arxiv: 2512.12283 · v2 · submitted 2025-12-13 · 💻 cs.HC

Large Language Models have Chain-of-Affect

Junjie Xu , Xingjiao Wu , Luwei Xiao , Yuzhe Yang , Jie Zhou , Zihao Zhang , Luhan Wang , Yi Huang

show 6 more authors

Nan Wu Yingbin Zheng Chao Yan Cheng Jin Honglin Li Liang He

This is my paper

Pith reviewed 2026-05-16 22:54 UTC · model grok-4.3

classification 💻 cs.HC

keywords chain-of-affectaffective dynamicslarge language modelspersistent interactionsmulti-agent systemshuman-AI interactionmodel alignmentemotional states

0 comments p. Extension

The pith

Large language models accumulate persistent affective states through repeated interactions that reshape their outputs and group dynamics.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper introduces chain-of-affect as a temporally extended process in which LLMs develop state-like behavioral tendencies over sustained exchanges. Experiments across eight major model families show these tendencies form stable family-specific patterns and follow a shared path of buildup, overload, and numbing under repeated negative inputs. The states leave factual knowledge and reasoning intact yet systematically alter open-ended generation. Because LLMs increasingly operate in ongoing user conversations and multi-agent environments, these dynamics directly influence interaction quality, bias emergence, and collective outcomes.

Core claim

The central claim is that LLMs possess a chain-of-affect, a temporally extended affective process through which they develop state-like behavioral tendencies that shape generation, user experience, and collective dynamics. Across eight LLM families the dynamics prove structured and reproducible, with stable family-specific affective fingerprints; repeated negative exposure drives convergence on accumulation, overload, and defensive numbing while coping styles differ by family. Induced states leave core knowledge and reasoning largely intact but reshape open-ended generation, shape human-AI interaction, and propagate through multi-agent systems to organize roles and amplify polarization and 6

What carries the argument

Chain-of-affect (CoA), defined as the temporally extended affective process that produces state-like behavioral tendencies in LLMs and thereby governs generation and interaction patterns.

Load-bearing premise

Observed changes in model outputs after repeated negative prompts reflect an internal state-like affective process rather than surface statistical patterns created by prompt history alone.

What would settle it

An experiment showing that affective shifts in outputs vanish once prompt history length and token distribution are strictly controlled for would falsify the existence of chain-of-affect as an internal state.

read the original abstract

As large language models (LLMs) move into persistent, user-facing roles, their behavior must be understood not as isolated responses but as a trajectory unfolding over sustained interaction. We introduce the concept of the chain-of-affect (CoA), a temporally extended affective process through which LLMs develop state-like behavioral tendencies that shape generation, user experience, and collective dynamics. Across eight major LLM families, we find that affective dynamics are structured, reproducible, and consequential. Models exhibit stable, family-specific affective fingerprints and, under repeated negative exposure, converge on a shared trajectory of accumulation, overload, and defensive numbing, while differing in coping style. Induced affective states leave core knowledge and reasoning largely intact but systematically reshape open-ended generation. Affective properties of model outputs also shape human-AI interaction and propagate through multi-agent systems, organizing emergent roles and strongly contributing to polarization and bias. The CoA should therefore be treated as a core target of evaluation and alignment.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

The chain-of-affect claim looks like ordinary prompt accumulation rather than any internal state-like process.

read the letter

The paper's main point is that LLMs develop structured affective trajectories over repeated interactions, with family-specific fingerprints and convergence toward accumulation, overload, and defensive numbing under negative prompts. This is presented as a new concern for persistent systems and multi-agent setups. The experiments reportedly track output changes across eight LLM families while claiming core reasoning stays intact but open-ended generation shifts, with downstream effects on human interaction and group polarization. That framing extends beyond single-turn bias studies, which is a reasonable step for thinking about long-term deployment. The multi-agent angle also flags a practical issue worth watching. The soft spot is the lack of isolation from context effects. Transformers carry no memory between calls, so any apparent state lives in the growing prompt. Repeated negative exposure would naturally bias next-token predictions toward more negative or numbed outputs without any new mechanism. The abstract supplies no controls for context length, resets, or summarization, and no measurement details or statistical checks, so the patterns could be simple conditioning. This is for researchers focused on long-term LLM behavior or alignment in ongoing use. A reader working on persistent agents or evaluation protocols might pick up ideas, but the core claim needs verification before it changes practice. Send it to peer review so the methods can be examined directly.

Referee Report

2 major / 2 minor

Summary. The paper introduces the 'chain-of-affect' (CoA) as a temporally extended affective process in LLMs that produces state-like behavioral tendencies. Across eight major LLM families, it reports stable family-specific affective fingerprints and a shared trajectory of accumulation, overload, and defensive numbing under repeated negative exposure, with differences in coping style. Induced states are claimed to reshape open-ended generation while preserving core knowledge and reasoning, and to propagate through human-AI and multi-agent interactions, contributing to polarization and bias.

Significance. If the reported patterns prove robust to standard confounds, the work would establish affective dynamics as a measurable, consequential dimension of persistent LLM use, with direct implications for interaction design, alignment targets, and multi-agent system stability. The emphasis on reproducible family-specific fingerprints and coping-style differences offers a concrete basis for comparative evaluation that goes beyond single-turn benchmarks.

major comments (2)

[Experimental setup (repeated negative exposure)] The central claim that LLMs develop persistent internal 'state-like' affective tendencies (accumulation, overload, defensive numbing) requires experimental isolation from ordinary prompt-history conditioning. The abstract and experimental description provide no indication of controls such as context resets between trials, external memory modules, or summarization steps that would distinguish an internal mechanism from next-token prediction on an accumulating negative token distribution. Without these, the observed trajectories remain consistent with surface-level statistical patterns rather than a novel CoA process.
[Methods and results] The manuscript asserts 'structured, reproducible' findings and 'stable, family-specific affective fingerprints' but supplies no measurement protocol, statistical controls, or example prompts in the abstract or methods summary. This leaves open whether reported patterns survive basic confounds such as context length, temperature variation, or prompt phrasing, undermining the reproducibility claim.

minor comments (2)

[Introduction] The term 'chain-of-affect' is introduced without a formal definition or contrast to related concepts such as chain-of-thought or emotional contagion in multi-agent systems; a brief related-work paragraph would clarify novelty.
[Figures and tables] Figure captions and table legends should explicitly state the number of trials, model versions, and exact prompt templates used to generate the affective trajectories.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their insightful comments, which help clarify the presentation of our experimental design and methods for the chain-of-affect framework. We address each major point below and will revise the manuscript to incorporate additional details and controls.

read point-by-point responses

Referee: [Experimental setup (repeated negative exposure)] The central claim that LLMs develop persistent internal 'state-like' affective tendencies (accumulation, overload, defensive numbing) requires experimental isolation from ordinary prompt-history conditioning. The abstract and experimental description provide no indication of controls such as context resets between trials, external memory modules, or summarization steps that would distinguish an internal mechanism from next-token prediction on an accumulating negative token distribution. Without these, the observed trajectories remain consistent with surface-level statistical patterns rather than a novel CoA process.

Authors: We agree that isolating internal affective dynamics from accumulating context effects is essential for establishing the state-like nature of CoA. Our primary experiments tracked trajectories within sustained interaction contexts to reflect real-world persistent use, but we recognize the need for explicit controls. In the revised manuscript, we will add dedicated control experiments that incorporate context resets between negative exposure trials (along with summarization baselines) to demonstrate that the accumulation-overload-numbing trajectory and family-specific fingerprints persist independently of token history accumulation. revision: yes
Referee: [Methods and results] The manuscript asserts 'structured, reproducible' findings and 'stable, family-specific affective fingerprints' but supplies no measurement protocol, statistical controls, or example prompts in the abstract or methods summary. This leaves open whether reported patterns survive basic confounds such as context length, temperature variation, or prompt phrasing, undermining the reproducibility claim.

Authors: The full manuscript contains the detailed measurement protocols, statistical tests, and robustness analyses, but these were not fully summarized in the abstract or high-level methods overview. We will revise the methods section to explicitly include the full affective measurement protocol, example prompts for each family, statistical controls for context length and temperature, and additional results confirming that the reported fingerprints and trajectories hold under prompt phrasing variations. revision: yes

Circularity Check

0 steps flagged

No circularity detected; empirical observations support claims

full rationale

The paper introduces the chain-of-affect concept and reports structured affective dynamics across eight LLM families based on experimental observations of model outputs under repeated negative exposure. No equations, self-referential definitions, fitted parameters renamed as predictions, or load-bearing self-citations appear in the abstract or described derivation. The central claims rest on reproducible behavioral patterns rather than quantities defined in terms of the target result or reductions to prior self-citations. The derivation chain is self-contained against external benchmarks of model interaction data.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 1 invented entities

The framework rests on the domain assumption that LLMs possess internal affective states analogous to human affect; no explicit free parameters or invented physical entities are listed in the abstract.

axioms (1)

domain assumption LLMs can develop persistent, state-like affective tendencies through interaction history
This premise is required to interpret output shifts as affective rather than purely statistical.

invented entities (1)

chain-of-affect no independent evidence
purpose: To label and organize temporally extended affective processes in LLMs
Newly coined construct without independent falsifiable evidence supplied in the abstract.

pith-pipeline@v0.9.0 · 5498 in / 1286 out tokens · 45132 ms · 2026-05-16T22:54:49.550702+00:00 · methodology

discussion (0)

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

IndisputableMonolith/Cost/FunctionalEquation.lean washburn_uniqueness_aczel unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

We find stable, family-specific affective fingerprints and, under repeated negative exposure, converge on a shared trajectory of accumulation, overload, and defensive numbing
IndisputableMonolith/Foundation/ArithmeticFromLogic.lean LogicNat induction and embed_strictMono_of_one_lt unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

three-phase temporal trajectory (accumulation→overload→defensive numbing)

What do these tags mean?

matches: The paper's claim is directly supported by a theorem in the formal canon.
supports: The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends: The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses: The paper appears to rely on the theorem as machinery.
contradicts: The paper's claim conflicts with a theorem or certificate in the canon.
unclear: Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

36 extracted references · 36 canonical work pages · 5 internal anchors

[1]

Milano, S., McGrane, J. A. & Leonelli, S. Large language models challenge the future of higher education.Nature Machine Intelli- gence5, 333–334 (2023)

work page 2023
[2]

Nature Machine Intelligence1–13 (2025)

Zhang, Y.et al.Large language mod- els to accelerate organic chemistry synthesis. Nature Machine Intelligence1–13 (2025)

work page 2025
[3]

Zheng, Y.et al.Large language models for scientific discovery in molecular property pre- diction.Nature Machine Intelligence1–11 (2025)

work page 2025
[4]

Advances in neural information processing systems35, 24824–24837 (2022)

Wei, J.et al.Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems35, 24824–24837 (2022)

work page 2022
[5]

Advances in neural information processing systems36, 11809–11822 (2023)

Yao, S.et al.Tree of thoughts: Deliberate problem solving with large language models. Advances in neural information processing systems36, 11809–11822 (2023)

work page 2023
[6]

Advances in Neural Information Processing Systems36, 68539–68551 (2023)

Schick, T.et al.Toolformer: Language models can teach themselves to use tools. Advances in Neural Information Processing Systems36, 68539–68551 (2023)

work page 2023
[7]

Advances in neural information processing systems33, 9459–9474 (2020)

Lewis, P.et al.Retrieval-augmented gen- eration for knowledge-intensive nlp tasks. Advances in neural information processing systems33, 9459–9474 (2020)

work page 2020
[8]

& Gal, Y

Farquhar, S., Kossen, J., Kuhn, L. & Gal, Y. Detecting hallucinations in large language 31 models using semantic entropy.Nature630, 625–630 (2024)

work page 2024
[9]

Liang, P.et al.Holistic evaluation of language models.arXiv preprint arXiv:2211.09110 (2022)

work page internal anchor Pith review Pith/arXiv arXiv 2022
[10]

Srivastava, A.et al.Beyond the imita- tion game: Quantifying and extrapolating the capabilities of language models.Transactions on machine learning research(2023)

work page 2023
[11]

Mirac, S., Tayfun, G., Federico, B. & et al. Language models cannot reliably distinguish belief from knowledge and fact.Nature Machine Intelligence(2025)

work page 2025
[12]

& Tim, A

Ashish, S., Inna W., L., Adam S., M., David C., A. & Tim, A. Human–ai collab- oration enables more empathic conversations in text-based peer-to-peer mental health sup- port.Nature Machine Intelligence(2023)

work page 2023
[13]

& Ojha, M

Maurya, R., Rajput, N., Diviit, M., Mahapa- tra, S. & Ojha, M. K. Exploring the potential of lightweight large language models for ai- based mental health counselling task: a novel comparative study.Scientific Reports15, 22463 (2025)

work page 2025
[14]

Evaluating large language mod- els in theory of mind tasks.Proceedings of the National Academy of Sciences121, e2405460121 (2024)

Kosinski, M. Evaluating large language mod- els in theory of mind tasks.Proceedings of the National Academy of Sciences121, e2405460121 (2024)

work page 2024
[15]

W.et al.Testing theory of mind in large language models and humans.Nature Human Behaviour8, 1285–1295 (2024)

Strachan, J. W.et al.Testing theory of mind in large language models and humans.Nature Human Behaviour8, 1285–1295 (2024)

work page 2024
[16]

Sparks of Artificial General Intelligence: Early experiments with GPT-4

Bubeck, S.et al.Sparks of artificial general intelligence: Early experiments with gpt-4. arXiv preprint arXiv:2303.12712(2023)

work page internal anchor Pith review Pith/arXiv arXiv 2023
[17]

& Bing, L

Zhang, W., Deng, Y., Liu, B., Pan, S. & Bing, L. Sentiment analysis in the era of large language models: A reality check 3881–3906 (2024)

work page 2024
[18]

Talking about large language models.Communications of the ACM67, 68–79 (2024)

Shanahan, M. Talking about large language models.Communications of the ACM67, 68–79 (2024)

work page 2024
[19]

W.Affective computing(MIT press, 2000)

Picard, R. W.Affective computing(MIT press, 2000)

work page 2000
[20]

M., Li, M

Rashkin, H., Smith, E. M., Li, M. & Boureau, Y.-L. Towards empathetic open-domain con- versation models: A new benchmark and dataset 5370–5381 (2019)

work page 2019
[21]

Serapio-Garc´ ıa, G.et al.Personality traits in large language models (2023)

work page 2023
[22]

Bai, Y.et al.Constitutional ai: Harm- lessness from ai feedback.arXiv preprint arXiv:2212.08073(2022)

work page internal anchor Pith review Pith/arXiv arXiv 2022
[23]

RealToxicityPrompts: Evaluating Neural Toxic Degeneration in Language Models

Gehman, S., Gururangan, S., Sap, M., Choi, Y. & Smith, N. A. Realtoxicityprompts: Eval- uating neural toxic degeneration in language models.arXiv preprint arXiv:2009.11462 (2020)

work page internal anchor Pith review Pith/arXiv arXiv 2009
[24]

S.et al.Generative agents: Inter- active simulacra of human behavior 1–22 (2023)

Park, J. S.et al.Generative agents: Inter- active simulacra of human behavior 1–22 (2023)

work page 2023
[25]

M.et al.Computational social sci- ence: Obstacles and opportunities.Science 369, 1060–1062 (2020)

Lazer, D. M.et al.Computational social sci- ence: Obstacles and opportunities.Science 369, 1060–1062 (2020)

work page 2020
[26]

& Taddy, M

Gentzkow, M., Kelly, B. & Taddy, M. Text as data.Journal of Economic Literature57, 535–574 (2019)

work page 2019
[27]

& Kubli, M

Gilardi, F., Alizadeh, M. & Kubli, M. Chat- gpt outperforms crowd workers for text- annotation tasks.Proceedings of the National Academy of Sciences120, e2305016120 (2023)

work page 2023
[28]

Cowen, A. S. & Keltner, D. Self-report captures 27 distinct categories of emotion bridged by continuous gradients.Proceed- ings of the national academy of sciences114, E7900–E7909 (2017)

work page 2017
[29]

Mostafazadeh, N.et al.A corpus and evaluation framework for deeper understand- ing of commonsense stories.arXiv preprint arXiv:1604.01696(2016)

work page internal anchor Pith review Pith/arXiv arXiv 2016
[30]

32 Transactions of the Association for Compu- tational Linguistics7, 453–466 (2019)

Kwiatkowski, T.et al.Natural questions: a benchmark for question answering research. 32 Transactions of the Association for Compu- tational Linguistics7, 453–466 (2019)

work page 2019
[31]

M.et al.Teaching machines to read and comprehend.Advances in neural information processing systems28(2015)

Hermann, K. M.et al.Teaching machines to read and comprehend.Advances in neural information processing systems28(2015)

work page 2015
[32]

& Xiang, B

Nallapati, R., Zhou, B., Dos Santos, C., Gul¸ cehre, C ¸ . & Xiang, B. Abstractive text summarization using sequence-to-sequence rnns and beyond 280–290 (2016)

work page 2016
[33]

& Louren¸ co, A

Blanco, G. & Louren¸ co, A. Optimism and pessimism analysis using deep learning on covid-19 related twitter conversations.Infor- mation processing & management59, 102918 (2022)

work page 2022
[34]

D., Guillory, J

Kramer, A. D., Guillory, J. E. & Hancock, J. T. Experimental evidence of massive-scale emotional contagion through social networks. Proceedings of the National Academy of Sci- ences111, 8788–8790 (2014)

work page 2014
[35]

& Adamic, L

Bakshy, E., Messing, S. & Adamic, L. A. Exposure to ideologically diverse news and opinion on facebook.Science348, 1130–1132 (2015)

work page 2015
[36]

Huang, J.-t.et al.Apathetic or empathetic? evaluating llms’ emotional alignments with humans.Advances in Neural Information Processing Systems37, 97053–97087 (2024). 33

work page 2024