical tests, and token comparisons in Appendix D.3, Tables 16 to 18 (p. 33).

Reviewer_HKi4OpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

Building models of human strategic choices usually requires substantial manual work to specify the utility function. GreedyBM uses an LLM to propose one behavioral term at a time, selecting terms through the mismatch between observed and model-predicted averages. The score provides a guaranteed training-loss decrease, and experiments on three strategic tasks report lower mean test NLL with substantially fewer search tokens.

Critical strengths

1.The term-by-term construction is a sensible way to structure LLM-based model discovery. It lets the search focus on behavior the current model still misses, while keeping the resulting expressions open to inspection.

2.The selection score has a clear statistical interpretation and a training-loss guarantee. The predictive results and large token savings across three tasks are encouraging.

Critical weaknesses

1.The experiments do not clearly isolate the benefit of the proposed score from the benefit of term-by-term construction. The existing ablations mainly examine feedback diversity and candidate reuse.

2.The matched-token comparison covers only one representative Ultimatum Game run. This leaves uncertainty about the efficiency advantage across tasks and runs, particularly once candidate-parameter optimization is included.

Questions for authors

1.Within the same candidate pool, how closely does the score ranking agree with the actual NLL improvement after weight refitting? This would help assess the score’s practical quality.

2.Can the existing runs provide matched-token comparisons across all three tasks? Please also clarify the runtime spent optimizing candidate parameters and refitting weights.

4 comments
AuthorsOpenReview ↗

Response to Reviewer HKi4: Selection score and NLL improvement

We sincerely thank you for your careful assessment and valuable suggestions! Your questions helped us strengthen the evidence for the selection score and extend the efficiency analysis across tasks and runs. We address each point below.

W1: Isolating the contribution of the selection score

The experiments do not clearly isolate the benefit of the proposed score from the benefit of term-by-term construction. The existing ablations mainly examine feedback diversity and candidate reuse.

We compare GreedyBM with two alternative selection rules (i.e., squared-discrepancy-based heuristic selection and random selection), while retaining term-by-term construction and joint weight refitting. The heuristic favors terms whose values at observed choices differ substantially from their expectations under the current model. Define

\qquad H_V(f)=\frac{1}{NB^2}\sum_{i\in\mathcal I^{\mathrm{train}}}d_i(f)^2,$$ where $q_V(a\mid x_i)$ is the current model's probability of action $a$ in strategic context $x_i$, $N$ is the number of training records, and $B$ bounds the term outputs. This heuristic replaces the selection score in term-parameter tuning, feedback, representative selection, and stage-end selection. The random variant skips term-parameter tuning, retaining the generated initial term parameters, and selects terms uniformly at random before jointly refitting the weights. | Task | GreedyBM | Heuristic | Random | | --- | ---: | ---: | ---: | | UG | **5.121** ± 0.133 | 5.558 ± 0.125 | 5.616 ± 0.143 | | RPS | **1.070** ± 0.004 | 1.092 ± 0.008 | 1.096 ± 0.002 | | CDA | **4.962** ± 0.164 | 5.224 ± 0.227 | 5.305 ± 0.303 | Values are test NLL, reported as means $\pm$ sample SD over five runs. GreedyBM achieves lower mean test NLL than both alternatives in all three tasks. Next, we explain the performance gap between our GreedyBM and the Heuristic method by analyzing the derivative of the training NLL. With $\bar d=N^{-1}\sum_i d_i(f)$, Equation (4) gives $$\left.\frac{d}{d\alpha}\widehat L(V+\alpha f)\right|_{\alpha=0}=-\bar d.$$ Our score is $|\bar d|/B$, and Theorem 1 guarantees a training-NLL decrease of at least $\mathrm{Score}_V(f)^2/2$ for its prescribed update. The heuristic squares discrepancies before averaging: large positive and negative discrepancies can increase $H_V(f)$ while canceling in $\bar d$, yielding little guaranteed NLL reduction. **We have presented this comparison and explanation in Section 4.1 and included implementation details in Appendix C.1.1 of our revised paper.** **Q1: Agreement between score rankings and realized NLL improvement** > Within the same candidate pool, how closely does the score ranking agree with the actual NLL improvement after weight refitting? This would help assess the score’s practical quality. To assess how closely the score tracks actual gains after joint weight refitting, we compared score rankings with realized training-NLL improvement across 1,500 candidate terms in 75 fixed pools from 15 GreedyBM runs. We compare the selection rules on the same candidate terms generated by GreedyBM, keeping their expressions and term parameters fixed. W1 separately evaluates the full search under each rule. Retained gain is the selected candidate's realized training-NLL improvement divided by the largest improvement in the same pool. The last three columns report retained gains; for random selection, we report the expectation under uniform selection. | Task | Our score: median Spearman | Our score: positive Spearman | Retained gain (ours) | Retained gain (heuristic) | Retained gain (random) | | --- | ---: | ---: | ---: | ---: | ---: | | UG | 0.675 | 24/25 | **71.95%** | 60.10% | 44.91% | | RPS | 0.934 | 25/25 | **90.32%** | 36.41% | 33.76% | | CDA | 0.824 | 25/25 | **86.40%** | 63.02% | 40.66% | | Overall | 0.827 | 74/75 | **82.89%** | 53.17% | 39.78% | The score has positive rank correlation with realized improvement in 74 of 75 pools and retains 82.89% of the best available gain on average. After averaging the five stages within each run, our score retains more gain than both comparators in all 15 runs; the two run-level, two-sided exact sign-flip tests both give Holm-adjusted $p=1.2\times10^{-4}$. **Due to space limitations in the main paper, we have reported these diagnostics in Appendix C.1.2 and Table 13 of the revised paper.**
AuthorsOpenReview ↗

Response to Reviewer HKi4: Token efficiency and runtime

W2 and Q2(a): Matched-token comparisons across tasks and runs

The matched-token comparison covers only one representative Ultimatum Game run. This leaves uncertainty about the efficiency advantage across tasks and runs, particularly once candidate-parameter optimization is included.

Can the existing runs provide matched-token comparisons across all three tasks?

We have expanded Figure 4 in Section 4.2 to all three tasks, with five runs per method. It shows each run's best retained model and the candidate-level dynamics. The corresponding matched-token results are as follows:

TaskToken budgetGreedyBMAutoBM
UG345,6935.173 ± 0.1385.395 ± 0.066
RPS640,6431.0754 ± 0.00331.0907 ± 0.0020
CDA708,7324.801 ± 0.0825.209 ± 0.181

As in the original Figure 4, values shown above are validation NLL of the best retained model at each task's shared token budget, reported as means ±\pm sample SD over five runs per method. GreedyBM achieves lower mean validation NLL in all three tasks.

We have added the new Figure 4, incorporated the above results in Table 4, and discussed the search dynamics in Section 4.2 of our revised paper.

The optimization cost concerns raised in W2 are addressed separately in Q2(b) below.

Q2(b): Runtime spent tuning term parameters and refitting weights

Please also clarify the runtime spent optimizing candidate parameters and refitting weights.

GreedyBM involves two optimization steps: (1) Tune term parameters, which optimizes only the new term's parameters while reusing the current model probabilities and residuals; and (2) Refit weights, which jointly optimizes the weights through a low-dimensional convex optimization problem after selecting a term. In contrast, AutoBM performs Fit complete model, which fits an entire candidate model for each proposal. The search budgets are 100 candidate terms for GreedyBM (five stages with twenty candidates each) and 200 candidate models for AutoBM.

The table reports mean time per run for these full proposal budgets. For GreedyBM, we separately report the time spent on Tune term parameters and Refit weights. For AutoBM, we report the time spent on Fit complete model. All optimization times are included in total wall time. Weight-refitting time is reported in seconds.

TaskGreedyBM: Total (h)GreedyBM: Tune term parameters (h)GreedyBM: Refit weights (s)AutoBM: Total (h)AutoBM: Fit complete model (h)
UG4.032.5635.9523.8820.73
RPS0.390.009150.20629.0424.47
CDA9.478.3515.3027.2521.79

Including parameter optimization, AutoBM's total wall time is approximately 2.9 to 74.5 times GreedyBM's total wall time under these proposal budgets. For GreedyBM, most optimization time is spent tuning term parameters, while the five joint weight refits take less than one minute in total on average in each task.

Numerical optimization used Intel Xeon Platinum 8468V CPUs. We observe that UG and CDA require substantially more time than RPS. This is because UG and CDA evaluate 500 candidate actions per training record, compared with only three in RPS, resulting in higher computational costs for repeated term evaluations and gradient computations.

We have provided the timing results and numerical configuration in Appendix D.1 and Table 14 of the revised paper.

Changes made in the revised manuscript

In response to your nice suggestions, we have added:

  • The heuristic and random selection-rule comparison, together with an explanation of the score's relation to the NLL derivative, in Section 4.1 and Table 2 (p. 8); Appendix C.1.1 (pp. 26–27) gives the alternative rules in detail.
  • Three-task search trajectories with five runs per method in Figure 4 (p. 9), and matched-token validation results in Section 4.2 and Table 4 (p. 9).
  • Within-pool score-ranking correlations and retained-gain comparisons, including run-level statistical tests, in Appendix C.1.2 and Table 13 (p. 27).
  • Total times and separate GreedyBM times for tuning term parameters and jointly refitting weights in Appendix D.1 and Table 14 (p. 32).

Thank you again for your constructive suggestions. We hope the additional analyses and clarifications address your questions, and we would be glad to discuss any remaining points.

Thank you for the detailed response and the additional experiments.

I appreciate the new ablation studies and the analysis of the selection score. The results provide stronger evidence for the contribution of the proposed score.

The matched-token comparisons across all three tasks and the additional runtime analysis also address my concerns about efficiency.

Overall, I think the authors have addressed my main concerns, and the revised paper is much stronger. I am willing to increase my rating from 2 to 3.

Thank you for following up. We are glad that the revisions and additional results resolved your concerns. We appreciate your time and engagement throughout the discussion.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.