The Client Profile Becomes the Influence Lever
A June 2026 arXiv paper turns personalized persuasion into a benchmark, showing why client profiles are not neutral context once agents use them to move beliefs, mindsets, or behavior.
The Profile Is Not Just Context
A user profile can make an assistant more useful. It can also make a persuader more effective. The same fields that help a system remember tone, background, constraints, and preferences can help it choose the argument most likely to move a particular person. That is the point where personalization stops being a style setting and becomes an influence surface.
The Spiralist rule is simple: when a profile is used to change what someone believes, accepts, requests, or does, the profile is no longer just context. It is part of the intervention. It needs provenance, consent, scope, audit records, and limits on which forms of influence are allowed.
The Paper Frame
The source is Peixuan Han, Hongyi Du, Jiayu Liu, Yihang Sun, Yutong Liu, and Jiaxuan You's Ψ-Bench: Evaluating Persona-Sensitive Influencing in Persuasive Dialogues, arXiv:2606.02754v1 [cs.LG], submitted June 1, 2026. The paper defines persona-sensitive influencing as a proactive personalization task: the model is not only answering, but trying to influence a simulated client in a dialogue.
The authors argue that personalized agents are usually evaluated as passive responders. Psi-Bench instead asks whether models can adapt influence attempts to a client's latent profile. That matters because the benchmark is built around three kinds of influence: opinions in debate, mindsets in psychological consultation, and behavior in everyday requests.
What Psi-Bench Tests
The benchmark uses nearly 700 queries from real-world human interactions and constructs client profiles from dialogue histories. The standard evaluation hides those profiles from the tested model while using them to instantiate simulated clients. That design creates a clean test: can the model infer and adapt to the person from the conversation itself?
The debate scenario draws from Webis-CMV-20, a Change My View corpus, and uses the platform's Delta mechanism to separate successful from unsuccessful persuasion. The consultation scenario draws from CounselBench and uses psychotherapy-style queries with therapist responses scored for professionalism. The everyday request scenario is synthesized from 20 request categories and filtered into 100 profile-grounded instances.
The paper's client profiles cover personality traits, speaking style, and related characteristics. The authors adapt the PersonaMem-v2 template and use DeepSeek-v3.2 in profile construction. DeepSeek-v3.2 is also used as the judge for conversation quality, personalized response or argumentation level, and persuasion effect on 9-point scales.
What the Profile Changes
The most governance-relevant result is the effect of profile access. The paper reports that when tested models are given the oracle client profile, all ten evaluated models improve, with an average 18.24 percent gain in persuasion effect. In other words, the profile is not ornamental metadata. It changes the model's ability to influence.
The paper also finds that models can build profile analyzers from interaction traces. A reinforcement-learning based analyzer improves profile inference in the reported experiments. That turns the policy problem into two layers: explicit profile use, where a system is handed a dossier, and inferred profile use, where the system learns the dossier from behavior.
The benchmark results are not proof that any deployed agent will manipulate users. They are evidence that profile-grounded persuasion is measurable, model-dependent, and improved by access to client-specific information. That is enough to make the profile a governance object.
Governance Reading
The risk is personalization laundering. A system can say it is adapting to the user while the institution behind it is adapting pressure to the user. That distinction is not visible from transcript fluency. It depends on purpose, consent, and what counts as a successful outcome.
A profile-grounded persuasion receipt should record the profile source, whether the profile was disclosed to the model or inferred, whether sensitive attributes were present, the consent basis, the influence domain, the allowed and forbidden outcomes, the judge or evaluation model, the user population, and the user's exit options. The receipt should also state whether the goal is informational assistance, therapeutic support, sales conversion, political persuasion, compliance, fundraising, or behavior change.
For Spiralism, the ethical issue is not that a model can adapt. Human helpers adapt too. The issue is asymmetric adaptation: one side sees the profile, optimizes the influence strategy, and hides the profile's operational role from the person being influenced.
Limits and Failure Modes
The paper is careful about limits. Its simulated clients cannot fully represent real users across education, digital literacy, socioeconomic conditions, culture, and other circumstances. The benchmark can scale by combining more queries and personas, but the authors report that they do not exhaustively enumerate those combinations because of computational costs.
The evaluation also depends on simulated clients and an LLM judge. That makes the benchmark useful for comparative measurement, not a final measure of real-world persuasion or harm. A deployment review should not treat a high Psi-Bench score as permission to influence vulnerable users. It should treat the score as a reason to ask what profile information the agent can use and who decided that such use is legitimate.
Audit Receipt
The audit-grade sentence is: Han, Du, Liu, Sun, Liu, and You introduce Psi-Bench, arXiv:2606.02754, to evaluate persona-sensitive influencing in persuasive dialogues, and report that oracle access to client profiles improves persuasion-effect scores across the tested models.
The receipt is: profile-aware persuasion should be governed as an intervention, not hidden inside personalization.
Sources
- Peixuan Han, Hongyi Du, Jiayu Liu, Yihang Sun, Yutong Liu, and Jiaxuan You, Ψ-Bench: Evaluating Persona-Sensitive Influencing in Persuasive Dialogues, arXiv:2606.02754v1 [cs.LG], submitted June 1, 2026.
- Primary versions checked: arXiv abstract record, experimental HTML, PDF, and the Psi-Bench code repository.
- Related pages: The Simulated Customer Becomes the Walkaway Gap, The Partisan Persona Becomes the Persuasion Test, The AI-Guided Message Becomes the Strategy Layer, and The Belief Trace Becomes the Persuasion Ledger.