Modelling Human Action

An actor model requires a defined prediction target. Whether an analyst publishes a revised estimate after an earnings announcement, whether a regulator opens an investigation and whether a supplier changes a delivery commitment are different tasks. They have different observation windows, available actions and evidential requirements. Treating them all as instances of a general response obscures the distinctions needed for evaluation. The first modelling decision is therefore to specify whose behaviour is being predicted, which actions count and when the observation period ends.

Three questions often appear together but should be assessed separately: whether an actor acts, which action the actor takes and how that action is expressed. Predicting the text of a reply conditional on a reply occurring does not establish the probability that the actor will reply. Similarly, ranking several supplied actions does not demonstrate that a system can discover the relevant alternatives without assistance. These distinctions determine both the information a model may use and the conclusions that its score can support.

The unit of analysis also matters. A person, an organisation and a representative population group cannot be substituted without changing the model. An institutional announcement may reflect an approval process involving several people, legal obligations and resource constraints. A group-level estimate may describe an average response that no particular individual exhibits. Throughout this essay, an actor means the explicitly defined unit whose observable action is the target; it does not imply that all such units share a psychological decision process.

We distinguish predictive performance from explanation. A model may predict a statement using recurring language without identifying why the speaker made it. Conversely, an account of incentives may be substantively informative while producing poor probability estimates. Establishing a mechanism requires evidence beyond successful prediction. Our concern here is narrower: how to construct and evaluate conditional predictions of observable behaviour, while retaining enough provenance to investigate their failure.

A persona is a representation of an actor, usually expressed through attributes, a biography or a description of preferences. Such representations can provide relevant constraints, but their detail is not an independent measure of validity. A generated biography may contain plausible statements that have no evidential basis. If subsequent predictions treat those statements as observations, the system introduces an unrecorded inference between its data and its output. The resulting explanation may be internally consistent while depending on an unsupported premise.

Evidence about persona conditioning must be interpreted at the level of the task studied. Schwager et al. (2026) evaluate conditioned comment prediction: generating a response to a supplied stimulus using previous comments or a generated biography. Their experiments distinguish linguistic form from semantic grounding and examine how conditioning interacts with fine-tuning. These findings concern response generation under the evaluated conditions. They do not directly measure whether a person will participate, which information they encounter or which non-linguistic action they select.

The practical implication is to test the contribution of each representation. A comparison might hold the prediction model and current context fixed while varying whether it receives documented attributes, relevant behavioural examples or both. The comparison must also account for input length and retrieval effort. Otherwise, an apparent advantage of history may reflect additional information or a larger resource budget rather than the superiority of one representational form. The appropriate claim is conditional on what the experiment actually holds constant.

Self-reports provide a different form of grounding. In the revised study by Park et al. (2026), agents constructed from interviews, surveys or both predict held-out responses from 1,052 participants. The authors compare agent performance with participant test–retest consistency. This normalisation helps interpret predictions of measurements that are themselves variable. It is not a universal upper bound on behavioural prediction, nor does success on the studied instruments establish accuracy for institutional decisions, public-market actions or unfamiliar populations.

Behavioural traces, self-reports and contextual records answer different questions. A trace records something an actor did through an observed channel. A self-report records what the actor said about an experience, attitude or intention. A contextual record describes the circumstances in which an action occurred. None should silently replace another. A stated intention may be informative without being fulfilled; an observed transaction may be real while its motive remains unknown. Evidence should retain its original type when incorporated into an actor representation.

Relevance is conditional on the prediction problem. Previous responses to earnings revisions may be useful for modelling an analyst after a new financial disclosure. Unrelated public statements may add little even when they are recent or numerous. Comparable history should be selected using criteria that can be applied before the target outcome is known. Choosing precedents because they resemble the eventual action would make the retrieval stage itself a source of hindsight leakage.

A historical information boundary requires more than filtering publication dates. Sources can be corrected, expanded or reinterpreted after the event. A present-day profile may include an appointment that had not occurred at the forecast time. Retrieval systems may return retrospective summaries whose wording reveals the outcome. A defensible replay records the version of each source available at the cutoff, distinguishes contemporaneous material from later reconstruction and identifies information that cannot be recovered reliably.

The model itself presents a further boundary. Restricting retrieved evidence does not remove outcomes that may already be represented in its training data. For retrospective studies, model release dates and documented training cutoffs provide useful context but do not prove the absence of contamination. Prospective predictions, saved before the outcome occurs, provide a stronger design for this purpose. Historical replay remains useful for diagnosing software and reasoning failures when its evidential limits are reported.

Evidence provenance also constrains the strength of an actor description. Several records may repeat a single source, and an inferred attribute may depend on a small number of exceptional episodes. A research representation should distinguish directly documented facts, recurring observations and tentative interpretations. It should also preserve counterexamples. Otherwise a summary can appear more certain than the records from which it was constructed, particularly when the same generated description is reused across several forecasting tasks.

A public event is not necessarily an observed event for a particular actor. Information may be available through a platform, a professional network or a news service without being attended to. A behavioural model should therefore distinguish documented exposure from proxies for exposure and from an absence of exposure data. Following an account is a proxy for possible receipt; it is not evidence that every subsequent message was read. The distinction becomes especially important when the target is a response to one specific item.

When exposure is uncertain, the probability of a response combines at least two questions: whether the actor encounters the information and how the actor responds conditional on encountering it. A system may model these stages explicitly or estimate the combined outcome directly. Either approach needs an evaluation target consistent with the available data. Training only on people who responded and applying the result to everyone potentially exposed changes the population on which the estimate is being used.

Non-action must also be defined operationally. No public statement within twenty-four hours is an observable outcome if the relevant channels were monitored adequately. It is not equivalent to no private discussion, no change of belief or no action through another channel. Missing records and genuine non-action must remain separate wherever possible. Observation quality, the duration of follow-up and the channels covered are therefore part of the outcome definition rather than incidental data-cleaning details.

Including non-action in the candidate set does not by itself correct participation bias. The model may still assign excessive probability to visible activity because the examples used for training or retrieval disproportionately contain responses. Assessment should compare predicted participation with observed participation in an appropriately sampled population, then examine action selection among participants. This separates an incorrect activity rate from an incorrect account of what active actors do.

A useful analytical decomposition separates candidate discovery, probability estimation and expression. Candidate discovery identifies the actions that the model will consider. Probability estimation assigns mass to those alternatives given the evidence. Expression produces any accompanying text. This decomposition is a proposed design discipline, not a claim that every actor or software system executes three independent stages. Its purpose is to make distinct sources of error observable.

Candidate discovery creates a coverage problem. If the eventual action is absent from the candidate set, better ranking cannot recover it. An open-ended generator may broaden coverage, but superficially different formulations can describe the same action. Evaluation therefore needs a rule for matching equivalent actions and identifying genuine omissions. That rule should be fixed independently of the predicted ranking. Adding an unspecified alternative can reserve probability mass, but does not supply the missing action itself.

Probabilities must correspond to a coherent outcome space. Mutually exclusive and exhaustive alternatives can form a categorical distribution whose probabilities sum to one. Events that can occur together require a different representation. For example, publishing a research note and revising a price target may overlap; treating them as competing outcomes would impose a false constraint. The granularity of the action vocabulary should follow the question being answered and the observations available for resolving it.

Generating an explanation after an estimate can improve inspectability, but an explanation is not proof of the computation that produced the estimate. It should identify supporting observations, counterevidence and missing information without inventing private motives. Where a model renders an imagined statement, that statement remains simulated output. It must not re-enter later stages as evidence that the actor actually spoke or acted.

The quantity of available history and the quality of its use are separate variables. Li et al. (2025) introduce FineRob, with 78.6k question–answer records derived from histories of 1,866 social-media users. Their task separates the object, type and content of behaviour. They report performance degradation as memory accumulates in their evaluated settings and investigate a method for using detailed memories more effectively. This supports examining retrieval and memory organisation; it does not establish that longer histories are generally detrimental.

Memory evaluation should distinguish insufficient evidence from a failure to retrieve or use available evidence. One diagnostic compares predictions from the ordinary retrieval policy with predictions from a carefully reviewed evidence subset. Another removes individual evidence categories while preserving the rest of the input. Such comparisons can identify where useful information is lost. They cannot justify deployment changes unless the selection procedure and resource requirements can be reproduced on new cases.

Summaries introduce an additional representation layer. A statement that an actor usually responds in a particular way should retain a connection to the observations, exceptions and period from which it was inferred. Otherwise, repeated summarisation can turn a tentative interpretation into an apparently stable fact. Updating a summary should preserve changes in the underlying evidence rather than merely reinforce earlier model-generated wording.

Behaviour can also change. A new institutional role, incentive structure or information environment may reduce the relevance of older observations. Recency weighting is one possible response, but it can discard rare precedents that remain important. Evaluation should examine different history windows and retrieval policies on later observations, rather than selecting whichever history best explains a known outcome. Report performance separately where a documented change in role or environment makes transfer particularly uncertain.

Consider a hypothetical equity analyst following NVIDIA immediately after a scheduled earnings release. The target is whether the analyst publishes a price-target revision within the following two trading days. This example specifies a modelling problem; it does not report an actual analyst, forecast or company event. The observable target is a published revision, not an inferred change in conviction. The deadline and eligible publication channels must be defined before the release.

Before forecasting, the researcher assembles prior dated revisions, the analyst's documented coverage mandate, the existing estimate and the disclosure available at the cutoff. Previous earnings episodes provide candidate precedents. A later interview explaining the analyst's decision is excluded from the prediction input. If access to the earnings call is unknown, exposure is recorded as uncertain rather than inferred from the analyst's professional role alone.

The task can first distinguish a numerical revision from no observed numerical revision. Conditional on a revision, it can distinguish an increase from a decrease. An unchanged target accompanied by a new research note is a different communication outcome and should not be counted as a revision. If communication is the intended target, the categories and resolution rule must be changed before forecasting. Resolving this ambiguity in advance prevents the evaluator from changing the meaning of success after reading the resulting note.

A useful comparison includes the analyst's historical revision frequency, a rule based on the earnings surprise and an evidence-conditioned model. Each uses the same cutoff and outcome definition. When a prediction fails, the investigation asks whether a relevant precedent was missing, an exposure assumption was incorrect, the action set was incomplete or the probability was poorly calibrated. The example illustrates why a single correct narrative cannot establish the quality of the underlying actor model.

Evaluation begins with the intended deployment population. Holding out later observations tests temporal transfer; holding out actors tests transfer to different individuals or institutions; holding out event families tests dependence on familiar circumstances. These designs answer different questions. A model evaluated on later actions of well-observed actors should not be described as validated for new actors with little history. Where several predictions concern one actor or event, uncertainty estimates should account for their dependence.

Sequential prediction introduces a further distinction. A system may predict the next action while receiving the true preceding history, or generate an entire sequence in which later predictions depend on its own earlier outputs. The latter allows errors to accumulate. Rui Li et al. (2025) study behavioural chains extracted from literary works, with recognition and generation tasks. Their benchmark is informative about sequence consistency, but fictional trajectories and model-assisted assessment are not equivalent to prospective observations of real people.

No single metric covers these problems. Candidate coverage assesses whether the observed action was available. Classification measures assess selection among defined alternatives. Proper probability scores assess the distribution assigned to outcomes. For binary events, Brier loss is the squared difference between the forecast probability and the observed zero-or-one outcome. Its interpretation and the distinction between calibration and discrimination are developed in Calibration and Collective Judgment, drawing on Gneiting and Raftery (2007).

Error analysis should retain difficult and unobserved cases rather than discard them silently. A system that predicts only well-documented actors may achieve a lower average loss while covering a narrower population. Report the number eligible, attempted, returned and resolved, together with the handling of missing predictions. Examine performance by action frequency, evidence availability and observation quality when the data support those comparisons. Small subgroups warrant uncertainty rather than confident rankings.

Class imbalance makes a simple baseline especially important. If most eligible actors do not revise an estimate during a short window, always predicting no revision may achieve high classification accuracy. A useful actor model must show what it adds beyond that prevalence. Evaluation can examine participation probabilities and performance on the less frequent action separately, while preserving the natural event rate in the main test set. Artificially balancing the test population changes the interpretation of probabilities unless that change is accounted for explicitly.

A representation ablation should also address the strongest simpler explanation for its result. If actor identity improves performance, the model may be learning a stable activity rate rather than a context-sensitive response. Compare identity alone, current context alone and their combination under the same evaluation design. Evidence that the combination improves later predictions would support the value of conditioning on both. It would still leave open whether a detailed persona, rather than a smaller set of observed attributes, is necessary to obtain that improvement.

Some applications justify a simpler representation even when a richer model performs better on average. The additional information may be unavailable at deployment, expensive to obtain or unsuitable for the intended use. Comparing error, coverage and evidence requirements together makes this trade-off explicit. The scientific question is not whether maximum detail can reconstruct a familiar actor, but which reproducible information improves prediction for the population and decision under study. A model that needs retrospective interviews to explain every case may answer a different question from a prospective forecaster.

Actor models can support the selection of scenarios, identification of information gaps and systematic comparison of behavioural hypotheses. Their usefulness is greatest when the action space and observation process are sufficiently explicit to permit correction. An accurate estimate for one task does not transfer automatically to a different decision, population or horizon. Integrating actors into a larger simulation adds further assumptions about interaction, discussed in Simulating Society.

The argument developed here is that actor fidelity should be evaluated through a sequence of testable relationships: evidence to representation, representation to action distribution, and distribution to observed outcome. Failures at these boundaries require different remedies. More descriptive detail may help one boundary while introducing unsupported inference at another. The appropriate research objective is to identify which information and modelling choices improve prediction under specified conditions, and where the evidence remains insufficient to support a conclusion.

These limits also govern responsible interpretation. Observable public behaviour does not license assertions about a person's private psychological state. Where consented self-reports are used, their collection setting and permitted purpose constrain subsequent reuse. An inferred motive should remain an inference, and a generated statement should remain attributable to the model rather than the person represented. Maintaining that distinction protects the validity of the research as well as the people whose records make it possible.

1. Schwager, N., Münker, S., Plum, A. and Rettinger, A. (2026). Towards Simulating Social Media Users with LLMs: Evaluating the Operational Validity of Conditioned Comment Prediction. WASSA, 208–221. doi:10.18653/v1/2026.wassa-1.16.
2. Park, J. S. et al. (2026). LLM Agents Grounded in Self-Reports Enable General-Purpose Simulation of Individuals. arXiv:2411.10109, version 3. Preprint first submitted in 2024.
3. Li, K., Dai, C., Zhou, W. and Hu, S. (2025). LLM-Based Behavior Prediction for Social Media Users with Continuous Memory. IJCNLP–AACL, 459–474. doi:10.18653/v1/2025.ijcnlp-long.27.
4. Li, R. et al. (2025). How Far are LLMs from Being Our Digital Twins? A Benchmark for Persona-Based Behavior Chain Simulation. Findings of ACL, 15738–15763. doi:10.18653/v1/2025.findings-acl.813.
5. Gneiting, T. and Raftery, A. E. (2007). Strictly Proper Scoring Rules, Prediction, and Estimation. Journal of the American Statistical Association, 102(477), 359–378.