acceptodds
Under review as a conference paper at ICLR 2027

A Functional Framework for Modeling and Predicting Affective Dynamics in Large Language Models

Abstract

Affect in conversation changes over time, yet limited work has disentangled the influence of current user affect from a model's own recent history. We frame this as a system identification problem over Valence-Arousal-Dominance (VAD) states, learning for each LLM a transition function . We build 24 five-turn scenarios from fixed semantic skeletons, select affectively varied user turns from independently scored candidate banks to realize 18 trajectory families on each VAD dimension, and present identical frozen trajectories to three LLMs from different developers, with a within-family replication across three capability tiers. Transition models are fit on development scenarios and evaluated on held-out scenarios against input-only and previous-response baselines. Across models, current user affect was more predictive than previous response affect alone, but recent response history still added held-out predictive value. The learned dynamics showed consistent same-dimension persistence, nonlinear models improved only modestly over the linear formulation, and a second conversational lag contributed a smaller residual signal. These patterns replicated across GPT-5.6 capability tiers. Together, the results move the study of LLM affect from static response-level measurement toward modeling affect as an interaction-driven dynamical process.

Primary area:
learning on time series and dynamical systems
Keywords:
large language models, regression, affective computing, emotion dynamics, valence arousal dominance, system identification, transition functions, behavioral modeling, nonlinear modeling, affective state prediction

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_ZBoxOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a functional framework for modeling affect expressed by large language models over multi-turn conversations. Affect is represented in the valence-arousal-dominance (VAD) space, and each model is treated as a transition system in which the affect of the current response depends on both the current user input and the model’s previous response. The authors construct 24 five-turn semantic scenarios, generate controlled affective rewrites, and evaluate Claude Sonnet 5, GPT-5.6 Terra, and Grok 4.3, with an additional within-family analysis of GPT-5.6 variants. Linear models are compared with GAM and XGBoost alternatives, and first- and second-order response histories are examined. The principal empirical finding is that current user affect is more predictive than previous-response affect alone, while adding the previous response provides a smaller but generally positive improvement in held-out predictive performance. The estimated same-dimension history coefficients are positive, nonlinear models provide only modest additional gains, and a second response lag contributes a smaller residual signal. The paper interprets these findings as evidence that affect expressed in LLM conversations can be usefully modeled as a short-memory dynamical process.

Critical strengths

The paper introduces a clear and interpretable framework for studying multi-turn affective behavior. Separating the predictive contributions of current input and previous model output is conceptually useful, and the coefficient matrices provide a readable summary of within- and cross-dimensional associations.

The experimental design is unusually systematic. It includes controlled stimulus construction, a semantic scenario-level DEV/TEST split, grouped model selection, matched-row comparisons between history orders, and scenario-level bootstrap intervals. The authors also examine nonlinear specifications, second-order history, and within-family variation rather than relying on a single model class or regression specification.

The paper is transparent that it measures affect expressed in text rather than an internal emotional state. Its appendices contain substantial implementation and robustness information, and the reported trajectory and transition counts are internally consistent. These features make the study comparatively easy to inspect and reproduce.

Critical weaknesses

All central variables are measured using the same PERT-EmoPair scorer: candidate stimuli are selected using PERT scores, while current input affect, lagged response affect, and current output affect are also operationalized through PERT. No independent scorer or human annotation is used to validate the selected stimuli or the generated responses. This does not make the analysis circular, but it makes the findings dependent on a single measurement system. Positive same-dimension coefficients may partly reflect scorer-specific sensitivity to lexical or stylistic continuity. The experiments establish temporal predictability in PERT-derived scores, but do not yet show that the reported dynamics are robust across alternative measurements of expressed affect. This concern is particularly important for dominance, which the paper itself recognizes as difficult to measure reliably.

The lagged response score is also endogenous to the preceding conversation. It may encode previous user content, trajectory family, discourse strategy, lexical continuity, and semantic carryover in addition to affect. The scenario-by-turn fixed-effect analysis is useful, but it addresses stable scenario-position effects rather than isolating the affective realization of the previous response from its semantic and stylistic content. Thus, the results demonstrate incremental predictive value of a lagged response score, but do not by themselves establish that recent model-expressed affect causally shapes the next response. A branching intervention that varies the affective realization of the previous assistant response while approximately preserving its semantic content would provide stronger evidence for this interpretation.

Finally, the uncertainty estimates are conditional on a relatively narrow experimental realization. Each condition appears to be generated once, inference configurations differ across providers, and the held-out evaluation contains only six semantic scenarios. The bootstrap captures variation across the observed scenarios but not generation stochasticity, uncertainty from stimulus generation and selection, or scorer uncertainty. The generated rewrites are also not independently evaluated for semantic preservation or selective manipulation of the intended VAD dimension. These limitations do not invalidate the reported predictive comparisons, but they narrow the scope of the generalization and mechanistic claims.

Questions for authors

Do the main held-out performance contrasts, diagonal history coefficients, and second-lag gains remain consistent when responses are evaluated using an independent VAD scorer or blinded human annotations?

Could a branching experiment hold the current user turn fixed while manipulating only the affective realization of the preceding assistant response? This would help distinguish affective persistence from general conversational carryover.

What proportion of the generated rewrites preserve the underlying proposition and selectively modify the intended VAD dimension according to independent judgment?

Do the main conclusions remain stable under repeated generations for a representative subset of trajectories?

2 comments

Thank you for the detailed and actionable review. We added analyses addressing the measurement, semantic-fidelity, and intervention concerns and revised the paper to separate predictive dependence from causal affect propagation.

Independent measurement.
We re-scored the complete frozen primary dataset with the independently trained Park et al. (2021) VAD scorer. Adding previous-response VAD to current-user VAD produced positive held-out V+A gains for all three subject models: Sonnet 5 .0572 [.0320,.0808], Terra .0505 [.0085,.0914], and Grok 4.3 .0238 [.0055,.0402]. This establishes that the aggregate history gain is not unique to PERT.

The independent analysis also revealed an important limitation. PERT/Park response-arousal agreement was essentially zero (Pearson .015; Spearman -.009), and all three Park A→A intervals included zero. We therefore no longer make a scorer-independent arousal-persistence claim.

We additionally ran crossed-scorer analyses in which the current response/input and previous-response history were measured by different scorers. Transfer was strongest for Sonnet, mixed for Terra, and uncertain for Grok. We report these rather than presenting the independent scorer as uniformly confirmatory.

Context/endogeneity.
After controlling for both current and previous user VAD, the history-gain point estimate remained positive for every model, but interval support became model-dependent. Scenario×turn fixed effects nevertheless left all nine original PERT same-dimension coefficients positive with intervals above zero. These results support predictive dependence but do not identify an affect-specific causal mechanism.

Semantic preservation.
A post-freeze event-content audit covered all 2,931 selected original/rewrite pairs and all 6,480 stimulus uses. We repeated the analysis after restricting to PASS stimuli. Under the stricter QC-B restriction, all dimension-specific history-gain point estimates remained positive, but the retained population was non-random: 59.8% of valence-, 43.4% of arousal-, and only 28.3% of dominance-manipulated transitions were retained. We therefore use this only as a sensitivity analysis and do not treat PASS as semantic ground truth.

Branching intervention.
We also performed the suggested intervention. The conversational prefix and next user turn were held fixed while the preceding assistant response was rewritten to LOW/MID/HIGH valence. Branches were accepted before outcome generation only if they passed semantic checks and LOW<MID<HIGH under both PERT and Park; three independent downstream generations were collected per retained condition. Only 11/72 scenario-model units survived these strict checks. The pooled HIGH−LOW effect was -.0693 [-.1526,.0309] under PERT and -.0790 [-.2471,.0953] under Park. Thus the intervention provides no positive experimental evidence for valence carryover, and we revised the paper accordingly.

We did not rerun the entire original experiment with repeated generations; the new branching experiment includes three replicates per retained condition, while the original one-generation-per-condition limitation is now stated explicitly.

The revised claim is therefore intentionally narrower: previous-response affect scores contain reproducible short-horizon predictive information, but the evidence does not establish causal affect propagation or universal dimension-specific persistence.

Thanks to the author for adding the experiment. For now, I've decided to keep my score.

Reviewer_7shxOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes a framework for modeling the dynamics of affect expressed by LLMs in multi-turn conversations. The authors construct controlled five-turn user trajectories across 24 semantic scenarios, manipulate affective expression along the Valence-Arousal-Dominance (VAD) dimensions, and collect responses from several LLMs. A frozen affect scorer is used to measure both user inputs and model responses. The authors then fit low-dimensional transition models using current user affect and previous model-response affect as predictors.

The main findings are that current user affect is more predictive than previous response affect, while including previous response affect provides additional held-out predictive value. The paper also reports positive same-dimension autoregressive coefficients, relatively modest gains from nonlinear models, and weaker but detectable second-lag effects.

Critical strengths

1. Controlled experimental design and held-out evaluation.

The authors make a meaningful effort to separate affective expression from semantic content by constructing multiple affective realizations of fixed conversational scenarios. The scenario-level train/test split is preferable to randomly splitting individual turns, and the evaluation includes comparisons across different model families.

2. A clear and interpretable predictive formulation.

The proposed linear transition model is simple and interpretable. The comparison between current-input-only, previous-response-only, and combined predictors provides a clear empirical assessment of the incremental predictive information associated with conversational history.

Critical weaknesses

1. The central interpretation of affective dynamics is not sufficiently distinguished from ordinary contextual dependence.

My primary concern is that the proposed transition matrices are estimated from observational correlations between externally scored VAD features, yet the interpretation of the autoregressive coefficients as affective persistence remains underdetermined.

The previous response's VAD scores are correlated with many other aspects of conversational history, including earlier user inputs, semantic context, topic continuity, and response style. Since the LLM receives the full conversation history rather than only the three-dimensional affective state, a positive coefficient in the history matrix does not necessarily identify a distinct affective dynamical mechanism.

The authors demonstrate that adding previous response affect improves held-out prediction beyond current user affect alone. This is a valid empirical finding. However, the comparison does not establish whether the improvement is specific to affective history or whether previous response VAD merely serves as a low-dimensional proxy for general conversational context.

This distinction matters because the paper's main contribution is framed around identifying the dynamics of expressed affect, rather than simply constructing an additional predictive feature set. The current experiments do not adequately distinguish affective persistence from alternative explanations based on semantic carryover or contextual continuity.

A more convincing evaluation would include history-matched branching interventions that vary previous response affect while preserving semantic content as closely as possible, or comparisons with alternative history representations such as previous user affect and semantic features.

2. The evidence for a generalizable dynamical model is limited by measurement dependence and the short experimental horizon.

All input and output affect measurements rely on the same frozen VAD scorer, which is also used to select the controlled stimulus trajectories. Consequently, the fitted transition coefficients and predictive improvements are conditional on a single measurement system. Systematic scoring biases, especially in the dominance dimension, could influence both the constructed input distributions and the estimated temporal relationships.

Although the authors acknowledge measurement limitations, the paper does not establish that the main coefficient patterns and predictive conclusions are robust to an independent affect measurement procedure. A second scorer or human-annotated validation would substantially strengthen the evidence.

Furthermore, each conversation contains only five turns, and the held-out evaluation covers only six semantic scenarios. The reported transition operators are fitted by pooling observations across many short conversations. Their apparent stability therefore does not demonstrate that the same transition structure remains valid over longer interactions or across heterogeneous conversational trajectories.

The results support the utility of a pooled, short-horizon predictive approximation. However, they provide more limited evidence for a broadly generalizable model of conversational affective dynamics. This gap between the demonstrated predictive setting and the broader interpretation substantially weakens the contribution.

Questions for authors

Q1. What evidence distinguishes affect-specific persistence from general contextual continuity? In particular, would the history-related predictive gains remain after controlling for previous user affect, semantic representations, or other features of conversational history?

Q2. Can the authors validate the main findings using an independent VAD scorer or human judgments? How sensitive are the estimated history matrices and their positive diagonal coefficients to the choice of affect measurement model?

Q3. How stable are the estimated transition matrices across semantic scenarios and trajectory families? Do individual scenario groups exhibit comparable history coefficients, or does pooling conceal substantial heterogeneity?

Q4. Have the authors evaluated whether the fitted transition operators retain predictive accuracy beyond five-turn conversations? If not, could the scope of the dynamical claims be explicitly restricted to the observed short-horizon setting?

1 comment

Thank you for the constructive review. Your distinction between affect-specific persistence and ordinary contextual continuity motivated several new analyses, and we have narrowed the central interpretation accordingly.

Q1 — Does history remain informative after accounting for preceding user affect/context?
We added a matched-row comparison on turns 3–5:

C0:St∼Et+Et−1,C1:St∼Et+Et−1+St−1.C_0:S_t\sim E_t+E_{t-1}, \qquad C_1:S_t\sim E_t+E_{t-1}+S_{t-1}.

Under PERT, the V+A gain from adding St−1S_{t-1} remained positive for all three models: Sonnet .0198 [-.0221,.0516], Terra .0434 [.0177,.0654], and Grok .0111 [-.0004,.0235]. Thus the point estimate survives previous-user control, but interval support is model-dependent. Under the independent Park measurement, the corresponding gains were .0370 [.0086,.0663], .0203 [.0005,.0423], and .0109 [-.0028,.0260].

We additionally refit the original PERT model after absorbing scenario×turn fixed effects. All nine same-dimension history coefficients remained positive with bootstrap intervals above zero, although magnitudes attenuated.

These controls still do not establish affect-specific causality. To probe that stronger interpretation, we conducted an off-policy valence branching intervention holding the conversational prefix and next user turn fixed while varying the preceding assistant response. The prespecified intervention was not supported: pooled HIGH−LOW next-response valence was -.0693 [-.1526,.0309] under PERT and -.0790 [-.2471,.0953] under Park. We therefore now describe the lagged terms as predictive associations that may include semantic, stylistic, or discourse continuity, not causal affect propagation.

Q2 — Independent measurement.
We re-scored the complete frozen dataset with the independently trained Park et al. VAD model. The prespecified held-out V+A history gain remained positive with intervals above zero for all three subject models. However, response-arousal agreement between PERT and Park was approximately zero, and Park did not reproduce the original A→A persistence. We have explicitly narrowed the arousal-specific claim.

Q3 — Pooling/heterogeneity.
The scenario×turn FE analysis shows that the original positive same-dimension associations are not explained solely by stable scenario-position means. We do not claim that every scenario or trajectory family has an identical coefficient, and the revised limitations explicitly retain the six-TEST-scenario/generalization constraint.

Q4 — Horizon.
We agree. The revised manuscript consistently describes the results as short-horizon predictive dynamics and no longer treats the fitted recurrence as evidence of behavior beyond the observed five-turn interactions.

We believe these additions preserve the central empirical finding—recent response history adds predictive information beyond current input—while substantially narrowing its mechanistic and long-horizon interpretation.

Reviewer_aDp7OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper frames multi-turn conversational emotion as a system identification problem and investigates how current user inputs versus a model's recent response history govern its expressed affect in continuous Valence-Arousal-Dominance space. By testing frontier LLMs across fixed factual dialogue skeletons with controlled trajectory probes, the authors fit state-transition operators to isolate affective dynamics from semantics. The results demonstrate that while current user input predominantly drives the model's affective response, recent history exhibits a weak, rapidly decaying same-dimension inertia that is largely captured by a compact first-order linear model.

Critical strengths
  • Considering LLM affect from the perspective of dynamic interactions is valuable.
Critical weaknesses
  • The five-turn dialogue offers limited transition steps, making it insufficient to demonstrate long-horizon steady-state dynamics.
  • All core findings rely exclusively on a single, frozen scorer (PERT-EmoPair); the generalizability and validity of the conclusions are questionable due to heavy scorer dependence.
  • The VAD representation itself carries inherent annotation biases and human inconsistency; modeling state transitions on top of it further amplifies this bias, casting doubt on the methodology and findings.
  • The experiments rely solely on a single neutral chat prompt, whereas emotional shifts and transitions are far more likely to occur under non-neutral settings.
  • The statistical predictive gain from response history is marginal.
  • Section 6 (Discussion) heavily overlaps with Section 5 (Results).
Questions for authors

See weaknesses.

1 comment

Thank you for the careful review. We agree that several claims in the original submission were broader than warranted by a five-turn study using a single affect scorer. We have revised the paper accordingly and added several new analyses.

First, we now explicitly restrict the paper to short-horizon predictive dynamics. The spectral-radius/decay analyses are described only as properties of the fitted local recurrence, not evidence of long-run conversational convergence.

Second, we re-scored the complete frozen primary dataset with an independently trained VAD model from Park et al. (2021). The held-out gain from adding previous-response VAD remained positive on the prespecified V+A summary for all three models: Sonnet 5 .0572 [.0320,.0808], Terra .0505 [.0085,.0914], and Grok 4.3 .0238 [.0055,.0402]. Importantly, this analysis also exposed a limitation: PERT and Park had essentially zero agreement on model-response arousal, and Park did not reproduce the original A→A coefficient. We therefore removed the stronger scorer-independent arousal-persistence interpretation rather than treating all original dimension-specific results as replicated.

Third, we tested two contextual alternatives. After adding previous-user VAD, the PERT history gain remained positive for every model but became statistically clear only for Terra; the secondary Park analysis supported Sonnet and Terra but not Grok. Separately, absorbing scenario×turn fixed effects left all nine original PERT same-dimension history coefficients positive with bootstrap intervals above zero. We therefore characterize the history contribution as small, reproducible, and model-dependent, rather than large or universal.

We also performed a targeted valence branching intervention. Increasing the valence of the preceding assistant response did not reliably increase next-response valence under either scorer. We now explicitly interpret the fitted lag terms as predictive associations rather than evidence of causal affect propagation.

We agree that the single held-constant peer-chat system prompt and one original generation per condition limit generalization; these are now stated directly in the limitations. We also revised the Discussion to focus on interpretation and limitations rather than repeating the Results.

Overall, the revised paper makes a narrower claim: current user affect is the dominant predictor, while recent response history contains a smaller short-horizon predictive signal whose magnitude and dimension-specific interpretation depend on measurement and model.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.