Empathy as Predictive Misalignment Tolerance: A Co-Regulation Framework and the Regime Structure of Dialogue Repair

Technical Summary: Empathy as Predictive Misalignment Tolerance

Problem Statement
Current theoretical and computational models of empathy predominantly rely on a synchronic framing, defining empathy as the immediate resonance, mirroring, or simulation of another's emotional or cognitive state. In artificial intelligence, this translates to systems that detect user affect and generate congruent responses. The authors argue this is a "thin" notion of empathy that fails to account for extended dialogue, where understanding is a temporal process characterized by drift, divergence, and repair. Existing frameworks treat misunderstanding as a failure to be eliminated, whereas the authors posit that genuine understanding in extended interaction requires the capacity to anticipate and regulate interpretive divergence over time, rather than collapsing it.

Theoretical Framework: Interpretive Error Tolerance (IET)
The paper proposes reframing empathy as predictive misalignment tolerance. This is operationalized through the concept of Interpretive Error Tolerance (IET), defined as the capacity to maintain a "viable band of divergence" between interacting agents.

• Core Mechanism: Instead of minimizing the distance between two interpretive trajectories, the goal is to keep that distance within a bounded, dynamically negotiated zone (the tolerance space).
• Formal Heuristic: The authors introduce a regulatory update rule (Equation 1) for the tolerance threshold ($\theta_t$):
  $$\theta_{t+1} = \theta_t + \eta (\sigma(\Delta_t) - \sigma^*)$$
  Where $\Delta_t$ is the interpretive divergence, $\sigma(\Delta_t)$ is the short-window variance of recent divergence, $\sigma^*$ is a homeostatic target variance, and $\eta$ is an adaptivity rate.
• Regulatory Logic: When divergence variance exceeds the target, the tolerance band widens; when it becomes too orderly, the band tightens. Repair actions (e.g., clarification, downweighting suspect turns) are triggered when projected divergence threatens to exit this band.
• Philosophical Stance: Drawing on predictive processing, enactive cognition, and phenomenology (specifically Levinasian ethics), the framework argues that empathy is a process of co-regulation that preserves the "irreducibility" of the other, rather than a fusion of minds.

Methodology: Computational Probes
To test the operational viability of IET, the authors conducted two computational probes using the DailyDialog corpus. The experiments employed a controlled noise-injection paradigm where time-varying noise was introduced into dialogue contexts. Three policies were compared:

1. No-Repair Baseline: No intervention.
2. Fixed-$\theta$ Policy: A static threshold for triggering repair.
3. Adaptive-IET Policy: Implements the variance-based update rule (Equation 1).

• Probe 1 (Signal Choice): Used cosine distance between turn embeddings and a running context summary. Repair was a "hard drop" of suspect turns.
• Probe 2 (Anchor-Based Drift): Redesigned to address Probe 1's limitations.
  • Signal: Drift measured against a fixed "topic anchor" (first two turns) rather than a running summary.
  • Action: Repair changed from hard dropping to continuous downweighting ($w = \exp(-k \cdot \max(0, \Delta_t - \theta_t))$).
  • Metrics: Evaluated on two distinct axes: Discriminative Fidelity (1-of-20 next-utterance retrieval accuracy) and Gist Preservation (cosine similarity to the clean, unnoised original context).

Key Results
The empirical results did not validate the Adaptive-IET policy as a superior controller over fixed thresholds, but they revealed a critical structural phenomenon in dialogue repair:

1. Failure of the IET Update Rule: In both probes, the Adaptive-IET policy was statistically indistinguishable from the Fixed-$\theta$ policy (paired McNemar $p=1.0$). The variance-based update was too gentle to drive meaningful differentiation in the tested embedding space.
2. Signal Conflation in Probe 1: Probe 1 failed because the "running summary" signal conflated natural topical evolution with noise-induced misalignment, causing false positives that degraded performance.
3. Regime-Structured Repair (Probe 2): The most significant finding was a regime-dependent trade-off between retrieval accuracy and gist preservation:
   • Low-to-Moderate Noise: Repair mechanisms degraded discriminative fidelity (retrieval accuracy) compared to the no-repair baseline.
   • High Noise: Repair mechanisms began to preserve gist meaning (cosine similarity to clean context) better than the no-repair baseline, while retrieval accuracy remained low.
   • The Crossover: The curves for gist preservation crossed between noise rates of 0.30 and 0.45. Repair only became beneficial for meaning preservation when noise was high enough to degrade the baseline's performance comparably.

Significance and Claims
The paper explicitly states that its primary contribution is conceptual, normative, and methodological, not the validation of a specific control algorithm.

• Conceptual Reframing: The authors argue that empathy in extended interaction is not the elimination of divergence but the regulation of its dynamics. The "success" of an empathic system lies in maintaining a workable interpretive relationship across difference, not in achieving convergence.
• Empirical Insight: The "regime structure" of repair suggests that existing evaluation frameworks (which often average performance or use single metrics) obscure the reality that repair strategies involve a fundamental trade-off: sacrificing discriminative detail to preserve shared gist under high noise.
• Design Implications: The paper advocates for a shift in AI design from "convergence-seeking" to "divergence-management." It calls for:
  • Personalized and Contestable Tolerance: Parameters defining the tolerance band should be visible and adjustable by users, rather than hidden defaults.
  • Regime-Aware Evaluation: Benchmarks must test performance across noise grids and use dual metrics (harsh vs. forgiving) to capture the trade-offs between detail and gist.
  • Ethical Humility: Systems should be designed to preserve the user's relational ecology and distinctness, avoiding the risk of "empathy-as-capture" where smooth regulation creates dependency.

In conclusion, the paper posits that while the specific IET update rule tested did not outperform fixed baselines, the phenomenon of regime-structured repair confirms the necessity of viewing empathy as a dynamic process of maintaining intelligibility across difference, rather than a static state of alignment.