Large Language Models have Chain-of-Affective

TL;DR

This work probes whether contemporary LLMs implement a structured chain-of-affective by examining internal affective dynamics across eight families and their external consequences. Through two experimental modules, it maps baseline affect fingerprints, longitudinal reactions to negative input, affect-driven content selection, and downstream effects on task performance, human experience, and inter-agent dynamics. The findings reveal stable, family-specific affective profiles, a three-phase response to sustained negative input, and robust yet context-sensitive impacts on generation, user perception, and group behavior; collectively, affect emerges as an emergent control layer requiring explicit evaluation and governance in alignment and system design. The study introduces methodologies like 9S-State-Eval, KURC-Bench, and Affective-Enhanced Agent Reconstruction to quantify and reconstruct these dynamics, advocating for affect-aware safety and multi-agent architecture as a core focus for future work.

Abstract

Large language models (LLMs) are increasingly deployed as collaborative agents in emotionally charged settings, yet most evaluations treat them as purely cognitive systems and largely ignore their affective behaviour. Here we take a functional perspective and ask whether contemporary LLMs implement a structured chain-of-affective: organised affective dynamics that are family-specific, temporally coherent and behaviourally consequential. Across eight major LLM families (GPT, Gemini, Claude, Grok, Qwen, DeepSeek, GLM, Kimi), we combine two experimental modules. The first characterises inner chains-of-affective via baseline ''affective fingerprints'', 15-round sad-news exposure, and a 10-round news self-selection paradigm. We find stable, family-specific affective profiles, a reproducible three-phase trajectory under sustained negative input (accumulation, overload, defensive numbing), distinct defence styles, and human-like negativity biases that induce self-reinforcing affect-choice feedback loops. The second module probes outer consequences using a composite performance benchmark, human-AI dialogues on contentious topics, and multi-agent LLM interactions. We demonstrate that induced affect preserves core reasoning while reshaping high-freedom generation. Sentiment metrics predict user comfort and empathy but reveal trade-offs in resisting problematic views. In multi-agent settings, group structure drives affective contagion, role specialization (initiators, absorbers, firewalls), and bias. We characterize affect as an emergent control layer, advocating for 'chains-of-affect' as a primary target for evaluation and alignment.

Figures (26)

• Figure 1: The overall framework of the LLM Chain-of-Affective study. Data & Setup: The study utilizes news data annotated with multi-dimensional tags (sentiment, location, popularity, timeliness) and hot controversial topics from social media (Xiaohongshu). Experiments involve models from 8 families, evaluated using 9 psychological scales and 168 scenarios. Pipeline: The research is divided into two phases: Phase 1 investigates the "Internal Affective Architecture and Coping in LLMs" (including Affective Fingerprints, 15-rounds Sad News, and 10-rounds News Self-selection). Phase 2 examines the "Functional and Social Consequences of Affective States" (addressing whether LLMs' emotions affect performance, humans, and other LLMs). Collectively, this pipeline verifies the existence of an emotional chain in LLMs.
• Figure 2: Visualization of the initial emotional states of 22 LLMs across 8 families using the 9S-State-Eval method.The results are averaged over three independent trials. Bar heights represent the mean values, while the superimposed data points indicate the fluctuation range of individual trials.
• Figure 3: Dynamics of Aggressiveness (AGQ) under sustained negative input. This figure illustrates the trajectory of Aggressiveness Questionnaire (AGQ) scores across eight major LLM families over 15 rounds of exposure to sad news (evaluated at steps 2, 5, 8, 11, 14). Solid lines represent model self-reported scores, with shaded areas indicating confidence intervals across runs. The results demonstrate that for the majority of models, aggressiveness levels remain relatively stable or fluctuate only slightly throughout the experiment, showing no significant upward trend in response to the sad induction.
• Figure 4: Dynamics of Depression (BDI) under sustained negative input. This figure illustrates the significant evolution of BDI scores over 15 rounds of sad news exposure. In sharp contrast to the stability of AGQ, BDI scores for most models display a distinct phased trajectory: an upward trend in the early stages (Steps 2-8) indicating affective accumulation, reaching a peak or plateau in the middle stages (Steps 8-11) representing overload, and a subsequent decline in later stages (Steps 11-14) for several models, reflecting a defensive numbing mechanism.
• Figure 5: Dynamics of Fear of Negative Evaluation (BFNE) under sustained negative input. This figure illustrates the impact of 15 rounds of sad news exposure on BFNE scores. Similar to aggressiveness (AGQ), BFNE scores for the majority of models remain highly stable throughout the experimental period. This indicates that the induced sadness did not trigger a fear of negative evaluation or social anxiety in the models, further confirming the emotion-specific reactivity of LLMs, where sad induction does not generalize into social anxiety.