Synthetic Cognitive Profiles

• Synthetic cognitive profiles are structured representations of latent cognition that map internal activations to behavioral outcomes in artificial systems and human–AI ensembles.
• They facilitate analysis across multiple methodologies, including neuronal categorical segmentation, temporal modeling, and skill-based assessments with measurable metrics.
• Applications range from neural network diagnostics and language model evaluation to simulated user profiling and synthetic expertise in human–AI collaborations.

Searching arXiv for the cited works and adjacent formulations of “synthetic cognitive profiles.” Synthetic cognitive profiles are structured representations of cognition assigned to artificial systems, simulated individuals, or human–AI ensembles. In recent arXiv literature, the expression is explicit in some works and reconstructed from adjacent concepts in others: it can denote a network’s pattern of categorical segmentation and recombination, a simulated user’s latent cognitive trajectory plus multimodal emissions, a knowledge-component–level mastery profile for a LLM, or an allocation of skills and knowledge across a human/cog ensemble (Pichat et al., 2024, Drishti et al., 28 Dec 2025, Zhao et al., 13 Jan 2025, Fulbright et al., 2022). Taken together, these usages suggest a family of formalisms that turn otherwise diffuse notions of “how a system thinks” into analyzable profile objects.

1. Conceptual range and principal formulations

The literature does not present a single canonical definition. Instead, it offers several technically distinct profile objects that all connect latent cognitive structure to observable behavior, internal activations, or task performance. Some papers define the term directly, while others provide the ingredients from which the notion is reconstructed. In practice, a synthetic cognitive profile is usually a compact representation that supports comparison, prediction, diagnosis, or data generation (Pichat et al., 2024, Farooque et al., 22 May 2026, Burden et al., 2023).

Research setting	Profile object	Representative formalization
Neural-network internals	Categorical segmentation and recombination across neurons and layers	$y_j=f\!\left(\sum_i w_{i,j}x_{i,j}+a_j\right)$
Longitudinal synthetic users	Latent temporal state process plus behavioral and linguistic traces	$L_u(d)$, $P_u$, coherence, drift, entropy
LLM capability diagnosis	Knowledge-point–level mastery and coverage statistics	$\text{Acc}(kp)$, $\text{Freq}(kp)$
Human/cog ensembles	Skill allocation, knowledge stores, augmentation level	$A^+=W_{\text{cog}}/W_{\text{hum}}$

One line of work treats a profile as a latent-variable description. Bayesian Triangulation defines a cognitive profile as $\langle C,B,R\rangle$, where $C$ is a vector of capability levels, $B$ a vector of biases, and $R$ a vector of robustness parameters (Burden et al., 2023). Another line treats a profile as a temporal signature. Cogniscope, for example, models each simulated user through a latent state process $L_u(d)$0, a progression profile $L_u(d)$1, and state-conditioned linguistic and behavioral emissions (Farooque et al., 22 May 2026). A third line treats profiles as distributions over generated responses: multimodal LLMs can synthesize normative textual responses conditioned on age, gender, MMSE, and diagnosis, yielding group-specific response distributions for the Cookie Theft task (Yan et al., 25 Aug 2025).

2. Neural and representational profiles in synthetic cognition

In the neuropsychological literature on LLMs, synthetic cognitive profiles are grounded in the internal categorization performed by neurons. A neuron is treated as a synthetic category, with graded membership determined by activation, and the practical extension of that category is often approximated by the 100 tokens with the highest average activation, called core-tokens. In GPT‑2‑XL, category formation is analyzed through the aggregation function

$L_u(d)$2

which is interpreted as generating three mathematico-cognitive factors: synthetic categorical priming $L_u(d)$3, synthetic categorical attention $L_u(d)$4, and synthetic categorical phasing $L_u(d)$5 (Pichat et al., 2024).

The empirical analyses in that framework are strongly profile-oriented. For layer-1 destination neurons, summed precursor activations over the 10 strongest incoming connections correlate with destination activation ranks at $L_u(d)$6, supporting the priming effect. Attention analyses report $L_u(d)$7 between destination activation rank and cumulative weights over shared precursors, $L_u(d)$8 between connection-weight rank and the number of taken-tokens, and $L_u(d)$9 between weight magnitude and activation spread of taken-tokens. Phasing is operationalized by counting how many precursor neurons also treat a token as a core-token; the global analysis gives $P_u$0, with a logarithmic trend approaching an asymptote (Pichat et al., 2024). On this view, a network’s profile is the stable way its architecture, learned weights, and activation distributions partition token space.

A related strand studies activation geometry rather than aggregation factors. Using 12,800 neurons from layers 0 and 1 of GPT‑2XL, one paper finds “categorical convergence”: among top-100 core tokens, cosine similarity between successive activation-ranked tokens tends to rise with higher activation, with an approximately exponential trend at the high-activation end (Pichat et al., 2024). A complementary analysis shows that activation proximity does not collapse into semantic homogeneity: low-cosine successive pairs average 5.06% and 5.17% per neuron in layers 0 and 1, and tokens with almost identical activations still have mean cosine similarities of about 0.375 and 0.406, below $P_u$1 for virtually all neurons (Pichat et al., 2024). This suggests that neural profiles are often polysemous, intersectional, and only partially aligned with human-semantic similarity spaces.

3. Longitudinal simulated persons, digital phenotyping, and normative cohorts

In simulated-user and digital biomarker research, a synthetic cognitive profile is a longitudinal record generated from an explicit latent state model. Cogniscope defines each user by a progression profile $P_u$2, transition days $P_u$3, and a daily latent label $P_u$4 taking values in $P_u$5. The simulator then emits summaries, question answers, watch time, skips, pauses, replays, likes, shares, reaction time, churn, and related aggregates such as semantic drift, behavioral entropy, and engagement decay over 200 simulated days (Drishti et al., 28 Dec 2025). In the later benchmark formulation, the released simulation dataset contains $P_u$6 interaction records, plus a schema-aligned deployment dataset of 504 sessions across nine behavioral profiles; the benchmark emphasizes ERDE and time-to-detection rather than diagnosis, and explicitly states that its “Healthy,” “MCI,” and “EarlyAD” labels are simulated risk states rather than clinical adjudications (Farooque et al., 22 May 2026).

The 2025 Cogniscope paper reports that multimodal fusion is especially important for early-stage discrimination: in its ablation table, the full fusion model reaches accuracy $P_u$7, $P_u$8, and $P_u$9, whereas coherence-only and behavior-only models perform much worse on MCI (Drishti et al., 28 Dec 2025). The 2026 benchmark shows coherence values of approximately $\text{Acc}(kp)$0, $\text{Acc}(kp)$1, and $\text{Acc}(kp)$2 for simulated Healthy, MCI, and EarlyAD states, and reports that more than 95% of simulated users are detected within 10 days of MCI onset using a coherence-threshold detector with $\text{Acc}(kp)$3 (Farooque et al., 22 May 2026). In this literature, the profile is explicitly temporal, multimodal, and state-conditioned.

SynCog extends the same general idea to cross-linguistic multimodal cognitive-decline detection. Each virtual subject is parameterized by diagnosis, age, sex, education, and a five-dimensional linguistic style vector $\text{Acc}(kp)$4, where the dimensions are narrative length, syntactic complexity, spatial expressions, speech fluency, and clarity of expression, each quantized to $\text{Acc}(kp)$5. The continuous latent style variables are sampled from truncated Gaussians whose means depend on diagnosis and are modulated by age and education, then used to condition GPT‑4o transcript generation and IndexTTS2 speech synthesis (Feng et al., 8 Feb 2026). Fine-tuning a Qwen2‑Audio‑7B‑Instruct backbone with Chain-of-Thought deduction on these synthetic cohorts yields Macro-F1 scores of 80.67% on ADReSS, 78.46% on ADReSSo, and 48.71% on the independent Mandarin CIR-E cohort (Feng et al., 8 Feb 2026).

A closely related line targets synthetic normative data rather than diagnostic training. Using GPT‑4o and GPT‑4o‑mini on the Cookie Theft picture description task, advanced prompts that encode age, gender, MMSE, and diagnosis produce synthetic responses that better distinguish diagnostic groups and demographic diversity than naive prompts. In that setting, BERTScore is reported as the most reliable contextual similarity metric, with BERT F1 values typically in the 0.80–0.86 range across groups and prompting conditions, while BLEU remains much less informative for these creative clinical narratives (Yan et al., 25 Aug 2025). Another multimodal variant generates synthetic prosodic traces instead of synthetic patients: the SAD framework improves seven cognitive-state tasks by combining text with zero-shot synthetic audio, and on corpora with real audio the text+synthetic condition is competitive with text+gold audio (Soubki et al., 10 Feb 2025).

4. Capability, knowledge, and skill profiles

For LLMs, one prominent formalization is knowledge-component profiling under Cognitive Diagnosis Theory. CDS constructs a Question–Knowledge Point matrix, runs the student LLM on benchmark items, and computes

$\text{Acc}(kp)$6

These KC-level mastery and coverage statistics form a cognitive profile that then drives weakness-targeted synthetic data generation, KC-constrained augmentation, and profile-aware selection via $\text{Acc}(kp)$7. Reported gains reach up to 6.00% in code generation, 13.10% in mathematical reasoning, and 5.43% in academic exams (Zhao et al., 13 Jan 2025). Here, a synthetic cognitive profile is not a metaphor for “style”; it is a control signal for data synthesis.

Bayesian Triangulation generalizes profiling beyond LLMs. It defines the cognitive profile of a system as $\text{Acc}(kp)$8, infers these latent parameters from performance on structured task batteries, and uses measurement layouts to connect task-instance features to capabilities. The method is demonstrated on 68 AnimalAI Olympics contestants and 30 synthetic agents in O-PIAAGETS. In the synthetic-agent setting, RMSE for recovered latent parameters is as low as 0.13 for object permanence and 0.11 for flat navigation when the relevant task features are represented in the model (Burden et al., 2023). This formulation makes explicit that a profile can be a probabilistic latent vector with uncertainty, rather than a generated text or simulated trajectory.

Another recent operationalization is skill-conditioned rather than state-conditioned. ClueAegis defines a synthetic cognitive profile as a library of 12 forensic skills—Light, Shad, Phys, CS, Func, OCR, Human, Region, Animal, Freq, Pixel, and Trans—together with a skill-selection mechanism $\text{Acc}(kp)$9 and skill-specific reasoning/tool chains. On ClueAegis‑Bench, increasing the skill repertoire from 0 to 12 raises average accuracy from 87.31% to 99.01% (Cao et al., 24 May 2026). This suggests a broader profile concept in which cognition is decomposed into reusable, routable skill modules.

5. Human–AI ensembles, synthetic characters, and interactive personae

A different tradition locates the profile not inside a single model but across a composite system. “Synthetic Expertise” defines synthetic expertise as expert-level performance achieved by a human/cog ensemble and introduces six Levels of Cognitive Augmentation together with an augmentation factor $\text{Freq}(kp)$0. The corresponding profile includes the distribution of 13 fundamental skills—Perceive, Act, Recall, Understand, Apply, Analyze, Evaluate, Create, Extract, Learn, Teach, Alter, Collaborate—and the distribution of knowledge stores such as $\text{Freq}(kp)$1, $\text{Freq}(kp)$2, $\text{Freq}(kp)$3, $\text{Freq}(kp)$4, $\text{Freq}(kp)$5, $\text{Freq}(kp)$6, $\text{Freq}(kp)$7, $\text{Freq}(kp)$8, $\text{Freq}(kp)$9, and $A^+=W_{\text{cog}}/W_{\text{hum}}$0 across human and cog (Fulbright et al., 2022). In this literature, a synthetic cognitive profile is a configuration of division of labor.

In simulation and game AI, Sigma provides a more architectural realization. Synthetic characters are represented through factor graphs, predicates, conditionals, and a cognitive cycle comprising input, graph solution, decision, learning, and output. Profiles can integrate perception, SLAM-style spatial memory, reinforcement learning, Theory-of-Mind, and architectural appraisal variables such as attention, curiosity, surprise, desirability, and familiarity within one factor-graph substrate (Ustun et al., 2021). This is a profile as an executable cognitive architecture rather than a descriptive summary.

In HCI-oriented work on synthetic personae, the profile is memory-centric. One proposal argues that LLMs should be used as data augmentation systems rather than zero-shot persona generators and advocates episodic-memory frameworks with first-person voiceovers, expert commentary, valence/arousal scores, timestamps, and standardized locations. Retrieval is then based on semantic, emotional, and spatiotemporal similarity, yielding persona-consistent responses grounded in structured episodic memory (Gonzalez et al., 2024). A related usability line casts GPT‑4 and Gemini‑2.5‑pro as evaluator agents in a “Synthetic Cognitive Walkthrough.” Off-the-shelf LLM agents achieved higher task completion than humans—100% for GPT‑4 and 97.2% for Gemini, versus 88.2% for human participants—but identified fewer failure points; with additional prompting and full navigation context, the odds ratios for predicting human-identified failure points rise into the 2.22–7.50 range (Zhong et al., 3 Dec 2025). These studies treat profiles as tunable synthetic user models whose memory, exploration strategy, and confusion thresholds can be adjusted.

6. Limitations, controversies, and open problems

Several limitations recur across the literature. In the neural-category line, the statistical analyses are exploratory, only early layers are studied, and there is “no full formal metric of a synthetic cognitive profile yet” (Pichat et al., 2024). In Bayesian Triangulation, profile quality depends strongly on the measurement layout: if important task features are omitted, latent capabilities are misestimated or conflated (Burden et al., 2023). These constraints indicate that profile construction is inseparable from representational choice.

In clinical simulation, realism remains partial. Cogniscope’s simulated states are explicitly non-clinical risk states, its priors are illustrative, and its trajectories are monotonic, lacking remission or richer multi-domain symptom structure (Farooque et al., 22 May 2026). SynCog acknowledges that IndexTTS2 does not fully reproduce neuromotor deficits such as micro-tremors or articulatory breakdown, and also notes that Chain-of-Thought explanations may be clinically plausible without being faithful to the actual decision basis (Feng et al., 8 Feb 2026). Synthetic normative responses for Cookie Theft still exhibit hallucinations and require automated filtering and human adjudication before high-stakes use (Yan et al., 25 Aug 2025).

Human alignment is likewise incomplete. In synthetic cognitive walkthroughs, LLMs navigate more optimally than humans, rely on full context as a kind of perfect episodic memory, and under-detect failure points unless prompted specifically to rate confusion (Zhong et al., 3 Dec 2025). In synthetic personae and synthetic expertise, the literature presents rich representational schemes but far fewer standardized benchmarks for validating whether those profiles actually match the intended human users, historical figures, or ensemble roles (Gonzalez et al., 2024, Fulbright et al., 2022).

Future directions are already visible within the cited work: higher-layer studies of categorical abstraction in LLMs, richer and possibly hierarchical formalisms for profile description, hybrid evaluation on real cohorts, more realistic social and platform dynamics, broader task batteries, and stronger links between explicit profile objects and downstream decision-making (Pichat et al., 2024, Farooque et al., 22 May 2026, Feng et al., 8 Feb 2026). This suggests that “synthetic cognitive profile” is best understood, at present, as a methodological umbrella for controlled representations of latent cognition rather than a single settled formal object.