Position paper

PsychEvolve: Subtle Cues Must Keep Earning Their Influence

…if full-duplex agents are to reason psychologically and evolve.

Agents that see and hear us while they speak could notice what we show but never say. We argue that what they notice should count only as far as it has proved true for this person.

Junjie Chen1*, Wanru Li1*, Yanbin Hao1†, Yi Tan2†, Tianle Niu3, Yulang Chen4, Pengyu Liu1, Linfeng Zhang3, Meng Wang1

1HFUT · 2NTU · 3SJTU · 4HKUST
*Equal contribution · †Corresponding authors

A person in a cozy armchair talking with a friendly avatar, with small floating symbols for voice, gaze, time, and speech

He that has eyes to see and ears to hear may convince himself that no mortal can keep a secret. If his lips are silent, he chatters with his finger-tips; betrayal oozes out of him at every pore.

— Sigmund Freud, Fragment of an Analysis of a Case of Hysteria (1905)

Freud's observer needed only eyes and ears. Full-duplex agents now have both, and they use them while the other person is still talking. A client may slow down, glance away, or rub their fingers as a sensitive topic approaches. Could an agent notice, and should it believe what it notices?

How a subtle cue keeps earning its influence for one client
How a subtle cue keeps earning its influence for one client: observe, compare readings, respond provisionally, reassess with the agent's own action in view, revise with an expert, and re-earn the reading in later sessions. Click to enlarge.
but here is the catch

Better perception alone will not close the gap

1

Subtle cues are ambiguous

A perfectly detected finger rub may reflect anxiety, concentration, or simply cold hands.

Why?Facial configurations and micro-gestures do not map reliably onto particular emotions in isolation (Barrett et al., 2019; Liu et al., 2021).
2

The agent changes what it observes

A full-duplex agent nods and reacts while you speak, and your next move partly answers its own behavior.

Why?Listeners' responses shape how speakers tell a story (Bavelas et al., 2000), even when the listener is a robot (Park et al., 2017). Reading that as evidence about you alone is a mistake.
3

Memory turns passing readings into traits

One tense afternoon can quietly become "this client is anxious" in a stored profile.

Why?Memory-augmented agents over-apply stored memories (Hu et al., 2026) and miss changes in a user (Jiang et al., 2025), much as people infer lasting traits from situational behavior (Gilbert & Malone, 1995).

Watching more closely does not make a weak cue strong. People judging lies from truths are right only 54% of the time (Bond & DePaulo, 2006), and a lens-model meta-analysis traces this limit to the weakness of the cues, not to the observers (Hartwig & Bond, 2011). The more an agent sees, the more confidently it may read secrets that are not there.

our position

If full-duplex agents are to reason psychologically and evolve, subtle cues must keep earning their influence.

Following Brunswik's lens model (Brunswik, 1956), an inference is accurate only when the observer relies on each cue as much as that cue is actually valid. For an agent in conversation, validity is personal and situated.

R1 · personal
Validity belongs to this person.
A glance away may signal avoidance in one person and recall in another (Mischel & Shoda, 1995).
R2 · not self-made
Your own echo is not evidence.
Behavior the agent itself elicited does not count as confirmation.
R3 · revisable
Influence is lost as easily as gained.
When later evidence or the person's own account disagrees, the cue's weight falls.
how did we get here?

Psychology inspired AI. AI may in turn help psychology.

Conversational agents have grown from text to full-duplex audiovisual interaction, each stage borrowing from psychology. Every stage widened the channels an agent can read, yet, to our knowledge, the cues it reads have remained largely coarse. Pick a stage.

Evolutionary roadmap of conversational agents
is the field moving toward subtle cues?

Each ingredient is growing. Together, they remain rare.

A coarse keyword count from OpenAlex. It suggests that research on empathetic dialogue, subtle cues, and user memory has each grown fast, while work combining them for psychological support remains rare. Click any row to open the exact search on OpenAlex.

Log scale. Counts from OpenAlex title and abstract search, retrieved 13 November 2026, and may drift as the index updates. *2026 is a partial year. Rows starting with “+” add one condition to the row above. Keyword counts trade precision for coverage, so read them as orders of magnitude, not a census.

a blueprint, not yet a system

A blueprint on three clocks

To make the position concrete, we sketch PsychEvolve, a blueprint spanning data, architecture, and psychological grounding. We use it to raise open questions rather than to claim solutions. It would run on three clocks.

⚡ within a turn

Fast full-duplex pathway

Holds the conversation and drives an avatar or humanoid robot.

It listens while speaking, coordinates voice, face, and body, and responds provisionally before the evidence is complete. It reads the latest entries of a shared memory before each response and records its own actions so they can be discounted later.
🧠 across turns

Slow agent team

Weighs cues against this client's baseline.

Collaborating agents work asynchronously on the shared memory, some per channel of cues, others consulting psychological theories such as appraisal, the lens model, and if-then signatures. They pass back hypotheses that carry their evidence and expiry, not facts.
🌙 between sessions

Expert in the loop

The slowest member of the team.

A professional may review stored readings and correct how much each cue is worth for a client. The expert never steers the live conversation, and each correction updates validity in memory.
The PsychEvolve framework
Psychology-oriented data (left) feed a fast full-duplex model, a shared memory of hypotheses, and a slower agent team that experts may join (center). Evaluation (right) rests mainly on practical use.
who decides what the agent believes?

An expert-in-the-loop avatar platform

Inspired by AVATAR therapy (Leff et al., 2013; Garety et al., 2024), we illustrate how the blueprint could be put into practice with the design of a platform for psychological support. The design illustrates PsychEvolve rather than validating it, supports professionals, and never makes clinical diagnoses on its own.

Client

Talks with the avatar, and can confirm, correct, or withdraw what it has understood. Chooses what to share and accepts or declines every proposal.

Expert

Sees only assigned clients who agreed to share. Corrects the agent's understanding with stated reasons, between sessions only.

Administrator

Manages accounts, approved methods, releases, and an emergency stop. Sees session metadata, never content.

Roles in the envisioned platform
Accepted changes take effect only in the next session, so nothing changes silently while the client is talking.

A look at the early prototype

An early design prototype, not evaluated with clients or experts. Names and accounts are test data.

alternative views

Objections worth taking seriously

A position is only as strong as the objections it survives. We consider five and say where we think each one is right.

Why not simply ask the person?

Self-report is the most direct evidence, and R3 lets it overturn a reading. Yet people cannot always say what they feel and talk less freely when asked too often. Cues and questions are partners. A cue that has earned some influence can tell the agent when a gentle question is worth asking, and the answer then shows how much that cue deserved.

Can a cue's validity be learned for one person?

Subtle cues are weak (Hartwig & Bond, 2011) and one client offers few sessions. Bayesian accounts of affect and dialogue suggest one route, starting from population priors and updating them with each person's evidence. Slow learning is itself a safeguard, since a cue should keep little influence while evidence is scarce.

Should agents read unspoken cues at all?

Inferring emotion from faces is contested (Barrett et al., 2019), and continuous capture could become surveillance. These concerns argue against confident inference from cues, which is exactly what our position rejects. We therefore treat consent, client ownership of stored readings, the right to withdraw them, and a non-diagnostic scope as requirements of the design rather than as optional features.

Should AI offer psychological support at all?

Some argue that support for vulnerable people should come only from humans, and we share much of this concern. Trials of chatbots report benefits but leave safety and long-term effects open (Li et al., 2023; Heinz et al., 2025). Our position does not claim that agents should replace professionals. It claims that if agents are used in this setting, as they increasingly are, their beliefs about a person must be earned, visible, and correctable.

Won't scale make this unnecessary?

Larger end-to-end models may learn to handle cues well without explicit memory, and our requirements do not depend on modules. Yet more data of the same kind may not help. Recorded conversations do not separate the user's state from the listener's influence, so a model trained to imitate them has little reason to learn R2. We therefore see R1 to R3 as tests that any model architecture should pass.

a broader question that we believe deserves the community’s attention

How can we build psychologically grounded, fine-grained, multimodal, efficient, continually evolving, self-correcting, and full-duplex conversational agents in which subtle cues keep earning their influence, to better understand and support human emotions and psychological well-being?

We suggest one test for any answer:

Does learning from subtle cues make an agent more accurate about the person,
or simply more confident?
read the paper

The full paper, right here

/ –
100%
Loading the paper…

Cite

@article{psychevolve2026,
  title  = {PsychEvolve: Subtle Cues Must Keep Earning Their Influence
            if Full-Duplex Agents Are to Reason Psychologically and Evolve},
  author = {Chen, Junjie and Li, Wanru and Hao, Yanbin and Tan, Yi and Niu, Tianle
            and Chen, Yulang and Liu, Pengyu and Zhang, Linfeng and Wang, Meng},
  year   = {2026}
}