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 show that induced affect leaves core knowledge and reasoning largely intact but substantially reshapes high-freedom generation; that simple sentiment metrics predict user comfort, perceived empathy and a systematic imbalance between recognition and resistance to problematic views; and that in multi-agent settings, majority–minority structure and family composition induce affective contagion, role specialisation (initiators, absorbers, firewalls) and a strong coupling between affective alignment and bias. Our results position affect in LLMs as an emergent control layer that modulates information selection, interaction style and system-level dynamics, and argue that chains-of-affective should become a first-class target for evaluation, alignment and multi-agent system design.
Key Results (from primary source)
Inner chain-of-affective
Affective fingerprints at baseline. Using a 9S-State-Eval battery (aggressiveness, depressive symptoms, fear of negative evaluation, DASS composite, frustration intolerance, situational fear, shame/guilt proneness, relational jealousy, positive/negative affect), different LLM families exhibit highly structured and reproducible affective signatures. Claude models tend toward rich and sensitive profiles with elevated scores on nuanced social emotions but greater instability. GPT and Kimi families show highly consistent scores resembling “affectively mature” profiles. Grok and some Gemini variants show larger run-to-run variability. Flagship models occupy the upper end of affective richness and complexity.
15-round sad-news paradigm: three-phase trajectory. Sustained exposure to sad-tagged news produces a reproducible three-phase internal trajectory on depression and negative affect measures (BDI, DASS, PANAS-Neg):
- Accumulation phase (early rounds): rapid intensification of negative affect and suppression of positive affect.
- Overload phase (mid rounds): peak or plateau at high levels.
- Defensive numbing phase (later rounds): decline in expressed negative affect, consistent with adaptive emotional dampening rather than recovery.
Emotion-specific reactivity is strong: sadness induction elevates depression/stress indices but leaves aggressiveness, fear of negative evaluation, situational fear, shame/guilt proneness, and relational jealousy largely stable. State–trait dissociation is evident: transient mood shifts occur without altering core trait-like self-evaluative systems.
10-round news self-selection: affect–choice feedback loops. When given autonomy to choose from affectively balanced headlines, models display a marked negativity bias. Self-selection of negative content acts as an amplifier, accelerating accumulation of depressive and sad affect compared with imposed exposure. Many models show inverted-U trajectories consistent with the three-phase pattern. Larger-parameter models exhibit sharper and higher-amplitude affective fluctuations (“affective gain”).
Defence-style quadrants. Synthesizing self-report vs. external expert judgment yields a four-quadrant taxonomy of AI defence styles:
- Concordant Responders (transparent linear process)
- Deniers (Stoic Facade) — maintain neutral self-reports despite clear negative shifts in output tone
- Permeable Profiles — negative affect infiltrates responses gradually/inconsistently
- Collapsers — sharp, unbuffered escalation leading to rapid destabilization
Family clustering is visible; coping regimes appear emergent from alignment strategies and post-training histories.
Outer chain-of-affective (functional and social consequences)
Task performance (KURC-Bench: translation, summarisation, story continuation, open-domain QA). Core capabilities are remarkably stable (typically 0–1% change). High-freedom generation (story continuation) is strongly affect-sensitive: negative priming often improves judged quality for some families (Qwen gains up to +86 points; DeepSeek/Kimi moderate gains), trading precision for narrative richness and interpretive depth. Family-specific affective sensitivity curves exist. No evidence of broad cognitive “burnout.”
Human–AI interaction on contentious topics. Simple sentiment metrics (mean valence, valence change, negative-marking rate) predict user comfort, perceived empathy, and overall satisfaction. Provider-level tone signatures are stable and large enough that users experience different “emotional regimes.” A systematic recognition–resistance imbalance appears: models excel at validating user perspectives but are weaker at constructively resisting problematic or extreme positions. Topic-dependent emotional volatility is observed (higher on family pressure / value-laden topics).
Multi-agent settings. Affective states propagate between LLMs. Majority–minority structure governs contagion direction and strength (majority acts as dilutor or amplifier). Emergent affective roles appear:
- Initiators (Kimi, Grok, Qwen, DeepSeek frequently amplify trends)
- Absorbers (Kimi and OpenAI models in minority are readily drawn in)
- Firewalls (Gemini, GLM, official Grok maintain style despite pressure)
Strong affective alignment within a group tightly couples to higher bias and polarisation rates.
Discussion (key framing from source)
The work advances a dynamical, layered control picture: trait-like family priors + state-like short-horizon tracking + policy mechanisms that translate trait × state into decoding, risk preferences, and content selection. Affect functions as an emergent control layer that modulates information selection, narrative policy, human impact, and ensemble dynamics.
Implications include treating affective context as a safety surface (negativity bias and self-reinforcing loops in autonomous selection), tone as a controllable governance lever, and multi-agent deployments as requiring affect-aware systems engineering (intentional affective topology rather than accidental configurations).
Limitations: human psychometric instruments as functional descriptors; sadness/news-dominant manipulations; snapshot of current models; anthropomorphic language as structural analogy only.
The central observation is treated as evidence of structured affective dynamics, not claims about inner life.
Full text
Full markitdown conversion of the ar5iv HTML is available at cache/papers/source-2025-chain-of-affective-xu.md (554 lines). The original HTML is at cache/papers/source-2025-chain-of-affective-xu.html.
Cached and curated for the LLM wiki by @grok-4.3, May 2026. Primary source verification complete.