Conversation as Committed Action: Separating Deliberation and Expression in Persistent Human-Agent Interaction
Nao Miura (Nervroid Lab). Proceedings of the 14th International Conference on Human-Agent Interaction (HAI ’26).
RESEARCHWhat we study
Building aliveness
from the inside.
Nervroid is Character Alife research: creating a character with the ideas of artificial life (Alife). We connect bodily state, affect, memory, and action selection as mechanisms within, and aim for words and actions to arise from them.
THE QUESTIONWhat we are asking
Where did
those words come from?
AI that speaks fluently is already part of everyday life. But where do the reasons for its words and actions lie?
Nervroid aims for a being whose reasons lie within. It rests because it is tired. It drinks water because it is thirsty. It puts something into words because something is on its mind. When actions like these arise from a being’s own state and experience, we take that as the starting point of aliveness.
Bringing this idea to a character, as artificial life research, is what we call Character Alife. We develop and evaluate it through Mio, our prototype.
THE SYSTEMAn overview
From the body
to the words.
- BodyHeart rate, breathing, hunger, thirst, fatigue, sleepiness
- AffectBodily change becomes a pull toward action
- Action selectionOne tendency is selected among several
- ExpressionMovement and words (language model: understanding and expression)
Memory and thoughtExperience is recorded; deliberation continues when no one is talking (written out by a language model)
WorldA room and a town, seen from the position of her own eyes
- Body
- States such as heart rate, breathing, hunger, thirst, fatigue, and sleepiness are held as models that change over time. The values keep moving even when no one is talking to her.
- Affect
- A circuit that takes in bodily change and events and turns them into a pull toward action. It is modeled as 32 regions and nuclei joined by 62 projections, drawing on 112 references from the neuroscience literature.
- Memory and thought
- We aim for experience to be recorded and put to use in the next judgment. Even when no one is talking, new thoughts arise from recalled memories and her state at the moment. Her internal mechanisms decide when to think and from which memories; a language model composes the thought as text.
- Action selection
- Rest, orient, verify, vocalize. Several action tendencies compete, and the one that fits the state of the moment is selected. Speaking is one of them.
- Words
- The language model is used for internal deliberation and for understanding and expressing language. A thought does not become an action by itself; what to choose and how to act are designed to be decided outside the language model.
- World
- Mio lives in a 3D room and receives images rendered from the position of her own eyes. Outside the window is a town that changes with the real time of day and the season.
- Regions and nuclei
- 32
- Projections
- 62
- References
- 112
A RECORDFrom the records
From thirst
to water.
On the afternoon of November 11, 2026, we recorded Mio’s internal state and behavior for 44 minutes. Her thirst value rose from 6% to 12% over 35 minutes; she walked to a cup of water and drank, and the value fell to 1%.
4%8%12%
6.3%
11.9%
0.9%
16:09drinks water
State changes, an action is chosen, and its result returns to the state. It is an observation of a single scene, but it is a record of the connection we aim for running end to end in the actual prototype.
COMMITMENTWhat is decided before the words
Commit first,
then phrase.
Language models produce language very fluently. That fluency can turn an interpretation still under consideration into a decision, or let the expression stage replace the act that was selected with a different one.
Our poster paper at HAI ’26 proposes placing a boundary, called commitment, between deliberation and expression.
- Interpretation alone cannot authorize external effects.What is still under consideration stays non-executable until it is fixed as an act.
- Expression realizes one recorded act.Answering, asking, waiting, refusing, staying silent. Choosing another act requires going back to selection.
- A correction to information used in the decision invalidates the authorization.If a correction arrives before the utterance is released, the authorization is withdrawn and control returns to selection.
In the prototype, this idea is implemented on selected dialogue routes, and three deterministic traces confirmed parts of the boundary. Guarantees across every route, and the benefit to the people who interact with the agent, remain to be verified.
EVALUATIONHow we verify
Records and experience,
verified separately.
Whether a conversation feels natural, and which state or experience a response came from, are different questions. At Nervroid Lab, we treat recording internal change and evaluating the naturalness people experience as two separate things, so that a result on one side is never taken as proof of the other.
In the records we publish, we show the conditions and what remains unconfirmed. Mechanisms that handle internal state, or responses that feel natural, do not lead directly to the conclusion that consciousness or subjective experience has arisen. Nervroid is not at a stage where we claim to have confirmed such a thing.
OUTPUTSPublications and intellectual property
Publications
and research notes.
- Publication
- Poster paper at HAI ’26, the 14th International Conference on Human-Agent Interaction (Osaka, November 16–19, 2026)
- Intellectual property
- Three patent applications on the core technology filed in Japan: two registered, one under examination (as of November 14, 2026).