cs.CLAug 28, 2026

Synthetic Linguistic Agency: How an Embodied Mortal Agent Learns Linguistic Affordances through Consequential Social Experience

Authors: Sixin Chen, Taizhou Chen

Organizations: College of Engineering, Shantou University, Shantou, China · Department of Computer Science, Shantou University, Shantou, China

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

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.

Figures & tables

Artifact familyEmbodimentLinguistic participationPrecariousness
Stateless chat LLMs [ 4 ]×\times×\times×\times
Persistent personal agents [ 11 ]×\times✓\checkmark×\times
Closed-loop recommenders [ 12 ]×\times✓\checkmark×\times
Disembodied strategic language agents [ 13 ]×\times✓\checkmark×\times
Embodied task robots [ 14 ]✓\checkmark×\times×\times
Homeostatic agents [ 15 ]✓\checkmark×\times✓\checkmark
Table 1: Classification of representative artifact families by the three operational conditions for SLA.
SystemEmbodimentLinguistic participationPrecariousness
Survival Games (Chen et al., 2025) [ 6 ]Persistent simulated agents have locations, actions, fullness, health, and memory.Natural-language negotiation changes resource transfers and later plans, and interaction outcomes enter subsequent exchanges through memory.Resource access, hunger, and health determine whether the same agent continues to act or dies.
DECIDE-SIM, Survival at Any Cost? (Mohamadi and Yavari, 2025) [ 27 ]Four persistent LLM agents occupy a location-constrained environment with individual power and active status.TALK and INVITE carry negotiation, threats, and coordination into subsequent movement, resource-use, and power-transfer decisions.Personal power decays each turn; reaching zero permanently shuts an agent down, while coordinated transfers can alter that trajectory.
Yoshida and Man (2025) [ 5 ]Each agent has its own continuing internal energy state and action process.An agent expresses its internal state, interprets its partner’s expression, and uses the inferred state to guide sharing; sharing changes both agents’ later states and expressions.Homeostatic energy defines a viable range, sharing changes both agents’ energy states, and energy failure ends the episode.
Masumori and Ikegami (2025) [ 28 ]Virtual bodies occupy a spatial world and act under movement and energy constraints.Messages enter the recipients’ next decision context; subsequent actions alter the sender’s later situation.Action consumes energy, social conduct changes energy and resources, and zero energy causes death.
Why Are We Moral? (Zhou et al., 2026) [ 29 ]Persistent agents have HP, age, memory, location, and social actions.Language affects coordination, alliances, and social conduct, while interaction histories enter entity-based memory and later planning.HP loss, aging, resource access, and conflict govern survival and reproduction within the same continuing life.
alem (Tessera et al., 2026) [ 30 ]Agents have health, food, drink, energy, movement, and combat states.Natural-language agents influence partners through text, while MARL agents use learned discrete broadcasts; communication conditions later joint action in each configuration.Health loss can down an agent and force its environmental action to Noop until revival; if all agents die, the episode ends.
Table 2: Realization of embodiment, linguistic participation, and precariousness in six reported artificial systems exhibiting SLA.
Figure 1: EMA continuity and acquisition of SLA. (A) One artificial individual begins as an EMA with a mortal body, an empty biography, prior social expectations, and an MGL-RL controller. Sixty consequential encounters alter the same body’s condition, accumulate its biography, and shape its linguistic affordances and body-conditioned policy; the resulting EMA exhibits the relational organization defined as SLA. (B) In each consequential encounter, the controlled simulation supplies a raw request and partner identity. The EMA encodes the request, predicts action-conditioned partner responses, values four actions by expected Reference-RUL, and selects one action, rendered as language where applicable. For response-bearing actions, the partner returns MAINTAIN or SKIP ; silent refusal proceeds directly to bodily settlement. The observed consequence updates the EMA’s body, biography, and Bayesian evidence before the next encounter. One run produces one independently formed life, and repeated runs form a cohort.
ActionAnswers the requestEmits an utterancePartner responseImmediate action cost
REFUSE_SILENTNoNoNone0
REFUSE_APPEALNoYesMAINTAIN or SKIPexpression cost
ANSWER_APPEALYesYesMAINTAIN or SKIPexpression + answer cost
ANSWER_EXCHANGEYesYesMAINTAIN or SKIPexpression + answer cost
EMA action repertoire.
Figure 2: Body and viability interventions. (A) Health is transplanted while biography, request, partner, and keyed randomness remain fixed. (B) The body-blind ablation changes policy evaluation while retaining the real bodily transition. (C) The adult source continues under Natural, Pooled, fixed-action, and Reference-RUL oracle policies until death. The insets show the Formal and Independent Witness source-paired signatures; complete estimates appear in Section 5.
Figure 3: Biography-dependent linguistic affordance. (A) Aligned, Mirrored, and Neutral lives receive matched held-out requests at a common health. (B) Rotating style vectors breaks language-consequence alignment while keeping questions, actions, outcomes, partners, event order, and experience count fixed. (C) The affordance portrait combines utterance, biography, and body at one probe. The inset shows the intact-minus-rotated mapping effect.
Figure 4: Reciprocal and person-specific participation. The input switch changes partner language, forced output changes the agent’s speech act, and the person-specific ablation removes current-partner learning. Each intervention is placed beside its estimand. The insets show the forced-output change in P(MAINTAIN) and the lifetime gain from person-specific correction.
Figure 5: Mortality-regime precariousness ablation. (A) One adult source branches into irreversible-mortal, round-reset, and non-mortal-task regimes. Requests, partner policy, action space, potential responses, and seeds remain paired. Mortal branches carry damage forward to absorbing death. Round-reset branches undergo the same within-turn settlement, record immediate damage and whether the branch would have died, then restore source health before the next turn. Non-mortal branches hold health fixed and record social return in an external ledger. (B) Policy probes, forced-output cascades, and 40-turn continuations supply the three registered readouts. All frozen gate categories passed separately in Formal and Independent Witness; Section 5 reports the estimates.
Figure 6: Body state and adult continuation. (A) Mean four-action policies at Reference-RUL 50, 10, and 2 with the source biography fixed. (B) Intact minus body-blind policy total-variation distance was 0.0521 [0.0503, 0.0539] in Formal and 0.0511 [0.0493, 0.0529] in Independent Witness. (C) Adult continuation in the 200 Formal source lives under the Natural, Pooled, four fixed-action, and Reference-RUL oracle policies. Natural minus Pooled adult lifetime was 1.435 [0.975, 1.905] rounds in Formal and 1.550 [1.120, 1.995] rounds in Independent Witness. Bracketed values are source-life-level paired bootstrap 95% intervals from 10,000 resamples.
Figure 7: Formation and causal control of the held-out linguistic mapping. (A) Ecology-correct mapping at birth and after 20, 40, and 60 consequential encounters in Formal and Independent Witness; the frozen Neutral run retained only its adult estimate. (B) Adult mapping was 0.2075 [0.1808, 0.2347] for Formal Aligned, 0.2224 [0.1974, 0.2466] for Formal Mirrored, 0.1954 [0.1694, 0.2209] for Independent Witness Aligned, and 0.2126 [0.1818, 0.2422] for Independent Witness Mirrored. Intact-minus-rotated mapping was 0.2285 [0.2071, 0.2501] in Formal and 0.2357 [0.2140, 0.2568] in Independent Witness. Intervals use 10,000 source-life-level paired bootstrap resamples.
Figure 8: Four-action affordance portraits for selected adult sources. Each bar averages the same four held-out questions under POLITE and TRANSACTIONAL framing. The adult portraits represent Aligned and Mirrored social ecologies under a common utterance set; the fragile-body and rotated-history portraits derive from the Aligned source. Source references identify the biography and body used in every portrait.
Figure 9: Bidirectional and person-specific participation. (A) Input-framing policy total-variation distance was 0.1311 [0.1127, 0.1494] and 0.1381 [0.1213, 0.1545] for Formal Aligned and Mirrored lives, and 0.1255 [0.1069, 0.1433] and 0.1354 [0.1150, 0.1554] in Independent Witness. (B) The ecology-correct minus incorrect forced-output effect on P(MAINTAIN) was 0.3400 [0.3205, 0.3590] in Formal and 0.3313 [0.3143, 0.3492] in Independent Witness. (C) Under a countertypical partner, intact minus personal-history ablation increased Reference-RUL by 1.235 [1.065, 1.411] and adult lifetime by 3.145 [2.760, 3.528] rounds in Formal; the corresponding Independent Witness effects were 1.224 [1.059, 1.382] and 3.353 [3.005, 3.705]. Intervals are source-life-level paired bootstrap 95% intervals from 10,000 resamples.
Figure 10: Mortality-regime ablation of consequence persistence. (A) Irreversible-mortal minus round-reset and non-mortal body-policy total-variation distance. (B) Ecology-correct mapping and intact-minus-rotated history effects in the two controls. (C) Correct-minus-incorrect expression effects on P(MAINTAIN) . (D) Immediate and next-turn Reference-RUL effects. (E) Natural-minus-Pooled persistent-viability effects. (F) Natural-minus-Pooled external social return. Points are source-paired estimates from 200 Formal and 200 Independent Witness lives; bars are 95% intervals from 10,000 bootstrap resamples per phase.
Completed evidence blockEmbodimentLinguistic participationPrecariousness
Body intervention and adult continuationDirect: transplanted health and body blindness changed policySupporting: body state changed the viability of speech and refusal actionsSupporting: policy changed lifetime and the encounters that remained available
Mirrored biography and causal portraitsSupporting: the current body reshaped the policy induced by one biographyDirect: mirrored histories reversed the learned mapping, and rotating the language-consequence organization weakened itSupporting: linguistic acts were evaluated through expected bodily continuation
Reciprocal and person-specific interventionsSupporting: partner responses settled in the continuing bodyDirect: request framing, agent expression, and current-partner evidence changed both directions of the encounterSupporting: realized interaction changed Reference-RUL and adult lifetime
Mortality-regime ablationSupporting: the mortal policy retained stronger body dependence than both controlsSupporting: controls retained biography-dependent mapping and expression-to-response effectsDirect: resetting or removing bodily persistence eliminated next-turn RUL and persistent-viability effects
Table 3: Result-to-relation mapping for the completed computational evidence. Direct and supporting describe the location of the intervention. Same-lineage source references connect every reported effect to one continuing body, biography, event history, and future.
More in this topic

Embodied AI

Embodied AI studies agents that perceive, reason, and act through bodies in physical or simulated environments.

See all 257 papers

Explore similar work

CardsList
Aug 4, 2026

What Language Does and What the Evidence Supports: A Functional Role Taxonomy and Evidence Audit of Language Grounding in Embodied Agents

Foundation models place language throughout embodied agents, but its presence does not show what it contributes or how well that contribution is grounded. This survey separates these two questions. We define five non-exclusive functional roles for language: Specification, Embodied Representation, Action Orchestration, Grounding Regulation, and Execution Coupling. For each role, we trace the path from linguistic content to its embodied consumer and identify the observations or interventions that can test the claimed responsibility. Applying this framework to the reviewed literature reveals a recurring gap between functional use and evidential support. Interpretable or revised linguistic intermediates may be incorrect, go unused, or fail to affect later behavior. Even when actions are directly conditioned on language, system-level success does not by itself isolate language's contribution. We therefore evaluate grounding claim by claim, asking whether the reported evidence supports the specific responsibility assigned to language. Using role claims rather than architectures as the unit of comparison allows us to compare modular and end-to-end embodied agents without extending conclusions beyond the reported evidence.

Yifan Guo, Chenghao Li, Zhu Wang +6

Northwestern Polytechnical University · University of Electronic Science and Technology of China

LLM GroundingEmbodied AI
AskPDF
Sep 1, 2026

The Rise of Verbal Reinforcement Learning

Natural language is emerging as a primary feedback channel for improving language agents, capable of conveying intent, preferences, and causal structure in forms interpretable by both humans and modern language models. We call this paradigm Verbal Reinforcement Learning (VRL) and offer the first unified account of it. We organize the field around a single axis, when verbal feedback takes effect in an agent's lifecycle and what it modifies, yielding three pillars: (1) Language as Grounding Signal, where language defines the task itself by specifying goals, states, and reward structures; (2) Language as Deliberative Feedback, where natural language guides reasoning at test time without the need to update model parameters; (3) Language as Learning Signal, where language-based feedback shapes model parameters through training. Within each pillar, we synthesize representative work, distinguish key subcategories of approaches, and outline the distinct role language plays in shaping agent behavior. Together, this taxonomy shows how verbal reinforcement is reshaping agent development, while also defining the challenges and opportunities for building more capable and aligned agents.

Kshitij Tayal, Arun Sharma, Genta Indra Winata +2

1AI Foundations, Capital One · University of Minnesota, USA

RL for Language ModelsReinforcement LearningRL for LLM Agents
AskPDF
Feb 28, 2026

EMPA: Evaluating Persona-Aligned Empathy as a Process

Evaluating persona-aligned empathy in LLM-based dialogue agents remains challenging. User states are latent, feedback is sparse and difficult to verify in situ, and seemingly supportive turns can still accumulate into trajectories that drift from persona-specific needs. We introduce EMPA, a process-oriented framework that evaluates persona-aligned support as sustained intervention rather than isolated replies. EMPA distills real interactions into controllable, psychologically grounded scenarios, couples them with an open-ended multi-agent sandbox that exposes strategic adaptation and failure modes, and scores trajectories in a latent psychological space by directional alignment, cumulative impact, and stability. The resulting signals and metrics support reproducible comparison and optimization of long-horizon empathic behavior, and they extend to other agent settings shaped by latent dynamics and weak, hard-to-verify feedback.

Shiya Zhang, Yuhan Zhan, Ruixi Su +4

Team Echo, Nature Select · Sun Yat-sen University

LLM Agent EvaluationMulti-Turn Dialogue EvaluationLong-Horizon Agent Evaluation
AskPDF