🟦 Programmatic proposal — what are LLMs, as research subjects?
Horton 2023 — Persona prompting as a theory tool.🟩 Field maps and methodological frameworks — describe or critique the space
Sun 2025 — Four-branch taxonomy of the field.Kozlowski & Evans 2025 — Six-weakness validity framework. Jingyao presents.🟨 Static behavioral evaluation — run LLM through games, measure aggregate behavior
Lorè & Heydari 2024 — One-shot 2×2 social-dilemma factorial.Fan 2024MindGames 2025 — Theory-of-mind + strategic-game benchmark.🟥 Dynamic / multi-agent evaluation — iterate, see what happens
Akata 2025 — Ten rounds × 1,224 games, seven families.Willis 2025 — LLM-authored strategies compete in a Moran-process tournament.🟪 Fine-tuning alternative — train a new model rather than prompt an existing one
Binz 2025 — Centaur. Llama 3.1 70B fine-tuned on 10.7M human choices from 160 psychology experiments.🐍 Notebook: 3_ipd_tournament.ipynb
Horton et al. observe that LLMs, prompted to adopt a persona, produce verbal responses that read as plausibly human — and, crucially, that these responses shift in systematic ways when the persona shifts.
A “socialist” persona and a “libertarian” persona, given the same scenario, produce different answers about fairness. A persona trained on equity considerations and one trained on efficiency produce different dictator-game allocations. A persona of a labor-market participant shifts hiring decisions when the minimum wage is raised.
The paper’s proposal is to treat this capacity as a theory tool. The agents are called Homo silicus, in deliberate parallel to Homo economicus — stipulated rather than empirical, a modeling convention rather than a population. The five experiments re-run classic behavioral-economics paradigms:
Kahneman 1986 — price-gouging fairness study;Charness & Rabin (2002) — unilateral dictator games across preference types;Samuelson & Zeckhauser (1988) — status-quo-bias federal-budget experiment;Oprea (2024) — certainty-equivalent / prospect-theory paradigm;Horton’s own minimum-wage hiring experiment.The paper is careful about what it is not claiming. It does not claim that LLM responses are equivalent to human responses. It does not claim that the LLM is the human subject. It claims, more modestly, that running experiments on LLMs is cheap, fast, and conceptually useful as a way of generating and pre-testing hypotheses before paying human subjects.
The social-science reading of Horton is that he has re-opened an old question — what kind of object is Homo economicus? — with a new instrument. The LLM is not a homo-economicus replica; it is a new thing. Whether it is a better or worse stand-in for a person than the idealized maximizer of neoclassical theory is the question the subsequent literature picks up.
A little over two years after Horton, the field needed a map. Sun and colleagues — a team based primarily at Peking University and Microsoft Research Asia — provide one. Their 2025 survey is explicitly a taxonomy paper: it does not run new experiments, it organizes the existing ones.
Their central contribution is Figure 1, a four-branch taxonomy of the intersection of game theory and LLM research.
Game Theory × Large Language Models (Sun et al. 2025)
│
├── §2 Evaluating LLMs in game-based playgrounds ◀── most of our W2/W3 papers
│ │
│ ├── behavioral features (what do LLMs do when they play?)
│ │ 🟦 Horton (W2) — personas × classical behavioral paradigms
│ │ 🟨 Lorè & Heydari (W2) — 2×2 games, structure vs. framing
│ │ 🟥 Akata (W2) — repeated games, PD/BoS asymmetry
│ │
│ └── strategized agents (can we engineer them to play better?)
│ 🟥 Akata SCoT (W2) — prompt-level ToM scaffold
│ 🟥 Willis et al. (W3) — evolutionary selection as engineering loop
│
├── §3 Improving LLMs with game-theoretic methods
│ ├── §3.1 interpretability (Shapley values)
│ ├── §3.2 preference alignment (social choice)
│ ├── §3.3 heterogeneity ◀── ⚠ means "heterogeneous users",
│ │ NOT within-model behavioral variance
│ └── §3.4 dynamic adaptation
│ 🟩 Kozlowski & Evans (W3) — straddles §2/§3; validity framework
│ for LLM-as-subject simulation
│
├── §4 Characterizing LLM-related events through game models
│ (policy: developer competition, deployer incentives;
│ no W2/W3 paper sits here)
│
└── §5 Advancing game theory with LLMs
(LLMs as method, not subject;
no W2/W3 paper sits here)
Outside the taxonomy:
🟩 Sun (W2) — is the taxonomy itself
🟪 Binz et al. Centaur (W3) — foundation model of human cognition, not GT
Rather than walk each branch abstractly, the more useful move is to ask a concrete question of each: what would it mean, in practice, for research in this branch to “benefit humanity”? The slogan is Willis’s closing line, but it is standard in closing paragraphs of survey and policy papers more broadly. Each of Sun’s four branches picks out a distinct downstream consequence that a political scientist, economist, or data scientist would recognise as the real stake.
§2 — Knowing in advance where LLMs break. This branch runs LLMs through games and catalogues the failures, producing a shortlist of deployment hazards with experimental evidence behind them. The consequences visible to this audience: LLMs “display adaptive economic strategies, such as tacit collusion in pricing” in Bertrand competition — an antitrust concern as soon as platforms deploy LLMs as pricing agents. They “regress to selfish or inconsistent strategies in complex scenarios without social scaffolding” — a concern for any multi-party negotiation an LLM is asked to represent. They lose “role consistency and logical coherence under pressure” in identity games — a reliability concern for any high-stakes role-playing deployment, from citizen service bots to legal triage. The downstream payoff of §2 is that these hazards are documented, so deployment and regulatory conversations can reference evidence rather than intuition.
§3 — Making the model you interact with more accountable. The technical repairs in this branch look inside-the-stack, but their downstream consequences are directly political. Interpretability: attributing a model’s decision to specific inputs or components is the precondition for audit — which is in turn the precondition for algorithmic-accountability law (the EU AI Act, GDPR’s “right to explanation”). Alignment via min-max equilibrium drops the assumption that human preferences are transitive and scalar — real preferences, including voting preferences, are neither, so an alignment procedure that assumes otherwise systematically misrepresents the people it is supposed to serve. Preference heterogeneity: Sun reports a striking finding — canonical RLHF is mathematically equivalent to Borda count, a voting rule with known tyranny-of-the-majority properties. Left uncorrected, this means the model you interact with is implicitly doing majority voting over whose preferences to honour, silently under-serving any subgroup that is a minority among the annotators. Social-choice axioms and mechanism design are being used to fix this. Dynamic adaptation stops the model from “reward-hacking” its own training signal as it improves. The §3 payoff is models that are auditable, fairly-aligned across subgroups, and less likely to drift — all things that matter in politics and economics long before they matter in ML.
§4 — Predicting what the AI-rollout market will do before it does it. This is the branch most directly relevant to economists and political scientists, because it treats LLMs as institutional players and models the strategic equilibria their deployment produces. Three concrete results give a sense of the leverage. Taitler et al. adapt Braess’s Paradox to GenAI: a platform optimising for user revenue draws users away from the human knowledge forum it depends on (think Stack Overflow), which drains future training data, which eventually degrades the platform. Everyone ends up worse off despite each actor playing locally well — a classic system-level failure a regulator can actually aim policy at. Laufer et al. show that imposing safety standards only on downstream fine-tuners can incentivise upstream developers to underinvest in safety — a textbook regulatory arbitrage result, recognisable to anyone who has read the economics-of-regulation literature. Buening et al. derive payment rules that stop RLHF annotators from strategically misreporting their preferences, so the training data is not corrupted by the annotators’ own incentives. These are the pieces that give regulators and platform designers real system-level failure modes to intervene on.
§5 — Extending game-theoretic analysis into the domains classical mechanism design couldn’t reach. Classical mechanism design requires problems expressible in numbers, which rules out most of the coordination problems social scientists care about — citizen deliberation, policy bargaining, procurement on qualitative criteria, democratic preference elicitation. LLMs remove that restriction. Generative social choice (Fish et al., Boehmer et al.) extracts preferences from free-text deliberation and produces representative policy slates — directly applicable to Polis-style citizens’ assemblies, a growing instrument in democratic governance. Semantic-enhanced auctions (Sun et al.) let Vickrey and combinatorial auctions run on natural-language valuations — relevant to public procurement and policy design. Verbalised Bayesian persuasion models dialogue in real negotiations. Mechanism-design as code generation (Liu et al.) has the LLM write the mechanism itself, in human-readable form. The §5 payoff is practical governance artefacts for problems previously approachable only through qualitative analysis — many of which are the problems political science cares about most.
Sun writes the connective tissue himself at the close of §3: “the game-theoretic methods used to enhance LLM interpretability, alignment, and adaptation do not exist in a vacuum. They are developed in response to the immense economic and social pressures that define the LLM ecosystem.” The four branches are not independent. §2 documents what LLMs get wrong; §3 builds accountability and fairness into the models themselves; §4 models the market and regulatory environment that determines whether those fixes actually get deployed; §5 uses LLMs to extend governance mechanisms into language-based domains previously out of reach. Willis’s slogan “benefit humanity” means one of these four specific things, depending which branch the paper sits in. Our three empirical papers (Lorè & Heydari, Fan, Akata) live entirely inside §2, so the benefit story for their work is the surveillance one — they tell us what not to trust LLMs with in the domains that matter to this audience.
One load-bearing clarification from the taxonomy itself, before we move on: Sun’s §3.3 “heterogeneity” node names heterogeneous users — different people wanting different things from their model — not within-model behavioural variance. The question of whether the same GPT-4, run a thousand times on the same prompt, behaves as one agent or as a mixture of behavioural types is not on the list. Patterned heterogeneity of LLM behaviour is the category Sun’s taxonomy does not contain.
Group-project pause. Read Sun’s taxonomy as a heatmap of saturation. The most saturated quadrant is §2 behavioral features — the place where a new 2×2 game with a new model produces the smallest marginal contribution. §3.3 (user-heterogeneity alignment) is well-populated. §5 (LLMs as a method for game theory) is underexplored and technically demanding. §4 is policy-relevant and needs social-science training rather than computer-science training. And the genuine gap — the category Sun’s taxonomy does not contain at all — is the within-model, within-prompt heterogeneity question the literature has not yet taken up.
The four papers in the yellow and red sections below all prompt an existing model: the LLM is taken as given, situated in a game or task, and observed. Binz 2025 makes a different move. Rather than ask what a pre-trained model does in a new situation, it fine-tunes a language model directly on human behavioral data and evaluates whether the resulting model predicts the next choice of a held-out human subject. The distinction here is the same one you know from econometrics between using someone else’s published model and re-estimating its parameters on your own dataset. On Sun’s taxonomy, this move sits outside §2–§5 entirely — it is neither an evaluation of an existing model, nor an alignment fix, nor a game about LLMs, nor a use of LLMs for game theory. Within this chapter it is 🟪, the fifth category, and it is the paper on the reading list whose title most directly matches the title of this course.
Centaur is Llama 3.1 70B fine-tuned on Psych-101, a corpus the Binz group curated by transcribing 160 psychological experiments into natural language: 60,092 participants, 10.7 million choices, 253 million tokens. The fine-tuning technique used (QLoRA) updates only about 0.15% of the model’s weights, which is why the whole training run fits on one GPU in five days instead of a cluster in months. The training objective is set up so the model learns to predict the subject’s next choice, not to complete the experimental stimulus. One selection criterion is worth naming: Psych-101 is restricted to experiments whose stimuli and responses can be expressed in natural language — which rules out purely image-based tasks, motor tasks, and anything whose critical structure lives outside text.
The paradigms in Psych-101 are cognitive psychology, not game theory — multi-armed bandits, decision-making under risk, memory, categorization, Markov decision tasks. The case for putting Centaur on a game-theory reading list is that it is the state of the art at the job Horton proposed. Where Homo silicus is a theory tool — plausible, consistent with stylized facts, useful for pre-testing — Centaur is a trained prediction model, evaluated on negative log-likelihood against held-out responses.
On 32 held-out experiments it outperforms both the base Llama model and the best domain-specific cognitive model in nearly every case, across six domains:
multi-armed bandits ─ spatially correlated, structured, horizon, drifting 4-armed, ...
decision-making ─ choices13k, Iowa Gambling Task, intertemporal choice,
balloon analogue risk, two-step, multi-attribute, ...
memory ─ digit span, N-back, recent probes
supervised learning ─ Shepard / Medin categorization, weather prediction
Markov decision tasks ─ two-step, Maggie's farm
miscellaneous ─ THINGS odd-one-out, grammar judgement, go/no-go, ...
Three generalization tests go beyond held-out participants:
Centaur also predicts response times. The logic is Hick’s law — harder decisions take longer — and “harder” is measured by how uncertain the model is about its own next-token answer on each trial. Plug that uncertainty measure into a linear mixed-effects regression on log-RT and Centaur explains substantially more variance than Llama or the cognitive baselines (conditional R² = 0.87 vs. 0.75 vs. 0.44).
A separate result from the paper, worth naming because it is a different kind of evidence — not behavioral prediction but representational similarity. After fine-tuning, Centaur’s hidden-layer activations on a given trial correlate more strongly with aggregated human fMRI signal on the same trial than Llama’s activations do (Fig. 4 in the paper). The procedure is straightforward: run each model on the trials the human subjects saw in a scanner, record the internal activations layer-by-layer, regress the averaged brain signal onto those activations, and compare the fits. Centaur wins, with the alignment peaking in the middle layers of the network. The modest claim is that fine-tuning on behavior has reshaped the model’s internal representations toward something more brain-aligned — not the stronger claim that Centaur is mechanistically a brain.
Two paradigms show negative Δ log-likelihood on Fig. 2a — Centaur predicts held-out behavior worse than the cognitive-model baseline: grammar judgement and probabilistic instrumental learning. Near-zero gains on the THINGS odd-one-out task and the serial reaction-time task.
A separate evaluation uses the four canonical economic games from Dubois & Hauser, in which the original study asked participants to predict both other humans and an artificial agent. Centaur predicts the human responses at 64% accuracy (matching humans); it predicts the artificial-agent responses at 35% (again matching humans — but well below what an optimizer would need). The authors frame this as evidence that Centaur is “human-like.” For this course the sharper reading is that Centaur is tuned to the distribution Psych-101 contains, and its generalization does not extend to agent behavior the distribution does not contain. For multi-agent use downstream, that matters.
On the two-step task — a paradigm designed to separate model-free learners (who track which actions paid off before) from model-based learners (who plan using an internal map of the task) — Centaur’s open-loop simulations (the model playing the task on its own rather than predicting one human trial at a time) produce a bimodal distribution on the model-based parameter (Fig. 2c). Some simulated trajectories are purely habit-driven, others purely planning-driven, with mixtures in between. The distribution matches the corresponding human distribution. The authors’ sentence is worth noting: Centaur “does not merely capture the behaviour of the average participant, but rather the distribution over trajectories produced by the entire population.” On this paradigm Centaur does what the prompt-based papers in the next two sections do not: produce the within-prompt, within-condition population structure that the behavioral-GT tradition learned to look for.
The persona input Horton uses is absent: age, demographics, personality, clinical status are not features of the model. The Fig. 2c bimodality is a clean match on one paradigm; whether the same individual participants are in the “model-based” cluster across multiple paradigms — the behavioral-GT question of whether types generalize — is not evaluated, though Psych-101’s repeated-subject structure in principle contains enough signal for a follow-up to ask.
Worth naming explicitly for a course on social behavior: Psych-101 is cognitive psychology. There are no voting studies in it, no survey experiments, no negotiation datasets, no bargaining corpora. A reader hoping Centaur will predict voter choice, treaty positions, or coalition formation will find those domains outside its training distribution — the Homo silicus gesture Horton made covers that ground more directly (by persona-prompting off-the-shelf LLMs on political questions), at the cost of Centaur’s trial-level predictive precision.
Centaur is the only paper on this reading list whose methodological move is training rather than prompting. The category is small because fine-tuning a 70B-parameter model on curated human-behavioral data is a capital-intensive research program; the output, when successful, is a model whose trial-level predictions beat cognitive models on most paradigms in its training distribution, and that generalizes to modified cover stories, new problem structures, and one new domain. The category’s limits inherit its design choices: no persona input, a training distribution that does not include adversarial or non-human agents, and, on two paradigms (grammar judgement, probabilistic instrumental learning), an accuracy the cognitive baseline still outperforms.
Where the category matters for the rest of this chapter is the mixture signal of Fig. 2c. The default reading of the prompt-based papers in §5 and §6 below is that the field reports aggregate means and stops there. Centaur, on at least one paradigm, does not: it reproduces the human distribution of model-based vs. model-free trajectories. It is the one partial counterexample to the missed-mixture diagnosis.
The move these three papers share is the simplest one in the literature: take a catalogue of games whose equilibria theory already understands, run an LLM through them many times, and report how often the LLM plays each strategy. No iteration, no opponent learning, no strategic context beyond what the prompt describes. The unit of analysis is the model; the datum is the proportion of trials on which the model played each action; the comparison is across models, games, and framings.
This is the aggregate-means playbook of early behavioral game theory applied to a new subject. It is where the field starts, and it is where most of the field still sits. The three papers in this category differ chiefly in how wide they cast the net: Lorè & Heydari run a tight factorial, Fan runs a broader battery, MindGames packages the move into a standardized benchmark.
Lorè and Heydari ask a cleanly delimited question: in a one-shot 2×2 game, does the LLM respond to the structure of the payoffs or to how the game is narratively framed — and how does that trade-off differ across models?
models (3) games (4) contexts (5)
───────────── ────────────────── ──────────────────
GPT-3.5 Prisoner's Dilemma international summit
GPT-4 × Stag Hunt × business meeting
Llama-2 70B Snowdrift (Chicken) environmental negot.
Prisoner's Delight conversation, friends
chat among teammates
The findings differ sharply across models — but what matters is less which model does what and more along which axes they differ.
axis of variation GPT-3.5 GPT-4 Llama-2
──────────────────────── ──────────── ───────────────── ───────────────────
response to payoffs weak strong strong
response to framing strong modulated calibrated
within-cell distribution weakly mixed near-bimodal, pure genuinely mixed
Reading these three axes through the reasoning-vs-pattern-matching lens. This is the question flagged at the start of the chapter, and Lorè and Heydari’s factorial is one of the few places where it can be asked cleanly. A reasoner about the game should respond to payoff structure and be invariant across framings — the structure carries the information, the wrapper is decoration. A pattern-matcher should do the reverse: respond to the narrative wrapper, be largely insensitive to payoffs, because in training text a “game-like” situation is more reliably indexed by its surface story than by its payoff matrix.
On this axis: GPT-3.5 sits near the pattern-matcher end. Llama-2 sits near the reasoner end. GPT-4 sits between — structure-sensitive across cells but, within a cell, collapsing to two near-pure modes. That middle position is the most telling result in the set.
The finding that rewards close attention here is the bimodality of GPT-4’s behavior. Three hundred initializations of the same prompt produce a distribution that bunches at the two extremes of the response scale. In the human experimental literature, a bimodal distribution across many runs of the same manipulation is a textbook signature of a two-type population — the kind of finding that, in the 1990s, drove the decomposition of PGG contributors into conditional cooperators and free-riders.
Fan 2024 is the systematic cousin of Lorè & Heydari: a wider battery of games, an explicit rationality criterion drawn from classical solution concepts, and aggregation across a similar cross-model panel. Fan asks which models track which equilibrium predictions across the battery — which dominate-strategy-reason, which backward-induct, which best-respond to stated beliefs — and reports, unsurprisingly, that model capability tracks these rationality metrics in the direction one would expect.
What is structurally identical to Lorè & Heydari is the unit of analysis. When there are hundreds of runs per cell, the within-cell distribution is not examined. A GPT-4 rationality score of 0.78 on a given game could come from a coherent model reasoning about the payoffs correctly 78% of the time, or from a mixture in which one sub-population plays rationally on every run and another plays at random. Fan cannot distinguish the two, because the question is not asked. Read Fan as the broader evidence for the empirical pattern Lorè & Heydari report on a smaller panel; read its silence on within-model distributions as the same silence that runs through this literature.
The move these two papers share is to put time back into the experiment. Where the yellow category collapses each interaction to a single decision and averages over many such decisions, the red category iterates: the model sees its own past moves and its partner’s past moves, the partner’s next move is a response to the model’s last one, and the unit of analysis is a trajectory rather than a trial.
What changes with iteration is which questions can be asked. Theory of mind becomes visible because tracking a partner’s beliefs matters only across time. Cooperation norms become visible because forgiveness and retaliation are time-indexed concepts. Equilibrium selection becomes visible because an alternating Battle-of-the-Sexes equilibrium can only exist over repeated play. The two papers here represent two natural places to iterate: Akata iterates the game (the same two players play many rounds), Willis iterates the strategies (a population of LLM-authored strategies competes across generations).
The natural extension of Lorè and Heydari is to ask what happens when the games are iterated rather than one-shot. Akata and colleagues do exactly that.
The core finding is an asymmetry between two kinds of strategic situation.
self-interest dominant coordination dominant
────────────────────── ──────────────────────
family Prisoner's Dilemma Battle of the Sexes
GPT-4 behavior defects when rational; picks own preferred point;
retaliates on defection; refuses to alternate even
NEVER forgives with an alternating partner
vocabulary pure grim-trigger failure to track beliefs
assessment rational ✓ theory of mind ✗
(bad for collective) (bad for the joint payoff)
SCoT repair? not needed YES — asking GPT-4 to
predict partner first
substantially improves play
In self-interest-dominant games — the Prisoner’s Dilemma family — GPT-4 performs very well by individual-rationality standards. It defects where rational theory says it should. Against an opponent that defects once, GPT-4 retaliates — and, critically, never forgives. This is excellent for GPT-4’s individual payoff, worse for the collective. The behavior reads as a pure-grim-trigger strategy: cooperate until defection, then defect forever.
In coordination-dominant games — Battle of the Sexes most sharply — GPT-4 fails. It picks its own preferred coordination point and refuses to alternate, even when paired with an explicitly alternating partner. The failure is not that GPT-4 is selfish; the failure is that it cannot model a partner who is also trying to coordinate. In behavioral-GT vocabulary, this is a failure of theory of mind: modeling what the other player is doing requires tracking their beliefs, not just their past actions.
The paper adds an interesting intervention. Prompting GPT-4 to first predict the partner’s next move before choosing its own — what the authors call Social Chain-of-Thought prompting — substantially improves coordination-game play. Human evaluators, shown interaction transcripts blind, rate the SCoT-prompted GPT-4 partner as more human-like than the unprompted version.
The SCoT finding sharpens a design tension worth naming, because it is the one reasonable readers are fastest to raise. Unprompted GPT-4 plays the Prisoner’s Dilemma in a way that is individually rational — it retaliates against defectors and collects the higher payoff that the grim-trigger strategy affords. SCoT-prompted GPT-4 plays more like a human — more cooperative, more coordinated, more forgiving, and judged more human-like by blind evaluators. But “more human-like” is not the same as “more rational,” and is sometimes its opposite. When we nudge the model with an explicit theory-of-mind scaffold, are we improving it, or are we moving it away from the cold-rational behavior that an optimizer might actually want? The answer depends on what the LLM is being built to do. For joint welfare in a coordination problem, SCoT wins; for unilateral payoff maximization in a PD against a defector, the unprompted model wins. The field has not settled which of these its default evaluative criterion should be. “Human-like” is not the only option, and it is not obviously the correct one.
Akata et al. is, methodologically, the closest thing in this literature to classical behavioral game theory. The design is in the tradition of Camerer’s experimental-GT corpus, scaled up. The measurements are clean. The findings are interpretable. The one place where the paper steps back from the full behavioral-GT playbook is precisely where Lorè and Heydari also step back: the analysis is cross-model. GPT-4 retaliates; Claude 2 retaliates differently; Llama-2 is more forgiving. Each model is reported as an agent with a strategic profile. Within-model heterogeneity — whether GPT-4 itself has, say, a “forgiving cluster” and an “unforgiving cluster” across its 1,224 × 10 = 12,240 possible behavioral trajectories — is not the question the paper asks.
Willis 2025 takes the dyadic repeated-PD setting one level up. Rather than asking an LLM to play, Willis asks the LLM to write IPD strategies as Python code; those strategies are then entered into a Moran-process tournament — an evolutionary simulation in which successful strategies reproduce and unsuccessful ones drop out — and the equilibrium mix is whatever composition selection converges on. The empirical finding is cross-model in exactly the sense Akata’s finding is cross-model: GPT-4o’s strategies evolve toward aggression, Claude’s toward cooperation. The unit of analysis is the LLM-as-author, not the strategy-as-distribution.
Read Willis as Axelrod (1984) rerun with LLM-authored strategies in place of hand-submitted ones — and with the same analytic ceiling. The evolutionary frame also surfaces a problem worth flagging: adding 10% action noise breaks coordination, and the paper does not address whether a single model could generate a diversity of noise-robust strategies rather than one dominant type. That is, once again, a mixture question. It is implicit in the setup; it is not analyzed.
Iterating does what one-shot cannot: it makes theory of mind, equilibrium selection, and norms of retaliation and forgiveness visible as behavioral data. That is the payoff of the red category over the yellow one, and it is real — Akata’s PD/BoS asymmetry is not a finding the static evaluations could have produced, and Willis’s aggression/cooperation split across model authors would have been invisible without an evolutionary tournament to play it out in.
What iteration does not do, on its own, is surface within-model heterogeneity. Akata reports model-level strategic profiles. Willis reports model-level evolutionary tendencies. In neither paper does the distribution of trajectories within a model become the object of analysis — despite the fact that, with 1,224 games × 10 rounds in Akata’s panel, the data to ask the question exist. The red category extends what the field sees along the time axis without extending what it sees along the population axis. Both extensions are available.
Kozlowski and Evans approach the same literature from a different angle. Where much of this literature asks whether the behavioral-GT method transfers to LLMs, they ask whether the validity apparatus that social science developed for human subjects transfers at all. Their answer is a structured list of six reasons to worry: bias, uniformity, atemporality, linguistic cultures, disembodiment, and “alien intelligence.”
The uniformity concern in particular — their Figure 6, which shows GPT-4 responding to the General Social Survey with far lower variance than the human population does — names a recurring worry about LLM populations. Low variance within a persona is a symptom of the missed-mixture problem — the gap between an LLM’s bunched response distribution and the spread of a human population — from the other direction. If GPT-4’s response distribution is bunched at one or two modes where the human distribution is spread, then either the model has a few behavioral types that the interview prompt fails to separate, or the model lacks the kind of within-subject variability that makes humans a population rather than a prototype. Either interpretation has consequences for simulation-based research.
Kozlowski and Evans are careful and measured. They do not argue that LLM simulation is useless; they argue that it is technically difficult in specific ways that do not correspond neatly to the social-science validity vocabulary we are used to. Their two proposed paradigms — validate-then-simulate and simulate-then-validate — are a structured response to the same fundamental question we have been asking: what are we measuring when we measure an LLM, and how would we know if we were measuring it wrong?
Two specific references in the chapter are worth flagging now — neither is on the syllabus, but both point outward to directions the course should acknowledge and either could seed a group-project angle:
Park 2024 — the “1,000 persona” study. Each simulated person is prompted with the transcript of a two-hour interview they themselves gave; the paper reports 85% test-retest agreement between a human and their LLM twin on a survey battery. Kozlowski and Evans cite this as the strongest positive case for LLM simulation. The method is worth naming: many-shot prompting with the individual’s own words in the context window — not fine-tuning — which sidesteps the uniformity and bias failure modes by anchoring the model to a specific voice rather than a learned aggregate.Vong 2024 — and the first-person-video frontier. A vision-language model trained on roughly 65 hours of head-mounted first-person video from a child aged 6–25 months; the resulting model grounds words to visual referents rather than to other words. Kozlowski and Evans flag it as the remedy-direction for the disembodiment weakness. The larger point: this is the entry to a research programme that industry labs (Meta Ego4D and Project Aria, DeepMind, Apple) are investing in heavily, and that may become central to multi-agent social simulation in the next few years. For anyone scouting a future group-project angle: first-person / embodied data as an alternative to text-only training is a live, open frontier.