Pith. sign in

REVIEW 5 major objections 6 minor 4 references

Inferring Hidden Motives: A Utility Bayesian Model of learning the values of others

T0 review · 5 major / 6 minor · reviewed 2026-08-03 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read People infer others' hidden social motives by continuously updating beliefs over a seven-parameter utility space, and this model predicts their judgments better than non-Bayesian or discrete-type alternatives.

desk verdict Worth a serious look, but don't trust the headline numbers yet: the static-IC selection step and internally inconsistent reported ratios need to be fixed before the seven-parameter utility claims can be used. read the letter →

arxiv 2511.07825 v2 pith:ZPKDKAEQ submitted 2025-11-11 q-bio.NC

Gregory Stanley , Jun Zhang , Rick Lewis This is my paper
classification q-bio.NC
keywords socialpreferencesBayesianbeliefupdatingdictatorgameutilityfunctiondiscretevscontinuousmodelsaltruisminequalityaversiontheoryofmind
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper tries to show that when people watch others make allocation decisions, they learn the others' moral motives—self-interest, altruism, envy, guilt—not by labelling them as types but by continuously updating a distribution over motive strengths. The Utility Bayesian Model (UBM) treats each observed choice as evidence reweighting a prior over the parameters of the chooser's utility function; the paper reports it fits participants' predictions in repeated binary dictator games better than non-Bayesian baselines and better than every one of more than 15,000 discrete typological models. A companion comparison of 476 candidate utility functions selects a seven-parameter form, and fitting it to choosers yields population-level distributions: self-interest outweighs altruism by roughly nine to one, guilt (aversion to being ahead) outweighs envy (aversion to being behind), and other-regarding terms grow super-linearly with payoff size, with higher-than-expected fractions of sadistic, masochistic, and competitive types. The payoff is a common parameter space for preferences and beliefs about preferences, making cross-study comparison and empirically grounded priors for future social-cognition models possible.

What carries the argument

The load-bearing mechanism is the Utility Bayesian Model (UBM): a probability distribution over the parameters of another agent's utility function, updated by Bayes' rule with a SoftMax likelihood after each observed choice. In the full model the hypothesis space is continuous and approximated by dense grids or particle filters; discrete typologies are special cases that zero out most of the space. The likelihood itself is anchored by the winning seven-parameter utility function—self-interest, altruism, envy, guilt, and term-specific exponents—so the same machinery simultaneously identifies what people value and how they update beliefs about what others value.

What would settle it

Rerun the 476-model information-criterion comparison using the full dynamic Bayesian updating model (or a particle-filter approximation) on the same human-human data. If a different functional form—say, one with distinct exponents per term replaced by a single exponent, or guilt and envy symmetrized—wins under dynamic updating, the seven-parameter model and the distributional results built on it lose support. Alternatively, if a discrete model with types placed on an irregular grid or with envy/guilt dimensions surpasses UBM's total negative log-likelihood on the human-bot predictions, the con

Watch

Extended reading notes

Core claim

The central claim, stated in the paper's own terms, is that social-preference learning is quasi-continuous Bayesian inference: observers maintain graded beliefs over a multidimensional space of utility parameters and reweight those beliefs after each observed allocation choice, rather than placing others into discrete categories. This is established structurally by comparing the UBM against three non-Bayesian models and against thousands of discrete Bayesian models—built by placing prior mass on subsets of a 5x5 grid of self-interest and altruism values—and showing the continuous model yields lower negative log-likelihood on participants' predictions. The paper also claims that the best util

Load-bearing premise

The seven-parameter utility function was chosen using a simplified, non-updating version of the model because full dynamic updating across 476 forms was computationally infeasible; if that static approximation changes which utility form wins, the reported parameter distributions and the envy/guilt ordering could shift.

Editorial extensions

If this is right

  • Separating predictors' explicit predictions from choosers' choices lets a single experiment estimate both actual social preferences and beliefs about social preferences, unconfounded by strategic motive.
  • The reported mean and variance parameters for the seven dimensions give other modelers ready-made Bayesian priors for social-preference inference.
  • Discrete social-value-orientation categories—altruistic, selfish, competitive—are, at best, coarse summaries; a graded representation is needed to capture how people actually predict others.
  • If guilt does outweigh envy, canonical inequality-aversion results that put envy first may have conflated self-interest with social comparison, and future models should treat the two directions of inequality separately.
  • The fitted exponents above 1 on other-regarding terms question the standard assumption of diminishing marginal sensitivity to payoffs in social contexts.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The paper's continuous-vs-discrete comparison used a two-parameter (self-interest, altruism) space; an editor-level extrapolation is that the same graded-updating advantage would hold in higher-dimensional spaces (e.g., adding envy and guilt), but that is not directly tested.
  • The static selection of the seven-parameter utility function is the hinge for the parameter-distribution claims; if full dynamic updating were computationally feasible and changed which form wins, the envy/guilt reversal and convex exponents could shift. A concrete follow-up is to rerun the 476-form comparison with particle-filter updating.
  • For applications, the parameter distributions imply an implementable recipe for AI or social agents: initialize beliefs from the reported population means and update with the UBM likelihood; a testable extension would be whether agents with such priors cooperate more effectively in repeated trust games.
  • Because matching probabilities were transparent and repeated encounters likely activated reciprocity, the high rates of sadistic or competitive types may partly reflect strategic punishment rather than pure preferences; paid, one-shot variants would clarify this.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

5 major / 6 minor

Summary. The paper introduces a Utility Bayesian Model (UBM) in which observers maintain a continuous probability distribution over another person's social-preference parameters and update it after observing binary dictator-game choices. In a human-bot experiment, the UBM is compared with non-Bayesian baselines and with thousands of discrete typological Bayesian models; in a human-human experiment, a seven-parameter utility function is selected from 476 candidates by BIC and then used in the dynamic UBM to estimate population distributions of self-interest, altruism, envy, guilt, term-specific exponents, and temperature, separately for choosers and predictors. The central claims are that people track others' motives as graded continuous positions rather than discrete categories, and that the fitted seven-dimensional parameter distributions reveal self-interest dominating altruism, guilt exceeding envy, super-linear (γ>1) sensitivity in altruism and social-comparison terms, and higher prevalence of antisocial types than previously estimated.

Significance. If the central claims hold, the paper would make a substantial contribution: it provides a unified seven-dimensional parameterization of outcome-based social preferences, an unprecedentedly broad utility-form comparison, and a large-scale continuous-versus-discrete model comparison. The explicit elicitation of predictions, the randomized payoff space, and the simulation-based optimizer checks are genuine strengths, and the authors are transparent about several limitations. However, three load-bearing issues currently prevent the claims from being accepted as stated: the utility function used throughout the dynamic model is selected under a static likelihood, the continuous-versus-discrete model comparison does not control for asymmetric model flexibility, and the headline numerical results contain internal inconsistencies. Each of these is addressable, but each affects the paper's main conclusions.

major comments (5)
  1. [§4.2 and footnote 11; Table 6] The seven-parameter utility function that drives all of Section 5 is selected by an IC analysis that uses a static, non-updating model, as the authors acknowledge. Predictor predictions in the actual experiment depend on posterior beliefs updated from previous observations, so the likelihood used for predictor decisions is misspecified. The margin over the next-best model is only ΔBIC ≈ 17 (Table 6), so even a modest shift in NLL under the true dynamic likelihood could reorder the candidates. If a different form wins, the population means, the guilt>envy reversal, the γ>1 results, and the negative-type frequencies would all need re-estimation. The paper needs either a dynamic-IC robustness check (e.g., on a subset of participants or a smaller candidate set) or a principled argument that the static approximation is conservative.
  2. [§3.3.8, §3.3.9, and Table 4] The headline comparison between the continuous UBM and the discrete/typological models uses in-sample NLL without penalizing for the much larger number of fitted parameters in the UBM. The UBM is fitted with participant-level priors, while the typological models are optimized only at the population level, with individual-level fitting performed only for the single best typological model (NLL ≈ 8470 vs. UBM ≈ 5811). This asymmetry alone could explain part of the UBM's apparent advantage. To support the claim that people represent motives continuously rather than categorically, the authors should report cross-validated NLL or a complexity-penalized criterion (e.g., BIC or AIC), and ideally fit the discrete models with comparable participant-level flexibility. The 5×5 grid comparison addresses resolution but not parameter-count differences.
  3. [§5.1.1; Abstract; §2; §6.4] The paper's headline self-interest-to-altruism ratio is internally inconsistent. Section 2 reports a mean ratio of 3.05; §5.1.1 reports 8.974 and states that the average ratio was 'roughly nine times'; the abstract says 'roughly nine times'; §6.4 says 'roughly seven times' and also 'about three times'. Moreover, the bracketed definition in §5.1.1, 'V_ii/(V_ii+V_ij): mean = 0.755', implies V_ii/V_ij ≈ 3.08, not 8.974. These numbers cannot all be correct. Because this ratio is a central empirical result, the authors should correct the definition, report the actual ratio used, and ensure consistency across the abstract, introduction, results, and discussion.
  4. [§5.1.1 vs. §2] The reported prevalence of negative social-preference types is inconsistent. Section 5.1.1 states that 50.68% of choosers exhibited negative altruism and 10.96% negative self-interest, while Section 2 states 'approximately 27%' negative altruism and 'roughly 16%' negative self-interest. These discrepancies are not explained by the exclusion criteria described in the text. Since the claim of 'more antisocial preferences than previously estimated' is a key finding, the authors must reconcile these numbers and report the exact computation and sample sizes.
  5. [§4.6.2 and Equation 1] The 'parent-fair L2 penalty' is asserted not to change which model wins, but no sensitivity analysis is provided. The penalty is applied during optimization and then subtracted from reported loss; this can affect the path taken by the optimizer and the final NLL. Given that the entire Section 5 depends on the utility selection, the paper should show that the ranking of the top models (especially the k=7 vs. k=9 contest) is stable across a range of λ values, or at least that the selected model remains the BIC winner under reasonable variations of the penalty weight.
minor comments (6)
  1. [Abstract; §4] The abstract says the seven-parameter function was selected from '526 candidate forms', but the text and tables consistently say 476. Please correct.
  2. [§3.3.2; §4.4 equation] The example utility function contains subscript typos: in the envy and guilt terms, the second payoff should be π_j^A, not π_i^A. A similar typo appears in the definitions of the envy/guilt terms in Section 4.4. These make the equations hard to read.
  3. [Table 6 caption] The caption states 'the k = 5 model with ΔBIC = 0', but the table shows k = 7 with ΔBIC = 0.00. The caption and table should agree.
  4. [Figure 11 caption] 'Guild' should be 'guilt'. The sample-size notation 'n participants ∈ [0, 9]' and '[0, 19]' is unclear and should be explained.
  5. [References] The Kerschbamer (2015) entry appears twice. Also, 'Charles & Rabin' is an idiosyncratic citation of Charness and Rabin and should be standardized.
  6. [§3.3.6] The simulation text says 945 simulated dyads but Figure 5 reports n = 946. Please check the count.

Circularity Check

1 steps flagged · score 6.0 of 10

The central 'superior predictive accuracy' of the continuous UBM over discrete models is an in-sample fit: the same prediction data are used to fit the model's prior parameters and to compute the NLL reported as predictive performance.

  1. fitted input called prediction [Section 3.3.8 (Table 4) with Section 3.3.5 and Section 3.3.7]
    "We evaluated models via total negative log-likelihood (NLL) calculated from participants' predictions, with lower NLL indicating higher predictive accuracy. ... Utility-Bayes achieves substantially lower NLL (5811.11 with a 21×21 grid) ... we fitted priors to participants across all their counterpart, rather than dyad by dyad."

    The UBM's prior parameters are optimized by minimizing the same cumulative NLL over participants' predictions (Section 3.3.5: 'minimize this cumulative loss across all observations'), and the model-comparison table reports that same fitted NLL as 'predictive accuracy.' No held-out split or cross-validation is described. The continuous model is compared to discrete models on the training data, so its better NLL is the fitting objective, not a forward prediction. The headline conclusion that people represent others as continuous rather than discrete types is therefore supported by an in-sample fit advantage, i.e., a fitted value renamed as prediction.

full rationale

The main derivation chain (prior -> likelihood -> posterior -> prediction) is internally coherent, and the simulation checks (e.g., parameter recovery r=0.755; grid/particle correlation >0.993) are genuine external validations. I found no self-definitional equation collapse, no load-bearing self-citation (the reference list contains no prior work by Stanley/Zhang/Lewis), and no imported uniqueness theorem. The Section 4.2 static-IC selection of the seven-parameter utility, later embedded in the dynamic UBM, is an acknowledged model-misspecification risk (footnote 11 and Section 2.1) but not circularity: the IC criterion is not equivalent to the later parameter estimates. The one substantive circularity is that the 'superior predictive accuracy' claimed for the UBM over 15,000 typological models is computed as negative log-likelihood on the same prediction data used to fit the model's priors; this is a fitted loss presented as predictive evidence. For that reason the continuous-vs-discrete conclusion is partially circular. The fitted prior parameters labeled as participants' 'beliefs' are standard model-based estimates rather than a separate prediction, so I do not count that as an additional circular step. Overall score 6 reflects one central in-sample-fit-as-prediction issue; the rest of the paper's modeling is not circular.

Assumptions & free parameters 9 free parameters · 7 assumptions · 0 invented entities

The central claims rest on the SoftMax/Bayesian framework, a static-IC-selected utility function, and a set of fitted parameters. No new physical or ontological entities are introduced. The most consequential choices are the artificial R=3 reference point, the L2 penalty, and the Gaussian prior—all of which shape the reported parameter distributions.

free parameters (9)
  • Self-interest weight μ(Vii), chooser = 0.67861 (population mean)
    Fitted from choices in the human-human experiment; central to the self-over-other weighting claim.
  • Altruism weight μ(Vij), chooser = 0.04689
    Fitted per participant; drives the altruism and sadism estimates.
  • Envy weight μ(Eij), chooser = 0.01541
    Fitted from choices; part of the social-comparison asymmetry claim.
  • Guilt weight μ(Gij), chooser = 0.25591
    Fitted from choices; central to the claim that guilt outweighs envy.
  • Exponents γ1, γ2, γ3 = 0.82552, 1.33617, 1.37846 (chooser means)
    Fitted per term; basis for the γ>1 convexity claim for altruism and social comparison.
  • SoftMax temperature τ = 0.97034 chooser, 1.95989 predictor
    Fitted separately for choice and prediction; determines stochasticity in the likelihood.
  • Predictor prior means and standard deviations = Table 9 values
    Fitted to prediction data and then reported as participants' initial beliefs; these are the 'belief' results.
  • Parent-fair L2 penalty weight λ = 0.05
    Hand-chosen to stabilize optimization; the authors acknowledge it 'slightly biases parameter estimates toward zero.'
  • Reference point R = 3
    Artificial midpoint of the payoff range {1..5}; not independently measured, and the authors note it is not a true loss reference point.
assumptions (7)
  • domain assumption Choices and predictions are generated by SoftMax utility maximization with an outcome-based utility function.
    Stated in §3.3.3 and used throughout; excludes intentions, reciprocity, and strategic reasoning from the likelihood.
  • domain assumption Predictors update beliefs exactly via Bayes' rule over a truncated Gaussian prior on utility parameters.
    The UBM's core assumption in §3.3.2 and §3.3.4; the paper does not test alternative updating rules besides non-Bayesian baselines.
  • domain assumption Social preferences are static within the experiment.
    The model does not allow preferences to evolve; the authors list this as a limitation in §6.7.
  • domain assumption The population prior is Gaussian and is shared across predictors before individual fitting.
    Stated in §3.3.2; the fitted prior means are later interpreted as empirical facts about participants' initial beliefs.
  • standard math Information criteria (AIC/BIC) computed from summed negative log-likelihood are valid for selecting among the 476 utility forms.
    Used in §4.2; the comparison is in-sample and uses a static model, which the authors acknowledge.
  • ad hoc to paper The parent-fair L2 penalty does not change which utility model wins.
    Introduced in §4.6.2 to prevent nesting violations; the authors argue it only stabilizes optimization, but this is an untested modeling choice.
  • domain assumption In the human-bot experiment, avatar choices are generated by the two-parameter utility U = Vii(πA_i−πB_i) + Vij(πA_j−πB_j).
    Figure 1 caption; the discrete-vs-continuous model comparison rests on this assumed avatar utility, not the seven-parameter utility.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Inferring Hidden Motives: A Utility Bayesian Model of learning the values of others." pith.science (2026). https://pith.science/paper/ZPKDKAEQ

@misc{pith2026251107825,
  author       = {Pith},
  title        = {Pith review of: Inferring Hidden Motives: A Utility Bayesian Model of learning the values of others},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ZPKDKAEQ}},
  note         = {Machine review of arXiv:2511.07825}
}
read the original abstract

Cooperation depends on judging whom to trust: identifying the strengths of other people's morally relevant social motives, called social preferences, and refining those beliefs after observing their behavior over time. We model this process with a Utility Bayesian Model (UBM): a continuous probability distribution over the parameters of another person's utility function, reweighted after each observed choice. In repeated binary dictator games with randomized payoffs, the UBM captured participants' predictions of preprogrammed agents and of one another better than non-Bayesian alternatives and over 15,000 models that represent others as discrete types: people track motives as graded positions in a continuous space, not as categories. Because participants alternated between choosing and predicting, the framework separately estimates preferences and beliefs about preferences. Choosers' own utilities, fit with a seven-parameter function selected from 526 candidate forms, showed that most people place reliably positive weight on a stranger's payoff but weight their own roughly nine times more; that aversion to being ahead exceeds aversion to being behind, reversing the canonical ordering; that sensitivity to others' payoffs and to inequalities grows more than proportionally with their size; and that antisocial preferences are more common than prior estimates suggest. Predictors' initial beliefs captured the dominance of self-interest but expected aversion to being behind to outweigh aversion to being ahead. Together, these parameter distributions locate any individual's social preferences, and beliefs about them, in a common seven-dimensional space, enabling commensurable cross-study comparisons, meta-analyses, and empirically informed priors for future Bayesian models of social cognition.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

4 extracted references · 1 linked inside Pith

  1. [1]

    Abbink, K., & Sadrieh, A. (2009). The pleasure of being nasty. Economics letters, 105(3), 306-308. Akaike, H. (1973). Maximum likelihood identification of Gaussian autoregressive moving average models. Biometrika, 60(2), 255-265. Aksoy, O., & Weesie, J. (2014). Hierarchical Bayesian analysis of outcome -and process -based social preferences and beliefs in...

    arXiv 2009

  2. [51]

    f a l s e c o n s e n s u s e f f e c t

    Rabin, M. (1992). Incorporating fairness into game theory and economics. University of California at Berkeley, Department of Economics, 1281-1302. Rand, D. G., & Nowak, M. A. (2013). Human cooperation. Trends in cognitive sciences, 17(8), 413-425. Rawls, J. (1971). An egalitarian theory of justice. Philosophical Ethics: An Introduction to Moral Philosophy...

    1992

  3. [337]

    A., Joireman, J., Parks, C

    Van Lange, P. A., Joireman, J., Parks, C. D., & Dijk, E. V. (2013). The psychology of social dilemmas: A review. Organizational Behavior and Human Decision Processes, 120, 125-141. Wang, Q.-H., Wei, Z.-H., Chen, W.-N., Na, Y., Gou, H. -M., & Liu, H.-Z. (2025). The impact of inequality on social value orientation: an eye-tracking study. Frontiers in Psycho...

    2013

  4. [426]

    Luce, R. D. (1959). Individual Choice Behavior: A Theoretical Analysis. Mlneola, New York: Dover Publications, Inc. Marr, D. (1982). Vision: A computational investigation into the human representation and processing of visual information. Marshall, A. (1920). Principles of Economics (8th ed.). Macmillan and Co. McFadden, D. (1974). Conditional Logit Analy...

    1959

Pith tools

Reviewed August 3, 2026 · model on record in the stance chip above.