---
title: 'CAPE: Context-Aware Personality Evaluation'
url: https://www.emergentmind.com/topics/context-aware-personality-evaluation-cape
type: topic
---

# CAPE: Context-Aware Personality Evaluation

Context-Aware Personality Evaluation (CAPE) denotes personality assessment or inference that conditions on context rather than treating answers, behaviors, or utterances in isolation. Across the literature, “context” includes conversational history, smartphone sensing and usage traces, fully-duplex turn-taking structure, dyadic partner behavior, task demands, stress, culturally grounded scenarios, and adaptive multimodal probes; “personality” is variously operationalized as Big Five traits, MBTI dimensions and types, HEXACO dimensions, Jungian psychological types, self-reports, peer reports, perceived personality, or country-level human trait distributions [2508.20385] [1807.01515] [2505.14356] [2506.05670] [2012.14259]. CAPE therefore spans at least two related problems: inferring human personality from contextualized behavioral evidence, and evaluating or controlling personality expression in language models under realistic interaction histories [2212.10276] [2601.10025].

## 1. Conceptual foundations and personality targets

In the most explicit formulation for language models, CAPE replaces context-free item answering with a history-conditioned response function. Let the conversational history at step $t-1$ be $H_{t-1}=\{(q_1,r_1),\ldots,(q_{t-1},r_{t-1})\}$, and let the model’s response to the current item be $r_t=f(q_t,H_{t-1})$; scoring trajectories are then aggregated into trait scores $S_d=\frac{1}{N_d}\sum s_i$ over the items associated with trait $d$ [2508.20385]. This formalization makes context an explicit variable in personality evaluation rather than a nuisance to be removed.

The personality constructs used in CAPE are not uniform. Big Five traits dominate work on smartphone sensing, speech dialogs, social media text, culturally grounded evaluation, conversational speech, and dyadic interaction datasets, usually as openness, conscientiousness, extraversion, agreeableness, and neuroticism or emotional stability [1807.01515] [2505.14356] [2302.00493] [2506.05670] [2507.19137]. Other systems instead adopt MBTI as four binary dimensions and a 16-type space, sometimes combined with adaptive questioning and confidence estimation [2607.06149]. A different line models personality by Jungian psychological types and their MBTI correspondences through distributions of “psychological energy” over eight functions such as Te, Ti, Fe, Fi, Se, Si, Ne, and Ni [2601.10025]. In asynchronous video interviews, personality assessment is framed around four HEXACO dimensions—Honesty–Humility, Extraversion, Agreeableness, and Conscientiousness—using prompts that embed task descriptions and subject metadata [2507.22367].

The target variable also differs across studies. Some works seek self-reported stable traits, as in UDIVA’s use of BFI-2 self-reported OCEAN labels converted to z-scores [2012.14259]. Others infer perceived or apparent personality from external judgments, such as conversational speech in neutral versus stressful work situations [2507.19137]. PersonaTAB operationalizes “conversational personality” as trait-aligned patterns of talkativeness, turn-taking frequency, backchannel type and rate, interjection success, laughter frequency, and emotion or sentiment distributions in fully-duplex speech [2505.14356]. In LLM studies, the assessed object is neither latent human personality nor clinical diagnosis, but perceived personality as expressed in model outputs, often through psychometric proxies or role-played behavior [2212.10276] [2508.20385].

This breadth has two consequences. First, CAPE is not a single model family but a design paradigm for conditioning personality assessment on situational evidence. Second, comparisons across CAPE systems must respect whether they estimate traits, states, perceived impressions, or controlled persona expression.

## 2. Sources of context and representational units

The context sources used in CAPE are highly heterogeneous, but they can be organized by the kind of interactional evidence they preserve.

| Domain | Context source | Representative systems |
|---|---|---|
| Mobile sensing | Smartphone sensors, OS events, app and core-function metadata | TYDR [1807.01515] |
| Text and social media | Evaluative targets, stance, sentiment intensity, author-level topic prevalences | Evaluative profiling [2302.00493] |
| Conversational speech | Timestamps, overlaps, backchannels, laughter, emotion, sentiment, sample turns | PersonaTAB [2505.14356] |
| Work-situation speech | Neutral versus stressful interaction context, acoustic and non-verbal behavior | APP across situations [2507.19137] |
| Dyadic multimodality | Partner signals, task type, workload, relationship, mood, fatigue, metadata | UDIVA [2012.14259] |
| Cultural evaluation | Country-specific norms and scenario grounding | CulturalPersonas [2506.05670] |
| Adaptive assessment | Dialogue history, low-confidence dimensions, photo-based probes | BlossomPsy [2607.06149] |
| LLM psychometrics | Prior question–answer history and prompt-sensitivity factors | CAPE for LLMs [2508.20385] |

In smartphone-based CAPE, a general context data model comprises four categories aligned to how users interact with phones: physical conditions and activity, device status and usage, core functions usage, and app usage. Representative signals include location, weather, ambient light sensor, ambient noise level, accelerometer, inferred activity, step counts, phone un-/lock events, headphone un-/plug events, battery level and charging status, Wi-Fi and Bluetooth connectivity, calls metadata, music playback metadata, photos metadata, notifications metadata, app usage, and app traffic [1807.01515]. The use of metadata rather than content is central to this formulation.

In text-based CAPE from social media, the core representational unit is the evaluative topic: a topic induced from evaluative text so that it captures the target domain, the attitude or stance toward that target, and the linguistic resources of evaluation. Author-level evaluative profiling then aggregates topic prevalences, optionally weighted by sentiment intensity, to represent “what you evaluate and how strongly” [2302.00493].

In fully-duplex speech dialogs, context is temporal and interactional. PersonaTAB derives conversation-level features from word-level timestamps, overlapping speech, laughter tokens, backchannel classes, emotional and sentiment distributions, and sample turns. It explicitly distinguishes “partial overlaps” from “fully overlaps,” treats fully overlapped regions as candidate backchannels, and classifies backchannels contextually with access to both past and future surrounding utterances [2505.14356]. This makes simultaneity itself a context signal rather than annotation noise.

In situational speech CAPE, the key context variable is the interaction setting. The contrast between a neutral interview and a stressful client interaction alters both perceived trait distributions and the acoustic or non-verbal cues associated with them. Loudness, sound level, and spectral flux predict extraversion, agreeableness, conscientiousness, and openness in neutral interactions, while neuroticism correlates with these features in stressful contexts [2507.19137].

Dyadic CAPE expands context further by incorporating interlocutor signals and session metadata. UDIVA models target-person local context, partner extended context, audio, age, gender, cultural background, session index, pre-session mood, pre-session fatigue, task order, task difficulty, and relationship status in collaborative and competitive tasks [2012.14259]. CulturalPersonas, by contrast, grounds context in country-specific norms and everyday scenarios validated by native annotators, such as punctuality and group harmony in Japan or Ubuntu in South Africa [2506.05670].

Taken together, these systems treat context as structured side information that changes the meaning of behavioral evidence. A plausible implication is that CAPE is best understood as a problem of context-conditioned trait readout, not merely feature accumulation.

## 3. Data acquisition, preprocessing, and aggregation

CAPE systems differ sharply in how they collect and structure evidence before any prediction stage. TYDR is an Android app designed to collect smartphone data for personality prediction while embedding privacy protections. At first launch, users must read and accept a terms and privacy policy; the app then starts five main processes without requiring a login: a main menu and tile-based dashboard, a permanent notification, an optional Personality Dynamics Diary module, a data collection engine that anonymizes on-device before local storage and secure upload, and a periodic upload process every 24 hours after backend registration [1807.01515]. The same paper states that TYDR “utilizes standardized personality questionnaires,” adopts the Big Five framing for traits, and uses the Personality Dynamics Diary for daily personality states, but does not specify the exact trait instrument, sample size, demographics, collection duration, or predictive results [1807.01515].

PersonaTAB begins from raw two-channel phone conversations in the Fisher English Training Speech corpus and applies a six-step preprocessing pipeline: speech transcription with word-level timestamps using Whisper Turbo; laughter event annotation; response boundary construction using a 700 ms inter-word gap threshold; overlap identification; backchannel classification via GPT-4o; and utterance-level emotion and sentiment classification [2505.14356]. Conversation-level descriptors are then derived by aggregating emotion and sentiment percentages per speaker and time-normalized behavior rates such as laughs per minute, backchannels per minute during the partner’s speech, and interjections per 12-minute conversation [2505.14356].

Social-media CAPE requires a different filtering pipeline. On Reddit comments from PANDORA, evaluative text is isolated by combining VADER sentiment, opinion and stance lexicons, and rule-based evaluative patterns. The initial strict filter yields 29k sentences, after which quasi-snowballing with Sentence-BERT all-mpnet-base-v2 expands the set to 310k evaluative sentences using $t_{\text{sim}}=0.70$ and $\gamma=1.05$ [2302.00493]. Topic induction is then performed with LDA, BTM, ABAE, and CTM, with CTM selected on the Pareto front of NPMI and IRBO and the number of topics set to $K=20$ [2302.00493]. Author profiles are aggregated as unweighted topic means or sentiment-intensity–weighted topic means.

CulturalPersonas uses a retrieval-augmented generation pipeline rather than passive behavioral logging. For each of six countries, 30 highly cited academic documents are retrieved; GPT-4o extracts approximately 10 salient norms, which are then validated by native annotators. Scenarios and Likert-style answer options are generated to elicit each OCEAN trait through plausible behaviors rather than abstract self-statements [2506.05670]. All items are written in English in order to isolate cultural effects from language variability [2506.05670].

UDIVA’s preprocessing is organized around synchronized multimodal recording. Videos are divided into non-overlapping 32-frame chunks with stride 2, corresponding to approximately 2.5-second windows; visual features are extracted from an R(2+1)D backbone, audio is represented by VGGish embeddings, and structured metadata are broadcast-concatenated into the attention mechanism [2012.14259]. In “Traits Run Deep,” audio is extracted at 16 kHz, transcribed with Whisper-small, visual faces are detected and cropped with Arc2Face, and frozen encoders—Emotion2Vec, SFR-Embedding-Mistral, and SigLIP2—supply modality-specific representations for later fusion [2507.22367].

These preprocessing choices are not interchangeable. Smartphone CAPE emphasizes passively collected metadata and privacy gating, speech CAPE emphasizes temporal segmentation and overlap modeling, social-media CAPE emphasizes stance-aware filtering and topic induction, and culturally grounded CAPE emphasizes prompt construction from curated external knowledge. This suggests that the representational bottleneck in CAPE is domain-specific: the decisive design choice is often how contextual evidence is preserved before any learner is applied.

## 4. Modeling strategies and context integration

The CAPE literature includes blueprint-style systems, prompt-based inference, adaptive confidence-driven assessment, multimodal fusion architectures, and structured control mechanisms.

TYDR is primarily a design and implementation paper. It does not report machine learning models, training schemes, hyperparameters, or evaluation metrics; model-building is positioned as future work, including static prediction for traits, daily prediction for personality states, and investigation of which permissions and context categories are most predictive [1807.01515]. Its contribution to CAPE is therefore infrastructural: a data model, a privacy model, and an end-to-end collection pipeline rather than a fitted predictor.

PersonaTAB implements late fusion by prompting GPT-4o with four categories of speaker attributes—Emotion, Sentiment, Basics, and Samples—derived from annotated dialogs. Each Big Five trait is classified into one of five alignment labels, mapped to scores $\{100,50,0,-50,-100\}$, and the final score is averaged over five independent LLM responses:
$$
\hat{y}_{j}=\frac{1}{R}\sum_{r=1}^{R}s_{j}^{(r)}, \quad R=5.
$$
Raw behavior statistics are converted into interpretable categories such as “Normal,” “Many/Few,” and “Very Many/Few” using the mean and interquartile range across speakers, specifically with thresholds based on $0.8\times \mathrm{IQR}$ and $1.2\times \mathrm{IQR}$ [2505.14356]. The model therefore fuses multimodal evidence through prompt engineering rather than learned feature concatenation.

BlossomPsy models CAPE as an adaptive MBTI assessment. A RoBERTa-base encoder feeds four binary MLP heads and one 16-class head, while decision-making is delegated to a modified upper-confidence-bound mechanism. The text logits are passed through a PID-tuned enhancement
$$
f(p_i)=\frac{1}{1+\alpha e^{-\beta p_i}},
$$
and confidence is measured by the overlap between per-dimension mUCB and mLCB intervals; if the overlap exceeds a threshold of 0.6 when checked every 5 iterations, the system triggers a photo-based question for that dimension [2607.06149]. Here CAPE is explicitly sequential and interventionist: the next question depends on the current uncertainty state.

“Traits Run Deep” addresses asynchronous multimodality in video interviews by making text the semantic anchor. Psychology-informed prompts combine a personality task description, ASR text, and subject meta information; the resulting instruction-following text embedding is reduced by a Chunk-Wise Projector, aligned with audio and visual streams through a Cross-Modal Connector using multi-head cross-attention, refined by a Text Feature Enhancer with gates and residual connections, and scored by an ensemble of 32 MLP regressors [2507.22367]. The explicit rationale is that text carries trait-relevant semantics while audio-visual cues are asynchronous and should modulate rather than replace the semantic backbone.

UDIVA likewise uses a transformer, but its cross-attention is organized around local and extended context streams. Query embeddings are generated from the target face and local metadata, while local and partner audiovisual streams act as keys and values in separate transformer units. The output is a continuous regression of self-reported OCEAN traits from chunked multimodal evidence [2012.14259]. By construction, partner behavior is a first-class context variable.

In LLM-agent control, JPAF represents personality as BaseWeight and TemporaryWeight distributions over eight Jungian types and updates them through three mechanisms: dominant–auxiliary coordination for coherent core expression, reinforcement–compensation for short-term adaptation, and reflection for long-term evolution [2601.10025]. By contrast, the CAPE framework for LLM psychometrics keeps the model architecture fixed and changes only the evaluation protocol: prior question–answer history is retained, sensitivity factors are varied, and trajectory-level consistency is measured across runs [2508.20385].

These systems imply different answers to the same question: where should context enter the personality model? In some systems it enters as sensor-side metadata, in others as scenario grounding, in others as attention keys and values, and in still others as accumulated dialogue history or adaptive uncertainty state.

## 5. Evaluation regimes and reported findings

The empirical literature on CAPE is uneven: some papers provide only design guidance, while others report extensive quantitative evaluations.

| System | Evaluation setup | Key reported result |
|---|---|---|
| TYDR | No predictive model reported | Model-building and evaluation are future work [1807.01515] |
| PersonaTAB | Trend alignment, correlation with human labels, cosine similarity | Average TrendScore 0.186; average correlation 0.183; cosine similarity 0.503 [2505.14356] |
| BlossomPsy | MBTI-M consistency, per-dimension Acc/F1/$\kappa$ | Average $\kappa \approx 0.66$; strongest on T/F, weakest on S/N [2607.06149] |
| CulturalPersonas | Wasserstein distance, KS, TTR | Over a 20% reduction in Wasserstein distance across models and countries in MCS [2506.05670] |
| Traits Run Deep | MSE on AVI Challenge 2025 | Validation average best MSE 0.1003 vs baseline 0.1796; ranked first on test set [2507.22367] |
| APP across situations | $R^2$, Pearson $r$, KS | Up to $\approx 37\%$ variance explained for extraversion in interview; stressful interactions more predictive of neuroticism [2507.19137] |
| JPAF | MBTI dimension accuracy, DAG, DAR, TAA, PSA | GPT and Qwen achieved 100% PSA overall; JP DAG on MBTI-70 increased by +9.75% and +13.38% [2601.10025] |
| CAPE for LLMs | TAR, ED, TC, OC | Context-dependent evaluation generally increased consistency but also induced personality shifts [2508.20385] |

PersonaTAB provides one of the clearest demonstrations that multimodal conversational context changes personality inference quality. Against text-only baselines such as BERT, MiniLM, and an LM baseline, its context-aware prompt achieves stronger alignment with human-derived trend tables and higher correlation with listening-test annotations; the average TrendScore is 0.186 versus 0.070 for MiniLM, 0.037 for LM, and $-0.011$ for BERT, while average correlation with human labels is 0.183 versus 0.096 for BERT [2505.14356]. The ablation study further shows that “Samples only” gives better correlation than “Basics only,” whereas “Emotion+Sentiment only” yields the highest TrendScore but poor cosine similarity, indicating that different context channels support different notions of agreement [2505.14356].

BlossomPsy reports both consistency relative to MBTI-M and user-experience outcomes. Across 45 participants, Match 4 occurred in 33.33% of cases and Match 3 in 42.86%; per-dimension results were E/I Acc 0.76, S/N Acc 0.67, T/F Acc 0.95, and J/P Acc 0.76, with corresponding Cohen’s $\kappa$ values 0.63, 0.50, 0.90, and 0.62 [2607.06149]. The paper also reports higher interactivity, enjoyment, satisfaction, and immersion than MBTI-M, while keeping clarity maintained [2607.06149]. Because the system is confidence-driven, these results are tied not just to classification accuracy but to the policy that routes low-confidence cases to photo-based probes.

CulturalPersonas evaluates CAPE at the distributional rather than individual-item level. Trait alignment is defined as closeness between model-generated and country-specific human trait distributions, measured primarily by Earth Mover’s or Wasserstein distance and the Kolmogorov–Smirnov statistic. The benchmark reports over a 20% reduction in Wasserstein distance across models and countries in multiple-choice selection, and ablations show that removing country and norm priming significantly degrades alignment; for Brazil, the average Wasserstein distance rises by 31.2% without cultural priming [2506.05670]. The same benchmark also reports higher lexical diversity in open-ended generations under culturally grounded prompting [2506.05670].

In asynchronous video interviews, “Traits Run Deep” reports a validation average best MSE of 0.1003 across four traits compared with a baseline of 0.1796, corresponding to approximately a 44–45% reduction, and an official test-set MSE of 0.12284 with first place in the AVI Challenge 2025 Personality Assessment track [2507.22367]. The ablations show that naive concatenation is poor, that CMC-only and TFE-only improve on it, and that CMC plus TFE is best; replacing the 32-head ensemble regressor with a single regressor increases the standard deviation of errors from 0.0031 to 0.0096 [2507.22367].

Situational APP with conversational speech provides a complementary result: context changes not just performance but the feature–trait map itself. In neutral interviews, energy-related and spectral features correlate with extraversion, agreeableness, conscientiousness, and openness, while in stressful client interactions the same features correlate with neuroticism; cross-scenario generalization is near zero or negative, but within-scenario cross-session generalization improves, with interview-session correlations up to approximately 0.57 for agreeableness [2507.19137]. This is direct evidence that CAPE cannot be reduced to a single global regressor without loss of validity.

For LLM evaluation, the CAPE framework introduces TAR, ED, and two new consistency metrics, TC and OC, based on Gaussian-process support overlap across multiple scoring trajectories. Statistical validation shows TC and OC are strongly correlated with TAR and negatively correlated with ED; Cronbach’s $\alpha$ is 0.91 for TC and 0.86 for OC, and construct-validity tests show significant differences between context-free and context-dependent settings, with $p=0.0006$ for TC and $p=0.0084$ for OC [2508.20385]. At the same time, context-dependent evaluation can induce large personality shifts. GPT-3.5-Turbo and GPT-4-Turbo exhibit extreme deviations under context, whereas Gemini-1.5-Flash and Llama-3.1-8B depend heavily on prior interactions, including adversarial history injections [2508.20385].

The overall empirical picture is therefore asymmetric. Some CAPE systems demonstrate clear improvements in alignment, consistency, or predictive accuracy when context is modeled explicitly; others establish the sensing, privacy, or evaluation infrastructure but leave predictive validation for future work.

## 6. Privacy, ethics, limitations, and open directions

Privacy is a first-order concern in CAPE whenever context is derived from personal behavioral traces. PM-MoDaC, developed for TYDR, specifies nine privacy measures: user consent, letting users view their own data, an opt-out option, approval by an ethics commission or review board, random identifiers, data anonymization, use of the permission system, secured transfer, and identifying individual users without linking to their collected data [1807.01515]. TYDR implements these measures through on-device anonymization, metadata rather than content, secure uploads, random unique identifiers from Google Play Services, UI transparency, an in-app opt-out, and separation of contact data from sensor data on the backend [1807.01515]. The same paper explicitly warns against misuse such as intrusive profiling.

Ethical issues also arise when context is cultural, conversational, or adaptive. CulturalPersonas states that the benchmark is for population-level evaluation, not individual stereotyping, and emphasizes human validation by native annotators together with mapping of norms to Hofstede dimensions [2506.05670]. BlossomPsy acknowledges that MBTI is psychometrically debated and presents the system as a research prototype rather than a diagnostic tool; it also notes that future deployments should enforce explicit informed consent, privacy-by-design storage, and strict handling of conversational data and visual preferences [2607.06149]. APP from conversational speech stresses that outputs are perceived personality, not self-reported traits, and advises against deployment in high-stakes decisions such as hiring without rigorous validation [2507.19137]. JPAF calls for examining the ethical implications of deploying evolving personalities, including manipulation risks, user trust, bias amplification, and transparency in personality change logs [2601.10025].

Methodological limitations are equally recurrent. TYDR does not report sample sizes, demographics, trait instrument details, model architectures, or quantitative prediction results [1807.01515]. PersonaTAB does not report ASR WER, inter-rater reliability statistics, confidence intervals, or significance tests, and it models laughter as its only acoustic cue [2505.14356]. BlossomPsy uses a small human sample, lacks feature-level multimodal fusion, and does not report formal calibration metrics such as ECE or entropy [2607.06149]. Evaluative-topic profiling is exploratory and correlational, with a small facet-labeled subset of 127 users and limited control for confounds beyond gender [2302.00493]. UDIVA notes the need for longer temporal modeling, and “Traits Run Deep” identifies dependence on transcript quality and lack of explicit temporal sequence modeling as limitations [2012.14259] [2507.22367]. CAPE for LLMs adds that context stabilization does not guarantee logical consistency [2508.20385].

The future directions proposed across the literature are comparatively consistent. They include developing a reusable library for unobtrusive smartphone-based personality prediction and determining which permissions and context categories are most predictive [1807.01515]; adding prosodic, temporal, and richer social-signal features such as pitch, energy, speaking rate, entrainment, politeness strategies, and repair structures in speech-based CAPE [2505.14356]; adapting adaptive assessment from MBTI to Big Five with continuous trait regression and richer context modeling [2607.06149]; extending cultural benchmarks beyond nation-state categories and adding multilingual evaluation and measurement-invariance tests [2506.05670]; incorporating time-aware transformers, adaptive prompt tuning, and uncertainty estimation in multimodal interview assessment [2507.22367]; expanding dyadic modeling to longer-range interaction dynamics and physiological signals [2012.14259]; and hardening LLM evaluation against prompt sensitivity, adversarial history, and logical inconsistency [2508.20385].

A plausible implication is that CAPE is converging on a common research agenda despite domain differences: explicit context modeling, uncertainty-aware prediction, stronger psychometric validation, fairness and privacy safeguards, and evaluation protocols that distinguish stable trait expression from context-induced shifts.

Source: https://www.emergentmind.com/topics/context-aware-personality-evaluation-cape