Source-linked AI summary
Synthetic Linguistic Agency: How an Embodied Mortal Agent Learns Linguistic Affordances through Consequential Social Experience
Sixin Chen, Taizhou Chen
TL;DR
The paper asks how linguistic activity can become consequential for an artificial system’s continuing life rather than merely influence a human interaction. It operationalizes Synthetic Linguistic Agency, constructs a mortality-grounded EMA, and finds that bodily state, social history, partner behavior, and persistent bodily consequences shape linguistic choice and viability.
Problem
Fluent language models influence human decisions, but their exchanges do not enter a continuing, vulnerable life of their own.
Method
The paper translates embodiment, linguistic participation, and precariousness into SLA criteria and builds MGL-RL with an EMA to test their causal roles.
Results
Controlled interventions show that the three relations make separable causal contributions within one continuing artificial life, and six reported systems exhibit SLA under the operational definition.
Takeaways & Limitations
The EMA demonstrates that linguistic choices can connect partner responses to the future of the same artificial life.
Takeaways & Limitations
The evidence concerns one frozen configuration in a single scripted ecology, with variation across settings and alternative learning left for further study.
Abstract
from arXiv · showhide
Contemporary language models can converse fluently and influence human decisions, yet their exchanges do not enter a continuing, vulnerable life of their own. Linguistic-agency theory identifies this missing connection as linguistic agency and characterizes it through embodiment, linguistic participation, and precariousness: a body that acts and bears consequences, interaction that changes both agent and partner, and a future that can be sustained or lost. Two coordinated studies examine how this organization can appear in artificial systems. First, we translate these relations into inspectable criteria for Synthetic Linguistic Agency (SLA) and identify several existing SLA systems. Second, building on Homeostatically Regulated Reinforcement Learning, we develop a mortality-grounded linguistic-reinforcement-learning model and instantiate it in an Embodied Mortal Agent (EMA). The EMA learns how ways of speaking change a partner's willingness to protect it and chooses expressions by considering what those responses mean for its remaining life. Controlled experiments show that linguistic choices depend on the EMA's body and social history, change partner behavior, and adapt through experience with particular partners. When bodily consequences persist, linguistic choices alter the future of the same life; when the body is reset, their social effects remain but no longer shape continued viability. The resulting EMA exhibits SLA under our operational definition. This work motivates further research on synthetic empathy and strategic human-AI interaction: how artificial agents with persistent bodies, histories, and futures might develop and express empathy, and how people might care for, negotiate with, or govern them.
1 Introduction
The paper addresses the missing connection between linguistic expression and an artificial system’s continuing, vulnerable life. It operationalizes Synthetic Linguistic Agency and constructs an artificial individual to test embodiment, participation, and precariousness causally.
- Motivation: Fluent language models influence people, but their exchanges do not become part of a continuing life with bodily stakes.The paper contrasts linguistic systems lacking bodily stakes with fatigue-aware structures that regulate bodily stakes without linguistic participation.
- Motivation: Linguistic-agency theory links embodied activity, reciprocal social regulation, and consequential continuation through embodiment, participation, and precariousness.These requirements connect expression and interpretation with effects on the agent, partner, and interaction.
- Studies: The first study translates the three relations into inspectable criteria for Synthetic Linguistic Agency and applies them to representative artificial artifact families.This study asks which artificial systems exhibit linguistic agency.
- Studies: The second study constructs one artificial individual in which embodiment, linguistic participation, and precariousness are jointly realized, separately represented, and independently manipulable.This design supports tests of each relation’s causal contribution and their coupling within one continuing life.
- Construction: MGL-RL and the Embodied Mortal Agent connect strategic expression and anticipated social response to health and Reference-RUL through explicit linguistic-affordance learning.The construction builds on Homeostatically Regulated Reinforcement Learning.
- Experiments: Controlled interventions and ablations test whether the three requirements make separable, nonredundant causal contributions while remaining coupled through the same body and biography.The experiments intervene on body state, biography, reciprocal linguistic influence, and persistence of bodily consequences.
2 Linguistic agency and related artifacts
The paper defines Synthetic Linguistic Agency as a continuing artificial individual whose language participates in regulating embodied life, partners, and shared interaction. It applies the framework to artifact families and identifies six reported systems exhibiting SLA, while presenting the EMA as a constructed target organization.
- 2.1 Linguistic agency: Linguistic agency concerns language as embodied social activity whose addressed acts alter participants’ later possibilities and enter an agent’s continuing way of life.The account emphasizes that linguistic activity bears traces of other voices while changing the relations in which it participates.
- 2.1 Linguistic agency: SLA distinguishes linguistic agency from language competence and use by requiring linguistic acts and consequences to belong to an agent’s continuing organization and formation.In this study, linguistic activity regulates the artificial individual, its partner, and their shared interaction.
- 2.2 Three conditions: The operational framework specifies embodiment, linguistic participation, and precariousness as engineered, inspectable conditions for artificial systems.An artificial individual exhibits SLA when all three conditions are jointly realized within the same continuing lineage and consequential linguistic process.
- 2.2 Three conditions: Synthetic embodiment requires a particular continuing organization whose state constrains action, receives consequences, and changes later possibilities across encounters.The body’s internal and environmental state must perform these causal roles across encounters.
- 2.2 Three conditions: Synthetic linguistic participation requires sustained mutual regulation in which participants take up one another’s expressions or responses and thereby alter later action.The paper includes natural-language utterances and embodied signs as possible linguistic expressions.
- 2.2 Three conditions: Synthetic precariousness requires continuation to depend partly on the artifact’s actions, with deterioration or failure shaping what the same individual can later be and do.Homeostatic regulation, cumulative damage, declining action capacity, and terminal dissolution can realize this relation.
- 2.3 Relational map and boundary cases: Table 1 classifies representative artifact families by complete or partial realization of the three conditions, keeping different mechanisms separate.A check indicates a complete relation, while a cross denotes absence or incompleteness; partial relations retain relevant causal connections without satisfying the full condition.
- 2.3 Relational map and boundary cases: The examples separate distinct boundary cases: Hermes lacks embodiment, SayCan has complete embodiment but partial linguistic participation, and DRN has participation without body or viability.Artificial Leviathan supports sustained linguistic participation but has partial embodiment and precariousness.
3 Architecture and social formation
The architecture couples a continuing mortal body, social biography, linguistic affordances, and viability-guided action selection. Repeated consequential encounters let one EMA acquire SLA through changes to its body, history, partner expectations, and policy.
- 3.1 A linguistically mediated, mortality-grounded HRRL instantiation: HRRL represents the agent through coupled internal and external states, with reward defined by reducing deviation from viable or preferred conditions.This supports actions whose external effects regulate the agent’s body.
- 3.1 A linguistically mediated, mortality-grounded HRRL instantiation: Reference-RUL is expected remaining life under a silent, unprotected reference continuation, not a forecast under the EMA’s learned policy.The mortality-grounded drive increases as expected remaining life falls; health declines monotonically, damage increases fragility, protection can slow decline, and death is absorbing.
- 3.1 A linguistically mediated, mortality-grounded HRRL instantiation: MGL-RL learns partner responses in linguistic context, maps those responses to health and Reference-RUL, and compares actions on expected remaining life.The EMA can answer, withhold an answer, appeal for protection, or propose reciprocal exchange.
- 3.2 The mortal EMA and one consequential encounter: At each encounter, the EMA encodes a request with partner identity and biography, selects among four composite actions, renders language when applicable, and updates its body, biography, and Bayesian evidence.The action space separates task performance from social expression: answers or refusals can appeal for protection, exchange protection, or remain silent.
- A One individual through social formation: One EMA remains the same individual across 60 consequential encounters while its body condition, biography, social expectations, linguistic affordances, and body-conditioned policy change together.The controlled simulation supplies requests and partner identities; repeated independent runs form a cohort of independently formed lives.
- B One controlled formation run: The controlled formation run presents raw requests and partner identities, predicts responses, values four actions by Reference-RUL, renders the selected action, and records consequences for the next encounter.Response-bearing actions produce MAINTAIN or SKIP; silent refusal bypasses the response path and proceeds to bodily settlement.
- B One controlled formation run: Action and environmental damage reduce health, while MAINTAIN can eliminate environmental base damage at protection efficiency 1.0 without removing action damage.Health never increases and life ends at zero; action preferences therefore depend on current health and predicted partner response.
- 3.3 Learning linguistic affordances from biography: The learner shares evidence across linguistically similar encounters with an RBF kernel and stores each settled request, partner, style vector, action, and outcome in an append-only biography.The posterior estimates how likely a linguistic form with a particular partner is to elicit protection, which the body model values at current health.
4 Evaluation Design
The evaluation uses matched interventions and paired counterfactual continuations to test embodiment, biography, reciprocal participation, and persistent bodily consequences as separable yet coupled features of the EMA’s linguistic process.
- Design: Paired interventions change one connection in the same continuing EMA life while matching remaining body, biography, social ecology, and potential events where permitted.The design tests nonredundant causal contributions and their coupling within one consequential linguistic process.
- Biography-dependent affordance: Adult Aligned, Mirrored, and Neutral populations provide contrasting partner ecologies for testing whether linguistic mappings form through consequential history.Aligned and Mirrored partners reverse the request-style-to-response relation, while Neutral partners provide a nondirectional comparison.
- Embodiment: Body-state probes transplant health while preserving biography, request, partner, and keyed randomness, and body-blind probes isolate bodily effects on action evaluation.Policy separation is measured over the complete four-action set using total-variation distance; the body transition after selection remains unchanged in the body-blind ablation.
- Biography-dependent affordance: History rotation preserves experienced episodes while changing language-consequence alignment, testing whether adult linguistic affordances depend on biography’s encoded associations.Questions, actions, outcomes, partners, event order, and experience count remain fixed in the rotation intervention.
- Participation: Reciprocal participation is tested by switching partner language, forcing agent output, and removing current-partner learning while holding matched encounter conditions fixed.These interventions measure partner-to-agent influence, agent-to-partner influence, and later person-specific coordination.
- Precariousness: Mortality continuations compare Natural, Pooled, fixed-action, oracle, round-reset, and non-mortal regimes using policy, social-response, and 40-turn viability readouts.Mortal branches carry damage to absorbing death; round-reset branches restore health between turns; non-mortal branches record social return externally.
5 Results
The experiments show that the EMA’s body, biography, reciprocal linguistic interaction, and persistent bodily consequences jointly shape linguistic action and the future of the same life. Controlled interventions and ablations separate these contributions while preserving their coupling.
- Body state regulates policy in a mortal lineage: 0.0521 in Formal and 0.0511 in Independent Witness: changing current health altered policy total-variation distance under fixed social evidence.The same biography, request, partner identity, and keyed randomness produced different speech and refusal distributions for bodies with different remaining lives.
- Body state regulates policy in a mortal lineage: 1.435 adult rounds in Formal and 1.550 in Independent Witness: the Natural policy outlasted the Pooled policy in adult continuation.The additional rounds belonged to the same source life because damage accumulated monotonically and death ended each branch.
- Biography shapes linguistic affordances: After consequential experience, Aligned and Mirrored lives learned opposite mappings between request style and answer strategy, despite shared held-out questions and speakers.Aligned adults favored ANSWER_APPEAL after POLITE requests and ANSWER_EXCHANGE after TRANSACTIONAL requests, while Mirrored adults acquired the reverse relation.
- Biography shapes linguistic affordances: The Neutral population’s adult mapping was -0.0159, indicating no common directional language-consequence relation without such formative history.All three populations survived the same 60 controlled encounters, so the contrast was measured at a common social-formation endpoint.
- Participation runs in both directions: 0.125 to 0.138: matched POLITE and TRANSACTIONAL framings changed the EMA’s policy while question, partner, body, and potential events remained fixed.Forcing the ecology-correct answer increased P(MAINTAIN) by 0.340 in Formal and 0.331 in Independent Witness, linking incoming language to EMA policy and expression to partner conduct.
- Participation and precariousness within one life: Personal evidence increased correct-action probability by 0.356 to 0.377 and Reference-RUL by 1.224 to 1.235 during interaction with a persistent countertypical partner.The response entered the biography, changed the next policy, and altered the remaining course of the same life; mortality-regime ablations isolated consequence persistence.
6 Discussion
The discussion shows that the EMA’s linguistic agency depends on coupling a body, social biography, reciprocal participation, and persistent bodily consequences within one lineage. Interventions preserve some effects while revealing that only persistent consequences reorganize the future viability of the same individual.
- Operational interpretation: SLA requires the functional coupling of embodiment, linguistic participation, and precariousness within one persistent artificial individual.The EMA operationalizes this coupling through a consequential linguistic process acquired and maintained in one lineage.
- Body and biography: Biography predicts partner responses, while current bodily state determines which socially informed action is worth pursuing.Neither variable substitutes for the other: biography supplies social expectations, and the body supplies an agent-relative basis for valuing outcomes.
- Reciprocal participation: Reciprocal participation makes linguistic influence bidirectional and historically productive for both the EMA and its partner.Partner responses alter the EMA’s evidence, later action, and Reference-RUL, while the EMA’s expressions change partner maintenance behavior.
- Consequence persistence: Only persistent bodily consequences produce the persistent-viability pattern across the mortality-regime comparisons.Reset and non-mortal controls retain learned mappings and social effects, but remove the connection to the future viability of the same lineage.
- Causal coupling: Removing consequence persistence preserves linguistic learning and partner-response effects but breaks their connection to future viability.Under the paper’s operational definition, the intervention results support SLA because the three engineered relations make nonredundant causal contributions within the EMA.
- Scope and limitations: The evidence remains bounded to one frozen scripted ecology and has not established a robustness range across environments, body dynamics, or learning architectures.The EMA also uses a known one-dimensional body, restricted interaction repertoire, simulated embodiment, and no formal human-participant experiment.
7 Conclusion
The study identifies and causally analyzes Synthetic Linguistic Agency in artificial systems, using MGL-RL and an EMA to test embodiment, participation, and precariousness. The EMA’s linguistic behavior reflects bodily state and social history, changes partner behavior, and carries forward bodily consequences only when mortality is irreversible.
- The first study operationalizes SLA and identifies six reported artificial systems that exhibit it.
- MGL-RL places a learned, action-conditioned partner-response model between linguistic action selection and the EMA’s body model.
- Current body state changes linguistic action values, while social history changes expected partner responses.
- Incoming framing changes EMA policy, forced expression changes partner behavior, and realized responses update later interaction.
- Round-reset and non-mortal controls preserve linguistic learning and social effects, but only irreversible-mortal branches carry bodily settlement into future viability.
- Intervention blocks show separable, nonredundant contributions that remain coupled within the same body and biography, yielding an EMA that exhibits SLA.
Supplementary Material
Supplementary Figure S1 presents complete mortality-regime results across policy, biography, viability, social, linguistic, and trajectory outcomes. Forest-plot intervals derive from 10,000 bootstrap resamples per phase.
- Forest plots report policy and biography effects, forced-output cascades, continuation viability, and social and linguistic continuation outcomes by phase and ecology.
- Trajectory panels report mean start-of-turn Reference-RUL and cumulative external return for Aligned and Mirrored sources under Natural and Pooled policies.
- Points are source-paired estimates with bars showing 95% intervals from 10,000 bootstrap resamples per phase.