Review Papers Against Their Own Results

 Quick Explanation Copied

 Long Explanation

What the paper claims and how strongly it is supported

Chen, Thomas, Tenenbaum, and Saxe propose that relationships act like latent features of a sociological environment, analogous to physical obstacles: a dyad's intimacy lowers the discomfort of vulnerable actions (shared utensils, bodily contact, disclosure), which observers invert alongside desire and physical effort to explain behavior. They operationalize this with a hybrid architecture — Llama-3.3-70B generates context-specific alternative actions and scores goal-satisfaction, effort, and risk; a softmax inverse-planning model then performs exact enumeration over these elicited features Model and Methods section.

The empirical support is unusually disciplined: six preregistered experiments (N=1,554 retained of 1,564), 16 food and 16 non-food vignettes spanning substance, space, and privacy domains, leave-one-scenario-out cross-validation, and public code/data on GitHub and Zenodo Preregistration and data availability. Key finding: only the full model reproduces relationship-dependent inference — high-risk sharing is stronger evidence of desire in formal relationships, and weaker evidence of intimacy when desire is high or low-risk alternatives are effortful.

Simulated note: the chart reproduces the reported correlations and CIs from the paper's Results text; no novel data are plotted Figure 2-4 correlations.

Critical assessment

Strengths. The novelty is real: prior computational relationship work covered social connectedness and vicarious utility, but the formality–intimacy axis as an environmental constraint in inverse planning had not been formalized. The cross-domain generalization is compelling — food-fitted utility weights predicted non-food judgments at r=0.977 vs 0.988 for domain-fitted weights, arguing the structure is not a saliva-specific contamination response Cross-experiment generalization.

Weaknesses and blind spots. The authors transparently report five preregistration deviations; critically, under the fully preregistered specification the full model's held-out log-likelihood advantage over the vanilla model was not reliable in Studies 1b and 3a (95% CIs crossing zero) — the primary-metric reversal and post-hoc comparison-set reweighting (fitted η up to 8.85 in Study 2a) are the load-bearing changes Deviations from preregistration. Other limits: US-only participants and contexts (LM-elicited features inherit Western norms); only the intimacy dimension tested; the 'we-agent' abstraction cannot capture asymmetric, reluctant, or nonconsensual action; and the very high correlations near the noise ceiling partly reflect vignette design with strong experimental manipulations rather than free-form everyday observation. BGPT inference (not authors' claim): the observed~90% explainable variance at the cell level (Table S7) is impressive but the paradigm measures third-party reading of explicit relationship labels, not implicit relational knowledge in live interaction.

What would change the conclusion. A replication where the full model's advantage disappears under the preregistered specification, or where intimacy no longer modulates desire inferences after controlling desire and effort, would falsify the framework. Cross-cultural tests (where sharing norms differ) are the sharpest future test the authors themselves propose.

 BGPT Paper Review

Study Novelty

First computational formalization of the formality–intimacy axis as an environmental constraint within inverse planning, plus a neurosymbolic method extending structured ToM models to open-ended vignettes; builds incrementally on well-established inverse-planning and intimacy literature.

Scientific Quality

Six preregistered experiments, cross-validation, noise ceilings, and open data are exemplary. Deducted for five preregistration deviations, primary-metric reversal, post-hoc reweighting that materially improves fit (η up to 8.85), unreliable preregistered advantages in two studies, and near-ceiling correlations that partly reflect designed vignette structure.

Study Generality

Generalizes across food, substance, space, and privacy domains within US contexts, but only one relationship dimension is tested and cultural generality is explicitly unverified; the framework is claimed as extensible to rank and roles but not demonstrated.

Study Usefulness

Provides a reusable method (LM-proposed comparison sets + symbolic inverse planning) for testing cognitive models in open-world situations, and a falsifiable framework distinguishing intimacy-modulated discomfort from vicarious value — useful for social cognition and interpretable AI.

Study Reproducibility

Full preregistration on OSF, public code (GitHub) and data (Zenodo), detailed prompts and parameters; minor deductions because LM elicitations are non-deterministic across API versions and reported deviations complicate exact replication.

Explanatory Depth

Gives a mechanistic, parameterized account (cost factorization, γ-shaped intimacy attenuation, joint inference) with explicit model comparison against ablations; does not explain how mutual knowledge forms, is negotiated, or how multiple relationship dimensions integrate.