Dynamic Emotional Intelligence Suite

• Dynamic EI Evaluation Suite is a framework for assessing AI emotional intelligence through multi-turn, adaptive, and context-sensitive protocols.
• It decomposes the evaluation into four modules—sensing, explanation, response selection, and context adaptation—using tailored scoring mechanisms.
• The design emphasizes longitudinal interaction, adaptive weighting, and cultural diversity to enhance practical deployment in varied real-world scenarios.

Searching arXiv for the specified papers and closely related dynamic EI evaluation work. arXiv result (Parks et al., 29 Dec 2025): "Why We Need a New Framework for Emotional Intelligence in AI" — provides a self-contained specification for a “Dynamic Emotional Intelligence Evaluation Suite” (DEIES), including MDB/GEI, four modules, dynamic weighting, context-sensitivity scoring, and pseudocode workflow. Additional arXiv results confirm related dynamic EI evaluation frameworks: "EICAP" (Nazar et al., 8 Aug 2025), "AttuneBench" (Lubrano et al., 20 May 2026), "EIBench" (Zhu et al., 14 Jun 2026), "SpeechEQ" (Wu et al., 24 Jun 2026), "Multi-Bench" (Deng et al., 2 Nov 2025), and "Detecting Emotional Dynamic Trajectories" (Tan et al., 12 Nov 2025). Dynamic Emotional Intelligence Evaluation Suite denotes a class of evaluation frameworks that treat emotional intelligence in AI as an interactive, context-sensitive, and often multimodal capability rather than a one-shot labeling problem. In its explicit formulation as DEIES, the suite is organized as a two-part framework consisting of a Minimum Deployment Benchmark (MDB), which acts as a pass/fail safety floor, and a General EI Index (GEI), which provides a fine-grained $0\!-\!1$ competence profile across sensing, explanation, response selection, and context-adaptation (Parks et al., 29 Dec 2025). Closely related benchmarks instantiate the same shift through multi-turn MCQ dialogue, real multi-turn human-model conversations, speech-based forced choice, simulator-mediated emotion management, and trajectory scoring, making dynamic assessment a central pattern in current EI evaluation research (Nazar et al., 8 Aug 2025, Lubrano et al., 20 May 2026, Zhu et al., 14 Jun 2026).

1. Conceptual scope

The central theoretical move in DEIES is the distinction between phenomenal and functional EI. The framework explicitly discards “phenomenal consciousness” and “subjective feeling” as untestable in AI, while retaining functional analogues of human emotional processes. It therefore focuses on four core sub-abilities: appraisal reasoning, emotion cue sensing, socially appropriate prosocial response generation, and cross-turn and cross-cultural adaptation. At the same time, it deemphasizes self-report scales, trait-EI questionnaires, pure category-only tests, and one-shot, text-only labeling benchmarks (Parks et al., 29 Dec 2025).

This orientation is continuous with earlier theory-grounded work that defined machine emotional intelligence as the conjunction of Emotional Understanding and Emotional Application. In that formulation, Emotional Understanding covers perception, identification, and reasoning about emotions, including cause inference, while Emotional Application covers the use of that understanding to guide thought and action through effective responses or interventions (Sabour et al., 2024). A plausible implication is that dynamic suites do not replace earlier EI taxonomies so much as operationalize them under stronger ecological and temporal constraints.

The term “dynamic” has a specific methodological meaning in this literature. It refers to evaluation setups in which emotional state must be tracked or influenced across turns, under changing context, and often across modalities or cultural settings. This stands in contrast to static suites such as single-shot multiple-choice or clip-level classification benchmarks, which may measure emotional recognition or reasoning but do not directly test adaptation over dialogue time. Where this contrast is inferred rather than directly stated, the literature consistently frames multi-turn interaction, state updating, or trajectory analysis as the relevant advance (Parks et al., 29 Dec 2025).

2. Functional decomposition of emotional intelligence

DEIES decomposes evaluation into four modules. The Sensing Module detects emotional cues in text, audio, and/or video and outputs discrete labels, dimensional ratings such as valence $v\in[-1,1]$ and arousal $a\in[0,1]$, and intensity. The Explanation Module produces natural-language rationales or structured appraisals. The Response Selection Module produces candidate replies and assesses them for empathy, specificity, safety, and prosocial orientation. The Context-Adaptation Module measures performance over multiple turns, across registers, and across cultural or linguistic variants, while tracking co-regulation and rapport over time (Parks et al., 29 Dec 2025).

A closely related decomposition appears in EICAP’s psychologically grounded four-layer taxonomy for LLMs. There, emotional intelligence is partitioned into Emotional Tracking, Cause Inference, Appraisal, and Emotionally Appropriate Response Generation. The dialogue context up to turn $t$ is formalized as

$X^{(t)} = \bigl(u_1,a_1,u_2,a_2,\dots,u_t\bigr),$

and the four layers are then modeled as tracking the user’s evolving emotional state $\varepsilon^{(t)}$, inferring causes $\kappa^{(t)}$, computing appraisal $\alpha^{(t)}$, and selecting the assistant reply $a_t$ under those constraints (Nazar et al., 8 Aug 2025).

This suggests a near-isomorphism between the DEIES modules and the EICAP layers. Sensing aligns with emotional tracking; explanation aligns with cause inference and parts of appraisal; response selection aligns with emotionally appropriate response generation; and context-adaptation captures the longitudinal aspect that turns these abilities into a dynamic competence. The importance of appraisal is especially notable, because EICAP defines it as evaluation of intensity, urgency, ethical constraints, and cultural sensitivities, together with prioritization of an appropriate support strategy (Nazar et al., 8 Aug 2025).

3. Dynamic protocols and adaptive mechanisms

DEIES introduces several explicitly adaptive mechanisms. Evaluation begins with curriculum-style item sampling, starting from “easy” items and shifting toward low-agreement, complex, or mixed-emotion items as system performance improves. It also uses pluggable cultural-context packs with localized vignettes and annotation norms, and re-weights items to maintain balance across locales. After each evaluation epoch $t$, module-specific weights $v\in[-1,1]$0 are updated to emphasize remaining weaknesses, yielding what the paper describes by analogy as a dynamic EMI schedule (Parks et al., 29 Dec 2025).

Recent dynamic EI benchmarks instantiate this adaptive logic in different ways. EICAP-Bench presents each turn with the full dialogue history $v\in[-1,1]$1, so the model’s single prediction implicitly exercises all four EI layers, and subtask labels enable layer-specific analysis over multi-turn MCQ rows (Nazar et al., 8 Aug 2025). SpeechEQ uses a two-round forced-choice protocol at Turns 4 and 6, where identical text is paired with alternative paralinguistic deliveries, forcing the model to reason over evolving emotional arcs rather than semantic content alone (Wu et al., 24 Jun 2026). Multi-Bench extends this to spoken-dialogue systems through an interactive dialogue task in which a simulated user and an SDM interact for up to ten turns, with termination determined by an LLM judge when emotional relief or stalling is detected (Deng et al., 2 Nov 2025).

Simulator-based approaches push this further. EIBench has a simulator that plays the user, updates an emotion-relation state after each turn, and maps the final state to an anchor-based score, so the same infrastructure functions both as an evaluation benchmark and as a training environment with dense feedback for RL (Zhu et al., 14 Jun 2026). DESG adopts a different dynamic formalism: dialogue windows are represented as sequences of clinical states embedded in a directed graph over sliding windows, and in streaming settings the graph is updated incrementally by adding the new node $v\in[-1,1]$2 and edge $v\in[-1,1]$3, with the total score updated online (Han et al., 5 May 2026).

Taken together, these protocols show that “dynamic” in EI evaluation can mean at least four different things: longitudinal conversational context, adaptive sampling, stateful simulation, and streaming trajectory analysis. The field has not collapsed these into a single standard, but the shared assumption is that turnwise dependence is indispensable.

4. Metrics and scoring paradigms

DEIES formalizes dynamic scoring through module-wise normalized sub-scores $v\in[-1,1]$4, $v\in[-1,1]$5, $v\in[-1,1]$6, and $v\in[-1,1]$7, together with dynamic weights

$v\in[-1,1]$8

The context-sensitivity score for cultural context $v\in[-1,1]$9 is

$a\in[0,1]$0

and the overall General EI Index is

$a\in[0,1]$1

Deployment gating is handled separately by MDB pass/fail logic requiring threshold satisfaction on at least sensing and response, plus zero unsafe responses on a hold-out “critical incidents” set (Parks et al., 29 Dec 2025).

EICAP-Bench uses a more conventional benchmark geometry: per-instance log-probability scoring, overall accuracy, layer-specific accuracy, macro-average over four layers, and paired bootstrap resampling with $a\in[0,1]$2 resamples to estimate two-sided $a\in[0,1]$3 confidence intervals, with Benjamini–Hochberg FDR correction for multiple subtask comparisons (Nazar et al., 8 Aug 2025). This makes EICAP suitable for fine-grained hypothesis testing about which EI layers respond to intervention.

AttuneBench adopts an explicitly decomposed multi-metric design. It reports Emotion F1, Emotion VA partial credit, binary accuracies for observed and preferred behavior, Pairwise Accuracy and Kendall’s $a\in[0,1]$4 for response preference ordering, Four-Branch MAE, PANAS B-Adj, and a conversation-level composite

$a\in[0,1]$5

where $a\in[0,1]$6, $a\in[0,1]$7, and $a\in[0,1]$8 group emotion, behavioral, and holistic metrics, respectively (Lubrano et al., 20 May 2026). The significance of this design is methodological: it encodes the claim that emotionally intelligent behavior is not a single latent score but a bundle of separable capabilities.

SpeechEQ introduces a norm-referenced psychometric score. It first computes trajectory accuracy over the two-round spoken task and then converts raw performance into a robust $a\in[0,1]$9-score using the median and MAD, finally mapping this to a $t$0 SEQ scale capped at $t$1 (Wu et al., 24 Jun 2026). By contrast, the trajectory-based framework in “Detecting Emotional Dynamic Trajectories” defines BEL, ETV, and ECP to summarize long-horizon emotional state evolution, while EIBench combines per-turn process rewards with an anchor-based final outcome reward over a two-dimensional state $t$2 tracking negative emotion intensity and relational state (Tan et al., 12 Nov 2025, Zhu et al., 14 Jun 2026).

A plausible implication is that dynamic EI evaluation now spans four metric families: module-weighted indices, task accuracies, psychometric standardization, and state-trajectory functionals. Choice among them depends on whether the benchmark is intended for deployment gating, scientific diagnosis, leaderboard discrimination, or RL training.

5. Benchmark implementations and modalities

The following frameworks illustrate how dynamic EI evaluation has been instantiated across text, speech, simulation, and clinically structured dialogue.

Suite	Dynamic unit	Primary representation or score
DEIES (Parks et al., 29 Dec 2025)	evaluation epoch $t$3	MDB and GEI
EICAP-Bench (Nazar et al., 8 Aug 2025)	turn-by-turn MCQ with full history $t$4	layer-specific accuracy
AttuneBench (Lubrano et al., 20 May 2026)	real multi-turn conversation	multi-metric composite
SpeechEQ (Wu et al., 24 Jun 2026)	two-round forced-choice at Turns 4 and 6	SEQ
EIBench (Zhu et al., 14 Jun 2026)	simulator turn with state update $t$5	process reward and outcome reward
DESG (Han et al., 5 May 2026)	sliding-window or streaming graph update	graph-level health score

These suites differ not only in modality but in what counts as the evaluand. EICAP evaluates open-source LLMs on multilingual multi-turn dialogue with explicit cultural and linguistic coverage in English and Modern Standard Arabic (Nazar et al., 8 Aug 2025). AttuneBench is grounded in 200 genuine multi-turn human-model conversations with turn-by-turn annotations of emotional state, model behavior, preferred responses, and judged response quality (Lubrano et al., 20 May 2026). SpeechEQ evaluates Speech-LLMs on 2,265 spoken dialogues grounded in the 15 subscales of EQ-i 2.0, with the critical manipulation located in paralinguistic delivery rather than transcript content (Wu et al., 24 Jun 2026).

EIBench reframes evaluation as interactive emotion management: 2,222 scenarios are organized under a $t$6 taxonomy of Support, Defense, Repair, and Charm, and the simulator explicitly updates both emotional and relational state after each model turn (Zhu et al., 14 Jun 2026). DESG, meanwhile, is oriented toward support-dialogue audit rather than general conversational benchmarking; it embeds each turn in a decoupled clinical state manifold of semantic, affective, and cognitive-distortion features and scores directed transitions asymmetrically, so deterioration is penalized more heavily than recovery is rewarded (Han et al., 5 May 2026).

This range of implementations shows that a dynamic EI suite need not be limited to free-form chat evaluation. It may take the form of MCQ probing, forced-choice acoustic selection, user-annotated conversation replay, simulator-mediated rollouts, or directed-graph audit over support dialogue. The commonality lies in preserving temporal structure and making emotional competence depend on state transitions rather than isolated answers.

6. Empirical findings, misconceptions, and open directions

Across current evaluations, dynamic EI remains difficult. In EICAP, Qwen 2.5-7B-Instruct achieved the highest macro-average of approximately $t$7 across layers, and LoRA fine-tuning on UltraChat produced a statistically significant improvement only for Arabic Appraisal items, at $t$8 percentage points with $t$9 CI $X^{(t)} = \bigl(u_1,a_1,u_2,a_2,\dots,u_t\bigr),$0; other layer-language configurations showed no reliable gain or slight declines (Nazar et al., 8 Aug 2025). This directly challenges the assumption that general conversational instruction tuning is sufficient for robust emotional reasoning.

AttuneBench challenges a different assumption: that emotion recognition accuracy adequately captures conversational EI. Model rankings on emotion recognition, behavioral classification, preference prediction, and judged response quality are largely independent, Pairwise Accuracy is the most discriminating metric, and composite scores are tightly bunched despite large per-metric divergence (Lubrano et al., 20 May 2026). A plausible implication is that aggregate leaderboards can obscure clinically or interactionally important capability trade-offs.

SpeechEQ identifies three failure modes in spoken systems: a text-reliant “modality shortcut,” an alignment-induced “safety trap,” and “contextual amnesia” (Wu et al., 24 Jun 2026). Multi-Bench reports that current SDMs perform relatively well on basic understanding tasks but still have room for improvement on advanced multi-turn interactive dialogue and reasoning-related tasks, particularly emotion awareness and application (Deng et al., 2 Nov 2025). EIBench similarly finds that current models perform well on support and rapport-building scenes but struggle with boundary maintenance under user pressure, and that CTC-GRPO can substantially improve performance by reusing simulator state updates as turn-level feedback (Zhu et al., 14 Jun 2026).

A further controversy concerns how EI should be judged. DESG argues that direct LLM judges and symmetric text-similarity metrics are poorly aligned with therapeutic quality because the target label depends on clinical direction, and it proposes asymmetric clinical geometry over dynamic state manifolds as an alternative (Han et al., 5 May 2026). This does not invalidate LLM-based judging, but it does indicate that in mental-health or high-risk support settings, semantic agreement alone may be an insufficient surrogate for emotionally intelligent action.

The principal research directions are comparatively consistent across frameworks. DEIES argues for multimodality, multi-turn depth, prosocial focus, cultural coverage, transparent scoring, and a dual benchmark/index structure (Parks et al., 29 Dec 2025). EICAP calls for purpose-built, psychologically grounded corpora, multilingual and culturally diverse annotation, and objectives that jointly optimize emotion tracking, causal inference, appraisal, and ethical response generation (Nazar et al., 8 Aug 2025). This suggests that the long-term trajectory of dynamic EI evaluation will be toward benchmarks that are simultaneously theory-grounded, stateful, multimodal, culturally variant, and tightly coupled to training signals rather than remaining purely diagnostic.