Social simulations built from language-model agents assume that the persona written into each agent's prompt shapes what that agent does. Recent evidence suggests it often does not. Disentangling Models from Personas in Heterogeneous LLM Simulations shows that in a mixed social network, the engagement an agent receives depends more on which base model runs it than on its assigned persona. What nobody has measured is where along the agent's decision the persona stops mattering. It could be lost in the free-text reasoning, in the conclusion the agent draws from that reasoning, or in the step from conclusion to executed action.
We propose to locate that loss point. Agents receive personas whose effect on the correct action can be computed exactly; for example, a stated risk tolerance shifts the right poker call or the right negotiation concession by a known amount. We then trace each decision through three checkable stages: what the agent reasons, what it concludes, and what it does. Finally, we transplant intermediate outputs between base models, for instance handing model B the reasoning that model A wrote. This tells us whether the base model's own habits re-enter at the reasoning stage or the action stage. The output is a stage-by-stage account of how a persona specification survives, or fails to survive, the trip into behavior, measured separately for each base model.
Core Claim
Our hypothesis is that personas are expressed in the stated reasoning but are overridden when the agent commits to a conclusion. The override pulls the conclusion toward the action the base model would take with no persona at all, which we call its "default prior." The executed action then faithfully follows that drifted conclusion.
If this is right, it reconciles two literatures that currently disagree about where the inference-action gap (a model's failure to act on what it has inferred) lives. Doing What They Say, Not What They Reason finds in poker that agents execute what they state, so the failure is upstream between reasoning and conclusion. Its rule misapplications for Claude Haiku 4.5 are "overwhelmingly risk-averse," and DeepSeek-Reasoner shows a different error mix. We read that model-specific skew as a candidate base-model prior. Other work places the gap downstream. Can Large Language Models Serve as Rational Players in Game Theory? finds GPT-4 predicts opponent moves but picks the wrong response. YC-Bench shows models writing a correct rule in their scratchpad and violating it on the next action.
Our claim is that both are instances of one phenomenon: the base model's prior winning a competition against context-specified constraints. The persona is one such constraint, and the stage where the prior wins depends on the model. This predicts that the persona-induced errors of a given base model point toward that model's no-persona action more often than chance. It also predicts that this "prior pull" is stable across tasks for a given model but differs between models. That would be the mechanism behind Roytburg and Ippolito's network-level base-model dominance.
What Is New Relative to Prior Work
Several papers motivate the question without answering it. Persona Prompting as a Lens on LLM Social Reasoning finds a rigid reasoning template, with high agreement across personas (Krippendorff's alpha of at least 0.80). It measures this on single-turn classification and never traces through to actions. OPeRA finds that removing persona information sometimes improves GPT-4.1's exact next-action accuracy on real shopping logs. Because it has no persona-specific reference action, it cannot say whether the persona was ignored or misapplied.
SimpleToM and How FaR Are Large Language Models From Agents with Theory-of-Mind? separate explicit inference from applied action. They do so in fixed vignettes, with no personas and no comparison across models of the internal pipeline.
Other work constrains the design. Position: Synthetic Persona Needs Explicit Grounding and Standardized Reporting argues that personas need stated provenance and enactment checks. We adopt this literally: every persona in our study is a grounded parameter with a declared behavioral implication, not a vivid backstory. Persona-Model Collapse in Emergent Misalignment shows that fine-tuning destabilizes persona maintenance. We therefore restrict ourselves to off-the-shelf models and treat fine-tuned variants only as a later extension.
Wang's poker harness supplies the method: machine-checkable conclusions, plus the warning that free-text parsing inflates apparent conclusion-to-action errors. We reuse that harness and extend it in three ways it does not cover: personas, multiple base models seated at the same table, and transplants between models.
Research Program
The environments are chosen so that a persona parameter changes the normatively correct action by a computable amount.
In Wang's Texas Hold'em harness, a persona specifies a risk attitude. The reference action becomes "call if equity exceeds pot odds plus a persona-specific margin," so a risk-seeking and a risk-averse persona have different correct calls on the same hand.
In the CaSiNo negotiation task (dividing food, water and firewood under private preferences), a persona specifies a reservation value and a concession style. The reference offer follows from the ledger-and-utility machinery of GENEGO.
In iterated matrix games from Position: Theory of Mind Benchmarks are Broken for Large Language Models, a persona specifies a cooperation disposition, and functional regret is computed against the persona-consistent best response.
Each persona follows Lu et al.'s template. It records where the parameter comes from, what population or stance it represents, and what claim it licenses.
At every decision we elicit four objects. The first is free-text reasoning. The second is a structured conclusion containing the stated inputs (equity, pot odds, partner's inferred preferences), the persona rule being applied, and the intended move. The third is the executed action. The fourth is the persona-specific reference action. Separately, each model is run on the same situations with no persona, which gives us its default action.
For each stage $s$ (reasoning content, conclusion, action), we fit a variance decomposition over a design that crosses base model $m$, persona $p$ and situation $x$. We report persona retention:
$$\rho_s = \frac{\sigma^2_{p,s}}{\sigma^2_{p,s} + \sigma^2_{m,s} + \sigma^2_{p\times m,s}},$$
where $\sigma^2_{p,s}$, $\sigma^2_{m,s}$ and $\sigma^2_{p\times m,s}$ are the variance at stage $s$ explained by persona, by base model, and by their interaction. A sharp drop in $\rho$ between two adjacent stages marks where the persona is lost.
For reasoning text, we measure the stage variable as the persona parameter recovered by a held-out classifier. For conclusions and actions, it is the implied parameter value inverted from the chosen move.
For the errors themselves, we define prior pull. This is the fraction of persona-inconsistent actions that lie closer to the model's own no-persona action than to the reference action. We compare it against the fraction expected if errors were symmetric around the reference.
The causal component is cross-model transplantation. A conclusion transplant gives model B the situation, the persona, and model A's structured conclusion, then asks B to act; this tests the conclusion-to-action link independently of who drew the conclusion. A reasoning transplant gives model B model A's free-text reasoning and asks B to conclude. If B's conclusions drift toward B's own default rather than A's reasoning, the prior acts at the conclusion stage and belongs to B. If B's conclusions follow A's reasoning faithfully, the loss happened inside A's reasoning.
These transplants tell us whether base-model identity enters as a property of the "reasoner" or of the "decider." This distinction matters for anyone building mixed-model populations.
The base models span capability tiers and reasoning styles named in the file: Claude Haiku 4.5 and DeepSeek-Reasoner (for continuity with Wang), GPT-4o, and Llama-3.1 at 8B and 70B. This range lets us test the base-model-bottleneck prediction that $\rho$ at the conclusion stage rises with capability. That prediction comes from scaffolding gains scaling with backbone strength in ReDel and CaPo, and from the competence collapse of small models in AgentSociety: Incentivizing Agentic Social Intelligence.
Multi-agent play uses mixed-model tables and negotiation pairs. This lets us ask whether prior pull changes when the opponent is a different model family, which is the heterogeneous condition Roytburg and Ippolito identify as driving dominance at scale.
Baselines
The naive reference point is the no-persona agent. Each model plays the same situations with no persona, and we score it against every persona's reference action. Any persona-conditioned condition whose actions are no closer to the persona reference than this is, operationally, ignoring the persona. The no-persona run also supplies the default action that the prior-pull metric needs, so it costs nothing extra.
The second baseline is the outcome-only variance decomposition in the style of Roytburg and Ippolito. It attributes variance in final actions to model versus persona and never looks inside the pipeline. This is the strongest existing approach to the "model versus persona" question. Our stage-wise analysis must show something it cannot, namely a measurable drop in $\rho$ between two specific stages that differs across models. If $\rho$ is flat across stages, the outcome-only analysis was sufficient and our decomposition adds nothing.
The third baseline is the strongest existing method for making agents act on persona-relevant constraints: structured elicitation with mechanical rule application. This combines Wang's "apply the rule mechanically" instruction with the explicit belief-before-action prompting that raised GPT-4 from 10% to 100% optimal action on the Ring-Network game in Can Large Language Models Serve as Rational Players in Game Theory?. For negotiation, the comparison point is the ProCoT prompt-only baseline, which GENEGO reports at 19.1 points against 22.0 for GENEGO on CaSiNo. For the poker misapplication rates, the file contains no reported number. Wang reports only that the instruction "roughly halves" Haiku's rate (no reported result found on Lacuna for the exact rates), so we re-measure them in the same harness.
All prompted conditions share the same situations, seeds, decoding settings and token budget. Each receives an equal prompt-tuning budget of a fixed number of prompt variants chosen on a held-out set of situations. "Beating" this baseline means one of two things. Either our diagnosis predicts which models and personas will remain persona-inconsistent after mechanical prompting, with prior pull significantly above the symmetric chance level. Or we show that a stage-targeted fix chosen from the transplant results closes more of the residual gap than mechanical prompting does at equal cost. That fix is routing the conclusion step to the model with the highest conclusion-stage $\rho$.
The key ablation removes the persona's computable implication. The persona is given as narrative only, with no stated parameter. This tests whether the structured, grounded personas we use are themselves what makes the gap measurable. It also tests Lu et al.'s claim that vivid ungrounded personas hide failures. A second ablation scrambles transplant sources, giving model B reasoning from a different situation. This confirms that the transplant effects come from content rather than formatting.
Why not just…?
Why not just compare final actions across models and personas, as the seed paper did? Outcome-level attribution can establish that the base model dominates, but not where it dominates. That location determines the fix: re-prompting the conclusion step, swapping the reasoning model, or abandoning persona-conditioned simulation for that model. The decisive check is whether $\rho$ differs significantly across stages within a model. If it does not, this alternative wins, and we will report that the stage view adds nothing.
Why not just instruct the model to apply its persona mechanically and output structured conclusions? This is a real threat, and it is our strongest baseline rather than a straw man. Wang finds that it halves Haiku's misapplications but does not remove them, and that DeepSeek-Reasoner's error mix differs. YC-Bench shows explicit scratchpad rules being violated over long horizons. If mechanical prompting drives prior pull to chance for all models, the honest conclusion is that the persona-action gap is an elicitation artifact. That would still be a useful result for simulation practice. We expect residual, model-specific pull, but the experiment is designed so that either outcome is reported.
Why not just use a stronger reasoning model and assume the gap disappears? The base-model-bottleneck literature predicts the gap shrinks with capability, and we test that directly across tiers. Even strong models leave large gaps on applied reasoning: SimpleToM reports o1-preview at 59.5% on behavior judgment. Social simulations also deliberately mix cheap and expensive models for scale. Knowing each model's stage-specific persona retention is what lets a practitioner decide which model may play which role.
Why not just compute actions with a solver from the persona parameters, as GENEGO's ledger does? In our games, that solver is the reference action, so it serves as an oracle upper bound rather than a competitor. Simulations use language-model agents precisely because most social actions have no solver. Decoding Echo Chambers shows numeric opinion models regress to the mean (polarization change around −0.2) where language agents entrench (+0.584). The point of the computable-reference games is to calibrate a diagnostic that can later be applied where no solver exists.
Proposed Experiments
The first experiment replicates Wang's poker finding without personas, confirming that conclusion-to-action fidelity is high once conclusions are structured. We then add risk-attitude personas and fit the stage-wise $\rho$. The decisive check is a drop in $\rho$ at the conclusion stage, combined with prior pull above the symmetric chance level for Haiku in the risk-averse direction. If instead persona-induced errors are symmetric around the reference, our prior-override hypothesis is wrong.
The second experiment runs conclusion and reasoning transplants in every ordered pair of the five models. The decisive check is whether a model's conclusions from transplanted reasoning drift toward its own default action, which would make the prior a property of the decider, or toward the source model's default, which would make it a property of the reasoner.
The third experiment repeats the analysis on CaSiNo with reservation-value personas and on the iterated matrix games with cooperation personas. This tests whether each model's prior pull is stable across domains, which would answer the file's open question of whether the upstream/downstream disagreement comes from the domain or the model. For calibration we report average negotiation score against GENEGO's 22.0 and ProCoT's 19.1 on CaSiNo, and functional regret against the persona-consistent best response in the matrix games.
The fourth experiment uses mixed-model tables to test whether prior pull strengthens as more model families are added. This mirrors the seed paper's finding that base-model effects grow with heterogeneity, and would connect our mechanism to their network-level result.
Finally, if a model's conclusion-stage retention predicts its persona consistency in a lightweight Roytburg-style network, we will have shown that a cheap game-based diagnostic forecasts which models will flatten persona diversity in an open-ended simulation. That is the result simulation builders would actually use.
Workshop authors to connect with
- Disentangling Models from Personas in Heterogeneous LLM Simulations: Dani Roytburg, Daphne Ippolito
- Doing What They Say, Not What They Reason: Locating the Faithfulness Gap in LLM Agents: Yufeng Wang
- Position: Synthetic Persona Needs Explicit Grounding and Standardized Reporting: Shirley Huang, Guanghui Min, Zonglin Di, Brihi Joshi, Muhammad Ahmed Mohsin, Keyang Xuan, Jiatong Li, Yunhan Zhao, Bing Hu, Zhiwei Zhang, Xupeng Chen, binghang lu, and 32 more
References
- Disentangling Models from Personas in Heterogeneous LLM Simulations
- Doing What They Say, Not What They Reason: Locating the Faithfulness Gap in LLM Agents
- Can Large Language Models Serve as Rational Players in Game Theory? A Systematic Analysis
- YC-Bench: Benchmarking AI Agents for Long-Term Planning and Consistent Execution
- Persona Prompting as a Lens on LLM Social Reasoning
- OPeRA: A Dataset of Observation, Persona, Rationale, and Action for Evaluating LLMs on Human Online Shopping Behavior Simulation
- SimpleToM: Exposing the Gap between Explicit ToM Inference and Implicit ToM Application in LLMs
- How FaR Are Large Language Models From Agents with Theory-of-Mind?
- Position: Synthetic Persona Needs Explicit Grounding and Standardized Reporting
- Persona-Model Collapse in Emergent Misalignment
- GENEGO: A Disciplined LLM Negotiation Agent with Pairwise Preference Inference
- Position: Theory of Mind Benchmarks are Broken for Large Language Models
- ReDel: A Toolkit for LLM-Powered Recursive Multi-Agent Systems
- CaPo: Cooperative Plan Optimization for Efficient Embodied Multi-Agent Cooperation
- AgentSociety: Incentivizing Agentic Social Intelligence
- Decoding Echo Chambers: LLM-Powered Simulations Revealing Polarization in Social Networks