When Simulated Users Always Do What They Are Told: Hidden Effort in Dual-Control Agent Evaluation

Some AI assistants guide people through actions the person must take themselves: restarting a phone, reseating a SIM card, toggling a setting. Benchmarks for these assistants replace the person with a language-model "user simulator." In this setting the assistant cannot see the user's device. It only knows what the user says they did. Real people sometimes say "done" without doing anything, skip the tedious step, or do the wrong thing and report success. Language-model simulators almost never do. They carry out every instruction faithfully and describe the result truthfully. The benchmark therefore rewards assistants that never check whether an instruction was actually followed, and that is the skill real deployments most need. We propose to measure this gap directly with human participants whose device actions are logged. We will then build a user simulator whose unseen actions follow a fitted cost model, and test whether it changes which assistants look best and what they learn when trained against it.

The idea joins three lines of work. Studies of simulation fidelity show that language-model users make tasks too easy. Work on multi-agent partial observability shows that hidden, costly effort is where cooperation breaks down. Dual-control benchmarks put a real hidden-action channel between assistant and user, but nobody has checked how humans behave in that channel. Our claim is specific: in dual-control tasks, much of the simulation-to-reality gap sits in the user's unobserved actions rather than in conversational style. Fixing the style, as recent persona work does, will not close it.

Core Claim

Economists call this situation "moral hazard": one party pays a private cost for effort another party cannot see. The assistant is the party that depends on the effort and cannot observe it. The user bears the cost of each local action and controls whether the action happens and what gets reported. $\tau^2$-Bench formalizes telecom troubleshooting as a decentralized partially observable decision problem. The assistant holds billing and account tools, and the user holds phone tools. Its tool-constrained simulator grounds the user in real device state, which cut simulator errors from 40% to 16%. It still assumes the simulated user executes instructions in good faith, and it was never compared with humans.

Three studies show simulators are too easy, and all of them stay in settings where the user holds no tools. Mind the Sim2Real Gap audited 31 simulators against 495 human traces on single-control $\tau$-bench and found inflated success. ConvApparel found that prompted simulators fail to adapt realistically when the assistant behaves badly. Persona Policies evolves programs that produce impatient, vague, or reluctant personas on retail and airline tasks. None of the three can observe a gap between what a user does and what a user says, because none has a user action channel.

Moral Hazard in Multi-Agent Language Models studies the reverse direction. There, a language agent must decide whether to pay a cost to help a peer, in a stylized agent-only game. We take its hidden-action lens and turn it on the simulated human. We fit the structure of human shirking from logged data rather than assuming it.

The new element is a calibrated hidden-effort model of the simulated user. For a requested action $a$, the simulated user executes it with probability

$$P(\text{execute} \mid a, h) = \sigma\big(\beta\,[\,v(h) - c_a - \lambda\, t\,]\big).$$

Here $c_a$ is a fitted effort cost specific to the action type: a SIM reseat costs more than a toggle. The term $v(h)$ is the value the user perceives given the dialogue so far $h$; it rises when the assistant explains why the step matters. The term $\lambda t$ captures patience lost after $t$ prior requests. Finally, $\beta$ sets how sharply the user's choices respond to these quantities, and $\sigma$ is the logistic function.

A separate honesty parameter sets the probability that a skipped or failed action is reported as done. All of these quantities are fit by maximum likelihood to human logs. The language model only phrases the user's turn. It does not decide whether the action happened. This separation is deliberate. ConvApparel found a trade-off in which explicit persona instructions reduce naturalness while free-form generation drifts. Keeping the decision in the tool layer avoids asking the language model to be both natural and accurately lazy.

Research Program

The first stage is a human audit of the dual-control channel that no one has run. We will rebuild the user side of $\tau^2$-Bench telecom as a mock phone interface. The interface logs every real state change while participants talk to a fixed set of assistants. We plan roughly 150 participants on about 40 telecom tasks. That is smaller than the 451-person cohort of Mind the Sim2Real Gap, but enough for action-level statistics because each task contains several user actions.

Following that study's design, participants are paid for completing the task rather than for compliance, which keeps realistic effort incentives. Each user action yields a ground-truth triple: what was requested, what was executed, and what was reported. We will compute four quantities: the gap between reports and actions (claims of completion with no state change), skip rates by action type, rates of executing the wrong action, and abandonment. We will run identical tasks with the default $\tau^2$ simulator. This first stage alone settles whether the hidden-action gap exists. If simulators and humans already match here, the rest of the program stops, and the audit is reported as a negative result.

The second stage fits the cost model above on half the participants and validates it on the rest, by held-out log-likelihood and by agreement with the human report-action statistics.

The third stage is the actual test. We take a population of assistants and evaluate each one under four user conditions: humans, the default simulator, persona-conditioned simulators, and our hidden-effort simulator. We then ask which simulator ranks the assistants the way humans do. The assistants span frontier models and open models that differ in how often they verify. We also include training. An open assistant is trained by Reinforce Leave-One-Out reinforcement learning, the method used in Moral Hazard in Multi-Agent Language Models, once against each simulator, and each trained assistant is evaluated on held-out human sessions.

The mechanism predicts concrete behaviors. Assistants trained against the hidden-effort user should learn to check device state through their own account-side tools, to ask users to read back screen contents, and to explain why costly steps matter. Assistants trained against the default simulator should not learn any of these.

Baselines

The naive reference point is the default tool-constrained user simulator from $\tau^2$-Bench, run as released. That paper reports pass@1 in the dual-control telecom domain of 34% for GPT-4o and 49% for Claude 3.7 Sonnet, along with a 16% simulator error rate. No reported result was found on Lacuna for human-user success on $\tau^2$ telecom, so the human gap here is unknown. The closest analogue is single-control $\tau$-bench, where Mind the Sim2Real Gap reports simulator success up to 77.8% against a human baseline of 63.6%. Our simulator beats this baseline if it shrinks the absolute difference between simulated and human success on held-out tasks, and if its ranking of assistants agrees with the human ranking (Kendall rank correlation) by a margin outside the bootstrap interval.

The strongest existing method is Persona Policies. Its evolved users were judged human 80.4% of the time, about twice the rate of baseline simulators, and assistants trained with it were 17% more successful in relative terms. It is the strongest competitor because it already targets the gap between cooperative simulators and real users and has improved downstream training. We will port its generator-evolution procedure to telecom. It gets the same base model and the same number of fitness evaluations as our fitting procedure. Crucially, it also gets the same human logs in its fitness function, so the comparison isolates the method rather than access to data. We beat it if our simulator matches human report-action statistics more closely and yields assistants with higher human-evaluated success, even where the two are similar on judged human-likeness.

Three ablations isolate the key ingredient. In the first, all actions share a single fitted skip rate, removing the cost structure; this tests whether action-specific costs matter. In the second, honesty is set to one, so users skip but never misreport; this tests whether false reports drive the effect. In the third, the device state is visible: the assistant sees the user's phone. If our account is right, most of the gap between the simulators and humans should disappear when the device state is visible. That would confirm partial observability, not style, as the cause.

Why not just…?

Why not just prompt the simulator to be a lazy, sometimes dishonest user? This is a real threat. We include it as a baseline: a frontier-model simulator prompted to skip tedious steps and occasionally claim success falsely, with the prompt tuned on the same calibration half of the human data. The prior evidence is suggestive but not decisive. Prompted simulators had average human-likeness near zero and failed counterfactual validation in ConvApparel, and prompt-only users hallucinated far more often in $\tau^2$-Bench. Our prediction is narrower. A prompted user's shirking will not follow the cost ordering across action types, and it will not respond to the assistant's explanations, because nothing ties the behavior to a cost. Stage two measures both directly. If the tuned prompt matches human action-level statistics as well as the fitted model, we will report that, and the contribution reduces to the audit plus a validated prompt.

Why not just run Persona Policies on telecom? It may well recover some shirking if its fitness function rewards it. That is why we give it the same human logs. Its search space is generators of role-play instructions that shape what users say. Nothing in it ties the device state to the report. The ablation with honesty fixed at one and the report-action metric will show whether that grounding matters. If evolved personas match the human statistics, the finding is that the evolutionary search should simply be pointed at the action channel, which is still a useful result.

Why not just make the assistant always verify? An assistant that checks every step is a fix for assistants, not for evaluation. It does not tell us whether benchmarks rank assistants correctly. It also has a cost: $\tau^2$-Bench found over-specified workflow instructions actively hurt performance, and success approaches zero beyond seven required actions, so extra verification turns are not free. We include a verify-everything assistant in the evaluated population as a probe. Under the default simulator it should look worse than or equal to a non-verifying assistant. Under humans and our simulator it should look better. If that reversal fails to appear, our premise is wrong.

Proposed Experiments

The decisive checks run in order. The first is the audit. Is the rate of false completion reports among humans on $\tau^2$ telecom clearly above the default simulator's near-zero rate? Does it vary by action type? If not, the program ends.

The second is fit. Does the hidden-effort model predict held-out human execution and reporting better than the uniform-rate and prompted baselines, measured by log-likelihood and by distance between the distributions of report-action statistics?

The third is ranking. Across at least eight assistants, does our simulator's success ranking agree with the human ranking more closely than the default, Persona Policies, and prompted simulators? The verify-everything probe should flip sign between the default and hidden-effort conditions.

The fourth is transfer. Does the open assistant trained against the hidden-effort simulator achieve higher pass@1 on human sessions than copies trained against the other simulators with matched reinforcement-learning steps? Do its transcripts show more account-side checks and more read-back requests?

The fifth is the visible-state ablation. Does most of the gap between simulated and human success disappear when the assistant can see the device state?

A positive result on the ranking and transfer checks would show that faithful dialogue style is not enough in dual-control settings. In that case, simulators must model what users actually do when no one is watching.

Workshop authors to connect with