Emotion Circuits in LLMs

• Emotion circuits in LLMs are algorithmic pathways that combine episodic memory retrieval with norm-based contextualization to evaluate and express affect.
• They implement appraisal theory by comparing current experiences against a dynamically derived norm and quantifying emotional responses via the PANAS protocol.
• Empirical results highlight enhanced nuance in detecting negative affect, while also revealing challenges with model bias and memory scalability.

Emotion circuits in LLMs refer to the structured functional pathways—implemented algorithmically or arising through learning—that enable these models to perceive, represent, compare, and express emotional states in response to textual events and context. The architecture described in "Can Generative Agents Predict Emotion?" (Regan et al., 6 Feb 2024) operationalizes an emotion circuit through memory-driven appraisal, contextual comparison, and explicit affect quantification. This system aligns LLM-driven generative agents more closely with human emotional cognition by introducing a pipeline that incorporates episodic memory, contextualized event appraisal based on psychological theory, and systematic measurement of affective states.

1. Memory-Driven Architecture for Emotional Appraisal

The architecture consists of a multimodule pipeline:

• Experience Encoding: Sequential textual events (time-series inputs) are ingested by the agent, simulating a stream of lived experiences.
• Memory Retrieval and Norm Construction: Past memories relevant to the current experience are retrieved and summarized into a “norm”—a condensed, context-representative memory—using prompt-based summarization. This summary captures expectations and regularities across prior similar events.
• Contextual Comparison: The current experience is compared against the computed norm, with deviations or alignments highlighted through a secondary prompt. This contextualization step is intended to simulate episodic memory-informed human appraisal.
• Affect Elicitation (PANAS): After contextual comparison, the Positive and Negative Affect Schedule (PANAS)—a psychometric test consisting of 20 affect items—quantifies the agent's affective state by having the LLM rate its emotions based on the context-adjusted understanding.

This architecture can be rendered as:

$$\text{Experience}_t \xrightarrow{\text{Memory Retrieval}} \text{Norm}_t \xrightarrow{\text{Comparison}} \text{Contextual Understanding}_t \xrightarrow{\text{PANAS}} \text{Affect}_t$$

The use of a “norm” derived from episodic memory distinguishes this method from traditional, static feature-based emotion classifiers by endowing the generative agent with context-sensitivity and temporal continuity.

2. Episodic Memory and Norm-Based Contextualization

Episodic memory retrieval involves factors such as saliency, recency, and relevance. These memories are summarized via prompt-engineered instructions into a “norm,” reflecting, for example, standard behavioral routines or expectations in a scenario. The agent does not react to isolated events, but anchors appraisals to the deviation from, or conformity to, the established norm.

For example, in scenarios where social interaction is the historical norm, an abrupt shift toward isolation triggers a corresponding shift in predicted emotional valence. These mechanisms instantiate an explicit memory-comparison loop, analogous to episodic appraisal circuits in biological systems.

3. Appraisal Theory as Theoretical Foundation

The appraisal theory of emotion posits that emotional experience emerges from contextual evaluations of events—not simply the events themselves, but the meaning they acquire against an agent’s prior knowledge and expectations. The LLM-based agent mirrors this by operationalizing appraisal as a norm versus experience comparison, calibrating its own affect in reference to accumulated behavioral context.

Within the system, the norm functions as a dynamic set of evaluation criteria and as a latent world model, allowing the model to differentiate between ordinary and emotionally salient deviations. This mechanism directly implements a function central to appraisal theory: making emotion a function of prediction error or norm deviation.

4. Affect Quantification via the PANAS Protocol

Each post-comparison step culminates in the administration of the PANAS, with the LLM providing discrete ratings (scale 1–5) for each of the 10 positive and 10 negative affect items, randomized in order to mitigate bias. This yields two aggregate scores:

• Positive Affect Score (PAS)
• Negative Affect Score (NAS)

The administration is implemented through a dedicated prompt that compels the LLM to respond exclusively based on the interpreted context. This empirical quantification allows fine-grained analysis of how the agent’s emotion circuit responds over sequences of emotionally salient events.

5. Experimental Outcomes: Context Effects and Model Limitations

Empirical evaluation on a broad selection of episodic stories (including anger, depression, and anxiety scenarios) demonstrates that background context—instantiated through norm creation—generally yields more nuanced emotional dynamics, especially in negative-affect situations. Specifically:

• Norm-augmented agents correctly capture stronger negative affect in conflict or isolation contexts, closer to human-annotated responses.
• In certain scenarios, however, model-intrinsic biases—such as over-optimistic or context-insensitive baseline responses—persist even with contextual scaffolding, revealing the need for model selection and better alignment protocols.

These results indicate that the integration of episodic memory and context-aware appraisal consistent with psychological theory can enhance the granularity and validity of LLM-based emotion circuits, but that full emotional alignment remains challenging.

6. Limitations and Prospective Improvements

Key limitations identified in the paper include:

• Intrinsic Model Bias: The LLM used for PANAS administration (here, GPT-3.5-Turbo) can exhibit positivity bias, yielding unrealistically optimistic affect ratings even in negative contexts.
• Memory Scalability: The memory retriever’s performance in weighting saliency, recency, and relevance may degrade with larger or longer memory banks, requiring optimization as the system scales.
• Human Alignment: Direct comparison with human PANAS scores is necessary to empirically calibrate and validate the agent’s emotional responses.

Planned improvements are:

• Exploring alternative LLMs (e.g., Llama, Mistral) for more human-aligned affect annotation.
• Enhancing memory retrieval mechanisms for more precise context selection.
• Systematic benchmarking with human evaluators for validation.

7. Theoretical and Practical Implications

The architecture introduces a means for generative agents to acquire context-aware emotional reactivity by formally embedding appraisal-based comparison and episodic memory into the emotion circuit. This design provides a reproducible framework that can be instantiated in any sufficiently capable LLM, enabling both systematic emotional testing and improved alignment to human affective reasoning.

The primary theoretical implication is the demonstration that LLMs, when equipped with explicit appraisal and memory mechanisms, can more closely approximate the cognitive processes underlying human emotional experience. Practically, this approach sets a foundation for more nuanced, context-adaptive emotion modeling in AI systems, with direct applications in empathetic dialogue agents, AI companions, and affective computing platforms.