Mind the Gaps: Mixture-of-Minds for Human Simulation

The paper, authored by Pranav Dahiya of Semilattice (arXiv:2608.06115v1, August 2026), introduces Anacreon, an audience simulation model designed to predict individual-level responses within a narrow, well-specified domain. The paper frames the work around the persistent individual–aggregate gap in human behaviour prediction: "Statistical methods succeed at the level of the mass but falter at the level of the individual. Large language model simulators inherit this gap. They recover a population's central tendencies while flattening its heterogeneity, and they carry social biases and prompt brittleness that distort individual predictions. Anacreon addresses this by learning an authorship embedding that separates individuals, clustering a real qualitative corpus around seed people, and training a dedicated adapter for each cluster—a mixture of minds"—on a Gemma 4 12B base. The model reaches a state-of-the-art ordinal alignment of 0.775 on an external survey, which the paper describes as the individual-level accuracy measure on which the field has converged.

The introduction traces the history of aggregate prediction, from John Graunt's 1662 life tables, through Quetelet's l'homme moyen in 1835, to scientific opinion polling exemplified by the 1936 Literary Digest failure and Gallup's successful quota sample. The paper then describes the Simulmatics Corporation's People-Machine (1959) as a cautionary tale: Its confidence eventually outran its evidence... Its overreach prefigures a failure of modern LLM-based simulation. The paper argues that "The wider the target, the less any prediction can be grounded against real data. Existing LLM simulators run headlong into this limit. They repeat the overreach of 'the People-Machine' when they ask one model to stand in for any population on any question."

A key motivation is that RLHF harms simulation fidelity: "To become useful assistants, these models undergo extensive reinforcement learning from human feedback, which sharply reduces the diversity of their outputs. A better assistant is a worse human simulator, since it answers as one agreeable average rather than as the many distinct people a population contains."

The paper reviews prior work on individual-level simulation: Argyle et al.'s silicon samples (aggregate-level algorithmic fidelity), Park et al.'s interview-grounded generative agents (85% of test–retest consistency), and the digital-twin literature. Toubia et al.'s Twin-2K-500 reached 0.717 individual-level accuracy but expressed markedly less response variance than humans. The mega-study follow-up (Peng et al.) found twin responses correlated only weakly with their human counterparts (average r ≈ 0.20) and fell back toward the population mean, with a large share of regression coefficients differ[ing] in size or even flipp[ing] in sign. Bias literature (Wang et al.) identifies two failure modes: misportrayal (models reflecting outsider views of a group) and flattening (ignoring within-group heterogeneity).

Anacreon treats records as (stimulus, response) pairs and constructs a chain-of-emotion (c), a short trace of the emotional states and appraisals the stimulus evokes, before producing the response. The loss is the expected negative log-likelihood with pθ(c, r s) = pθ(c s) pθ(r s, c).

Clustering. A subset of data-rich individuals forms the seed population. A transformer is trained with the contrastive authorship-representation objective (InfoNCE) of Rivera-Soto et al., pulling two text samples from the same individual together and pushing different individuals apart. The trained model embeds the full corpus, and Every record is assigned to its nearest seed individual, which partitions the corpus into clusters. Each cluster is temporally split: The newest 20% of records is held out for testing, and the remaining 80% is used for training.

Data Augmentation. An LLM-as-judge extracts demographics and psychological traits, using a proprietary harness [to] reduce hallucination, and a majority vote across multiple runs stabilises each attribute. The harness favours accuracy over coverage. Survey questions are mined from text, creating multiple-choice and ordinal questions with known answers; these are mined separately on the train and test splits, so that test answers cannot leak into training. Because LLM responses show a positivity bias—skewing toward agree/strongly agree in 90% of cases—every question with a non-neutral response was included twice in the training and test sets, once framed positively and once negatively, addressing both sycophancy and measuring susceptibility to prompt brittleness.

Chain-of-Emotion. An agent harness takes the stimulus and selects the harvested attributes relevant to it. Only a few attributes matter for any given stimulus, not all, consistent with the constructive-choice view of consumer behaviour. The chain maps the evolution of emotional state from stimulus to response.

Training. A separate high-rank LoRA adapter is trained per cluster on Gemma 4 12B via QLoRA. At rank 64, the 12B base yields about 260M trainable parameters per cluster. Summed over the 420 trained clusters, Anacreon learnt about 110B parameters in total. Response order is randomly shuffled each epoch to ensure the model is not learning to predict responses purely based on their position. Incoherent clusters that fail to converge are pruned; At inference the 400 retained per-cluster models unpack to 4.9T parameters.

Checkpoint Selection. The best checkpoint maximizes θ⋆ = arg max α·Amc(θ) + β·Aord(θ) − γ·PPL(θ), balancing multiple-choice top-1 accuracy, ordinal alignment, and chain-of-emotion perplexity.

The evaluation domain is a B2B audience model built for a Semilattice customer. The modelled individuals are SME merchants who take physical, in-person card payments.

Bias. The signed prediction error pools to a distribution concentrated at zero but retain[ing] a small positive mean of +0.437—a positivity bias that remains but is much reduced relative to a base LLM. Within-cluster standard deviation of the signed error has a median near 1.85 scale positions. Demographic composition of the full modelled population and of the 50 lowest-scoring clusters is similar: The weakest clusters are therefore not concentrated in any single demographic group. This suggests that poor performance reflects the data quality within a cluster rather than the demographic it represents.

Survey Accuracy. Ordinal alignment (Aord = 1 − MAD/R) has a median of 0.775, exceeding prior systems: Twin-2K-500 at 0.717, SocratesQwen-14B at 0.740, and the digital-twin mega-study at 0.748. This is the state of the art on this measure. Top-1 accuracy on multiple-choice questions reaches a median of 0.679, comparable to the 0.657 exact-match accuracy of the interview-grounded agents of Park et al. (whose human test–retest baseline was 0.795).

Chain-of-Emotion Perplexity. The per-cluster perplexity distribution is unimodal and concentrated, with a median near 6.0, indicating the models reliably reconstruct the emotional chain that leads to each answer within the narrow, structured space of emotional states.

Narrow domain. Anacreon deliberately targets narrow populations; the training data act like a belief network... well supported near the training domain and sparse away from it. Conclusions extend to topics adjacent to the training data... They cannot be trusted far beyond it. At inference, a query can be checked for support in the training data, so an answer is offered only when the records provide some basis for it.

Bias and variance. While the positivity bias is small, model variance is high. The paper contextualizes this against human reliability: By classical test theory, a reliability of 0.5 corresponds to a test-retest standard deviation near 1.27 scale positions on a five-point item. Anacreon's median standard deviation of about 1.85 is above the human level but only moderately so, and since A faithful simulator should reproduce this human spread rather than collapse to a point, so the target is the human level, not zero variance. However, models at the tail end are only marginally better than a random guesser.

From individuals to aggregates. The paper proposes applying post-stratification weighting to Anacreon's per-cluster models, "reweighting the per-cluster models so that their pooled prediction matches a target population. However, this calibration is only sound when the individual predictions are sound. Individual fidelity is therefore the primary objective, and distributional alignment follows from it, not the other way around."

Ethics. The design focuses on groups rather than persons: "Records are pooled into clusters, and every model is trained on a cluster rather than on any single individual. The seed individuals only anchor the clusters... Anacreon is therefore built to describe how a segment tends to respond, not to identify, track, or manipulate a specific individual."

The paper claims several contributions: the seed-population clustering with authorship embeddings, the mixture-of-minds architecture, harvesting of attributes and survey responses from public text, shuffled options to reduce prompt brittleness, and positive/negative framing to reduce positivity bias. The measurable outcome is the 0.775 ordinal alignment—the state of the art among the systems that report the measure—with a small residual positive bias. The closing reflection: "For three and a half centuries, prediction has paid out at the level of the mass and faltered at the level of the person. Anacreon does not dissolve that boundary, but it does move it. By grounding a mixture of minds in what individuals reveal about themselves, it simulates a narrow population well enough to set the state of the art on individual-level accuracy. Bias and variance remain, and population-level fidelity is still to be shown."

• Replace a single monolithic simulator with a mixture-of-minds architecture. Instead of one LLM standing in for everyone, train a separate high-rank LoRA adapter per subpopulation cluster. This reduces the individual–aggregate gap, avoids flattening heterogeneity, and improves per-person ordinal alignment from 0.717–0.748 to 0.775.

• Learn authorship embeddings with a contrastive InfoNCE objective. Pull text samples from the same person together and push different people apart, then cluster all records by nearest seed individual. This partitions the corpus into behaviourally coherent groups rather than demographic stereotype buckets, preserving within-group diversity.

• Generate a chain-of-emotion before each response. Model pθ(c∣s)·pθ(r∣s,c), where c is a short trace of emotional states and appraisals evoked by the stimulus. This makes responses more accurate, interpretable, and grounded in a constructed decision process.

• Harvest person-specific attributes and survey questions from public text with an LLM-as-judge. Use a hallucination-reducing harness and majority voting across runs; mine questions separately from train and test splits to prevent answer leakage. This produces grounded multiple-choice and ordinal labels with known answers.

• Shuffle response order every training epoch. This prevents the model from learning position-based patterns and reduces prompt brittleness when answer options are reordered at inference.

• Train on both positive and negative framings of every non-neutral question. This counters the base LLM's positivity bias (90% skew toward agree/strongly agree) and explicitly measures sycophancy susceptibility.

• Select checkpoints using a combined objective. Maximise α·Amc + β·Aord − γ·PPL, balancing multiple-choice top-1 accuracy, ordinal alignment, and chain-of-emotion perplexity, rather than relying on a single metric.

• Prune incoherent clusters and check query support at inference. Drop clusters that fail to converge; for new queries, verify that nearby training records exist before generating an answer. This limits the system to domains where evidence exists and prevents People-Machine-style overreach.

• Apply post-stratification weighting only after individual-level fidelity is established. Reweight per-cluster models to match a target population for aggregate estimates, while keeping individual predictions as the primary objective.

• Predict individual-level survey responses within a narrow, well-specified population—e.g., SME merchants taking in-person card payments—with state-of-the-art ordinal alignment of 0.775 and multiple-choice top-1 accuracy of 0.679.

• Preserve realistic response variance instead of collapsing to the population mean; its per-person spread approaches human test–retest reliability rather than zero variance.

• Generate a transparent chain-of-emotion for every predicted response, showing the emotional progression that led to the answer, with a median perplexity near 6.0.

• Resist prompt brittleness: reordering answer options or reframing a question positively/negatively no longer flips the prediction arbitrarily.

• Detect out-of-domain queries and abstain when the training records do not support an answer, rather than confidently simulating any population.

• Produce segment-level and population-level estimates by post-stratifying per-cluster models, but only after validating that each individual-level prediction is sound.

• Avoid demographic misportrayal: the weakest clusters are not concentrated in any single demographic group, so poor performance reflects data quality within a cluster, not bias against a group.

Predicting how a population will answer a new question is a long-standing goal. Statistical methods succeed at the level of the mass but falter at the level of the individual. Large language model simulators inherit this gap. They recover a population's central tendencies while flattening its heterogeneity, and they carry social biases and prompt brittleness that distort individual predictions. This paper introduces Anacreon, an audience simulation model that targets the individual level within a narrow, well-specified domain. Anacreon learns an authorship embedding that separates individuals, clusters a real qualitative corpus around seed people, and trains a dedicated adapter for each cluster, a mixture of minds, on a Gemma~4 12B base. It harvests demographics, psychological traits, and survey responses from public text, and augments each record with a chain-of-emotion. It reduces prompt brittleness by shuffling response options and reduces positive bias by balancing the training distribution. On a large, externally sourced survey, Anacreon reaches a state-of-the-art ordinal alignment of 0.775, the individual-level accuracy measure on which the field has converged, with a small residual bias. The work is a step toward drawing aggregate insight from faithfully simulated individuals.