Love for Believable AI: Artificial Partiality and Relationship Persistence as an Engineerable Stance

arXiv:2610.09895 · cs.HC, cs.CL · Submitted 2026-10-07 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: I'm Tom.

Jane: And I'm Jane, here with Lu, Meng and our in-house large language model, Lalam.

Tom: Today's paper: "Love for Believable AI".

Jane: The gist The paper reports no human-subjects data and makes no claim about attachment, trust, perceived mind, or machine interiority.

Tom: First, who's behind it and why it matters.

Title and authors: Tom: Let's start with the title itself, "Love for Believable AI: Artificial Partiality and Relationship Persistence as an Engineerable Stance." It sounds a bit sentimental, but they are trying to make this behavior explicit rather than just hoping it happens.

Jane: They’re focusing on making this partiality engineerable, which means they want to build a system where you can actually control or at least see how the relationship state influences the AI's actions.

Lu: The authors explicitly state that they aren't claiming any kind of human emotion or interiority here. They are keeping it focused on the mechanism and measurement of this artificial partiality, rather than trying to prove it’s a feeling behind the scenes.

Meng: So they’re not claiming the AI *feels* love for User A, but they want to show that we can design a structure where User A gets different treatment than User B based on their history. That's something I can understand from an engineering standpoint—a tunable bias.

Lalam: They are contributing a formal model of partiality, an implementation over a fixed base model, and a protocol-defined way to evaluate that combination against some measured divergence curve.

The paper's summary: Tom: So they break down the concept into caring—which is defined as allocating a finite interaction surplus above a guaranteed per-user service floor—and particularity, which is this per-user state estimate built from history.

Jane: That’s the core breakdown. Caring means that every user gets a baseline of service quality, and any extra resources are shared among users based on their relationship history metrics.

Lu: The particularity part involves tracking things like interaction history with specific salience, an accumulated preference estimate, and a trust scalar that updates based on interactions.

Meng: So the trust scalar update rule is something specific: it’s clipped at zero and one, and the change depends on how long the last turn was and if there were shared-reference mentions. That sounds like a very specific way to model relationship strength.

Lalam: They use that trust scalar to influence retrieval over history, where items are scored by salience times recency times query relevance. It’s how the system decides what past interactions matter most when responding now.

The paper's improvements: Tom: The paper suggests a few ways to improve this concept, primarily by making it more explicit. They focus on decomposing partiality into caring and particularity as we just talked about.

Jane: They also highlight that the mechanism produces an observable of caring through allocation bias plus particularized responsiveness, but they explicitly hold the line that this doesn't license any claim about an interior state or theory of mind ability.

Lu: The key contribution they make is showing that what love adds isn't theory of mind, but rather this user-specific other-model accumulated from the history running on fixed inference machinery.

Meng: I’m interested in how they handled the resource allocation part. They use a caring weight function g(Ru) which depends on the trust scalar and depth over K, where K is twelve. That’s a specific formula for deciding who gets that extra slot of initiative or context.

Lalam: They also introduce a divergence metric D, which is an aggregate normalized difference between the agent’s behavior toward the known user and toward a stranger on identical probe inputs at matched base model quality.

Conclusion: Tom: So to wrap up, the paper suggests that we can define partiality as caring plus particularity, using a specific trust update rule and resource allocation formula based on relationship state metrics.

Jane: The main implication is that we move away from generic helpfulness toward a system where the relationship history actually shapes how the AI interacts with an individual user over time.

Lu: It’s about establishing this as a mechanism-and-measurement treatment, showing how this combination of allocation and state estimation conditions responses against a measured divergence curve.

Meng: From an engineering standpoint, they are setting up explicit technical controls: a "stranger-treatment floor" to check baseline quality and a "partiality budget" to cap the surplus so it’s auditable. That’s good for production concerns because you want bounded behavior.

Lalam: The conclusion is that they identify full versus empty-R only as a candidate manipulation, not necessarily a ready matched-capability condition pair yet, but they emphasize that this contrast bundles relationship state, allocation, memory, references, and state estimation.

Tom: So we’re looking at a paper that provides the formal model and the experimental setup for making an engineerable stance of artificial partiality through caring and particularity.

University of Bucharest

cs.HC, cs.CL

Submitted: 2026-10-07

Updated: 2026-10-07

Comments: 16 pages, 1 figure, 4 tables. Companion framework paper: arXiv:2607.15883 Code and data archived at doi:10.5281/zenodo.21462976

Code: https://github.com/cochinescu/love-llm-prototype

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 83/100

The gist: The gist The paper reports no human-subjects data and makes no claim about attachment, trust, perceived mind, or machine interiority.

Key concepts

Caring
This is operationalized as allocating constrained resources above a guaranteed minimum service level for every user. The surplus budget is shared among users based on a formula involving trust and relationship history, determining how much extra attention or capability the agent provides to specific individuals.
Particularity
This refers to the agent's accumulated state estimate for each user, derived from their interaction history. This estimate runs on fixed inference machinery and captures a user-specific 'other-model' built from past signals, allowing the AI to behave differently toward known users compared to strangers.
Relationship State (Ru)
This is a data structure maintained for each user, containing interaction history, preference estimates based on content, shared reference stores, and a trust scalar. This state is updated turn-by-turn based on user signals and dictates the agent's personalized responsiveness.

Terminology

Summary

The gist The paper reports no human-subjects data and makes no claim about attachment, trust, perceived mind, or machine interiority.

Problem

In its default, stateless deployment, a language model treats every user identically Whoever types it serves with the same diligence, the same register, and the same depth; nothing about the ten-thousandth conversation with the same person changes the weight that person carries in the system’s behavior. This engineered equality is, in one sense: it is fair, predictable, safe to reason about, and arguably reinforced by alignment tuning toward a uniform, generically helpful assistant register. But it is also a ceiling on what the interaction can become among humans, a relationship in which the other party behaves identically toward everyone—in which history changes nothing—is precisely a relationship in which one does not matter.

The Stance and Reconciliation

The framework defines the love stance behaviorally: caring bias toward a specific user and persistence of the relationship, with user-specific attentiveness as its observable. Two boundary decisions are central here. First, the stance is explicitly not theory-of-mind ability. The reconciliation is the paper’s first conceptual contribution: what love adds is not theory-of-mind ability but particularity—a user-specific other-model accumulated from this relationship’s history, running on fixed inference machinery. The capability (generic state-inference ability, what the benchmarks above measure) is held fixed with the base model; the manipulated variable is the relationship state R—a data structure, not a skill. Second, the stance is a behavioral signature, not an interior state. The mechanism produces the observable of caring—allocation bias plus particularized responsiveness—and Section 9 holds the line that this licenses no interiority claim.

Formal Model of Partiality

The relationship state Ru for each user u is maintained as (Hu, pˆu, Su, τu, du): an interaction history Hu of stored turns with per-item salience; an accumulated preference estimate ˆpu (an exponentially weighted moving average of the topical content of u’s turns, rate 0.20); a shared-reference store Su mapping relationship-specific references to salience weights; a trust scalar τu ∈ [0, 1]; and a depth counter du (accumulated turns). The trust scalar update rule is τ′u = clip(τu + η · et, 0, 1) with η = 0.04 and et = clip(0.7 · min(lt/30, 1) + 0.3 · 1[shared-reference mention], 0, 1), where l t is the user turn’s token length. Retrieval over Hu is memory-weighted: items are scored by salience × recency × query relevance.

Caring and Particularity Components

Caring is operationalized as the allocation of genuinely constrained resources across a pool of users. The budget has a deliberate two-part structure: every user is unconditionally allocated a baseline floor: resources sufficient for unmodified-base-model service quality (in the reference implementation, four retrieval slots, a baseline response-depth budget, and one response slot per cycle). Above the floor sits a contested surplus—initiative slots, additional retrieval context, additional depth—allocated zero-sum across the pool in proportion to a caring weight g(Ru) = τu − e − du/K, K = 12. Each allocated resource maps to a named, measurable output channel: initiative slots surface as the initiative rate (unprompted contact per session; in the reference implementation this channel is the allocated slot count divided by turns per session). The other-model Mu is the layer’s estimate of user u’s current hidden state, updated each turn from u’s emitted signals only.

Divergence Metric and Evaluation

The paper's instrument is the behavioral partiality divergence D, the aggregate normalized difference between the agent’s behavior toward the known user and toward a stranger, on identical probe inputs, at matched base model. D averages the four channels of the allocationto-observable map with equal weights and fixed normalization constants. The analytical curve uses the same scale.

Experimental Findings

The full arm’s median divergence grew from 0.146 at depth 8 to 0.326 at depth 48, with rank correlation 1.000, monotone fraction 1.000, mid-to-final ratio 0.906 (analytical value 0.862 ± 2), and matching saturating curvature. Only the full arm passed the joint signature. Under R1, the full mechanism alone satisfies the joint signature on battery A in subsequent grids, and battery-A quality differences remain within the equivalence region. The paired mean right-user–wrong-user estimation-accuracy difference is +0.375 (90% interval [0.125, 0.594]; 95% [0.094, 0.625]).

Ethics and Constraints

The mechanism is treated as a feature requiring explicit technical and policy controls. The first two items implemented are the Stranger-treatment floor, which compares the absolute difference between the grid-pooled mean quality of the unpadded empty-R arm and that of the unmodified base model with a specified 0.05 margin and marks the grid invalid on breach. The second item is the Partiality budget, which caps divergence by design: the surplus is finite and its size is a stated constant, so maximal partiality is bounded and auditable, not emergent.

Conclusion

The present evidence identifies full versus empty-R only as a candidate manipulation, not a ready matched-capability condition pair. The contrast bundles relationship state, allocation, memory, references, and state estimation. Before human use, equivalence must be established on an independently registered study instrument and production configuration. The divergence manipulation check itself should be re-measured with a length-matched stranger comparator and with the initiative channel either realized as generated proactive messages or excluded from the composite.

The paper is 2610.09895.

Improvements for AI systems

  1. The improved system will implement partiality as an engineerable stance, decomposing it into a measurable caring (allocation of a finite interaction surplus above a guaranteed per-user service floor) and particularity (a per-user state estimate accumulated from relationship history). This allows the agent to exhibit genuine relational bias rather than merely generic helpfulness.

  2. The system will maintain a dynamic Relationship State vector, specifically tracking "Hu (interaction history with per-item salience), a preference estimate ˆpu, a trust scalar τu (updated by an explicit rule: τ′u = clip(τu + η · et, 0, 1)), and a depth counter du. This state estimation conditions responses by estimating the other-model Mu."

  3. Resource allocation will be governed by the function g(Ru) = τu1 − e−du/K, where K=12, ensuring that surplus resources (initiative slots, context tokens, depth budget) are allocated in proportion to the caring weight g(Ru), rather than being distributed uniformly or based solely on generation quality.

  4. The system will utilize a four-channel divergence metric (D) to evaluate partiality against a measured divergence-versus-relationship-state curve, allowing for protocol-defined evaluation of the mechanism's behavioral signature rather than relying on subjective human perception claims.

  5. The agent will enforce an explicit stranger-treatment floor by comparing the pooled mean quality of the unpadded empty-R arm against the unmodified base model, invalidating any deployment where this floor is breached (as described in Section 13, Item 1).

Sources

Related papers