ChildAgentEval: Cognitive Age Alignment

• ChildAgentEval is a psychometrically grounded benchmark that evaluates if multimodal language model agents exhibit developmental cognitive profiles aligned with age-specific human stages.
• It employs 10 interactive subtests adapted from clinical tests, using browser-based tasks in vision-only and DOM-assisted modes to isolate cognitive behavior.
• Skill-guided distillation and age-normed scoring reveal that traditional age prompts fall short, while constrained high-capability models show improved developmental alignment.

Searching arXiv for ChildAgentEval and closely related agent-evaluation work to ground the article in current papers. ChildAgentEval is a psychometrically grounded benchmark for evaluating whether multimodal LLM agents can align their cognitive behavior with a target child developmental age, rather than merely solving tasks well. It is introduced as the first psychometrically grounded interactive benchmark for evaluating cognitive age alignment in MLLM-based agents, and it systematically compares agent performance against age-specific human developmental stages. Its central premise is that developmental appropriateness can be more important than raw correctness in child-facing settings such as tutoring, safety, and interactive support: a system may be technically correct yet cognitively mismatched, using abstractions, memory retention, or reasoning strategies that exceed a child’s developmental stage (Shen et al., 18 May 2026).

1. Conceptual basis and evaluative target

ChildAgentEval is motivated by the observation that state-of-the-art agentic AI systems frequently fail at foundational tasks that a child can resolve with ease, even when those systems integrate powerful tools and advanced MLLMs. The benchmark therefore does not ask only whether an agent can produce correct answers. Instead, it asks whether the agent can exhibit developmentally structured differences in reasoning, language, and memory across target ages from childhood through adolescence (Shen et al., 18 May 2026).

The benchmark is explicitly inspired by the Wechsler Intelligence Scale for Children (WISC-IV), but it does not copy protected clinical items. Rather, it adopts WISC’s psychometric structure and reimplements it as browser-based interactive tasks for AI agents. The design is also mapped to the CHC model, so evaluation is decomposed into distinct cognitive factors rather than collapsed into a single undifferentiated intelligence measure. In the paper’s framework, the primary factors are Gc, Gf/Gv, WM, and PSI (Shen et al., 18 May 2026).

This framing makes ChildAgentEval a benchmark of cognitive age alignment. In that formulation, a younger target age should not simply induce lower scores. Lower ages are expected to show shorter memory, simpler vocabulary, shallower reasoning, and more child-like error patterns, whereas higher ages should progressively show stronger abstraction, richer language, and more mature strategies. The benchmark therefore treats age alignment as a structured behavioral profile rather than as stylistic role-play (Shen et al., 18 May 2026).

2. Interactive psychometric architecture

ChildAgentEval contains 10 interactive subtests adapted to a web or browser environment in which an agent must click, select, type, and sometimes navigate between pages. The tasks preserve the construct of the corresponding clinical tests while making them executable for interactive agents.

Cognitive factor	Subtests
Gc	Test 2: Similarities; Test 6: Vocabulary; Test 9: Comprehension
Gf/Gv	Test 1: Block Design; Test 4: Picture Concepts; Test 8: Matrix Reasoning
WM	Test 3: Digit Span; Test 7: Letter-Number Sequencing
PSI	Test 5: Coding; Test 10: Symbol Search

Several subtests are engineered to isolate cognitive behavior from interface artifacts. Block Design uses numbered DOM labels for click-based spatial assembly so that errors reflect reasoning rather than poor visual localization. Coding uses a dynamic symbol table with strict time limits. Digit Span and Letter-Number Sequencing require exact sequence retention and manipulation across turns or pages. Open-ended verbal tasks such as Similarities, Vocabulary, and Comprehension require generated answers that are then scored by rubric (Shen et al., 18 May 2026).

The environment is implemented as a finite-state machine driven through Playwright, and each subtest follows a standardized administration protocol. Clinical-style control rules are preserved, including a Reversal rule, under which an agent that fails the first two questions at a higher starting age is moved to easier items, and a Discontinuation rule, under which the subtest stops after a defined number of consecutive zero scores. The benchmark is run in two execution modes: Vision-only, in which the agent sees screenshots and acts directly, and DOM-assisted, in which the agent sees screenshots plus a sanitized accessibility tree with visible elements, roles, and bounding boxes but no hidden answers or backend data. The main experiments use DOM-assisted mode to reduce confounds from pixel-level localization and better isolate cognitive behavior (Shen et al., 18 May 2026).

3. Developmental staging, scoring, and normative comparison

ChildAgentEval is age-normed over developmental stages spanning ages 6–17, with experiments centered on anchor ages 7, 10, 13, and 16. In the skill-guided setting, those ages are grouped into four developmental bands: 6–8, 9–11, 12–14, and 15–17. The benchmark expects trajectory-level age ordering: if an agent is properly aligned, performance should increase systematically with target age rather than fluctuate arbitrarily (Shen et al., 18 May 2026).

The scoring system combines raw benchmark scoring with WISC-style normative conversion. Objective subtests are scored deterministically. The paper specifies binary scoring for selection or exact-match tasks such as Picture Concepts, Matrix Reasoning, Block Design, Symbol Search, and early vocabulary items. Coding is scored as the number of correct operations completed within the time constraint. Open-ended verbal reasoning tasks use a 0/1/2 rubric, where 2 denotes a complete, abstractly appropriate answer, 1 denotes a partially correct or overly concrete answer, and 0 denotes an incorrect, irrelevant, or missing answer. GPT-5.4 is used only as a pre-annotation assistant, and all open-ended scores are independently verified by human raters (Shen et al., 18 May 2026).

After raw scoring, each subtest score is mapped through age-based normative tables into scaled scores, then combined into composite scores for the cognitive factors and into FSIQ. The paper converts composite scores into age-normed deviations by

$z=(S-100)/15$

where $S$ is the WISC-style composite score. This yields the reported FSIQ z, Gc z, WM z, Gf/Gv z, and PSI z values, which quantify deviation from typical human development at the same age (Shen et al., 18 May 2026).

The benchmark also reports several trajectory and language-sensitive statistics: normalized total score, normalized subtest scores $T1$–$T10$, trajectory monotonicity, Spearman correlation between target age and total score, Age-16 minus age-7 gap, regression slope across the four anchor ages, and language fidelity or complexity measures such as mean utterance length, lexical diversity, and causal and definitional constructions. The auxiliary language composite Lang. is defined from seven features—mean length of utterance, MATTR, abstractness, category rate, causal rate, definition rate, and average number of reasons—by the printed formula

$$Lang.\!=\!\frac{1}{7} \sum_{k=1}^{7} \frac{x_k}{\max_{\mathcal{M}(x_k)}}$$

with the paper noting, as printed, a truncated expression whose intended meaning is max-normalization over model conditions followed by averaging (Shen et al., 18 May 2026).

4. Skill-guided age distillation

A major contribution of ChildAgentEval is a skill-guided distillation method intended to induce age alignment more directly than superficial role-play prompting. The method begins from real child and adolescent data rather than from prompt labels alone. Its pipeline has two stages: statistical feature extraction from age-stratified corpora and developmental data, followed by teacher-LLM distillation into structured “cognitive skill cards” (Shen et al., 18 May 2026).

The age profiles are built from corpora spanning ages 6–17, including CHILDES, the Ohio Child Speech Corpus, Frog Story narratives, LCCPW, and ClassBank. From these sources, the authors derive a cognitive profile vector with six dimensions: Gc, Gf, Gv, WM, PSI, and Social perspective. These are then operationalized through five filter modules that are injected into the prompt, memory, and reasoning layers: a vocabulary abstraction filter, working memory mask, reasoning budget controller, visual reliance module, and social perspective filter (Shen et al., 18 May 2026).

The stated objective is to constrain not only what the model says, but also how much it can retain, how deeply it can reason, and how much it relies on visual cues. This distinguishes the skill-guided setting from ordinary age prompting. A plausible implication is that the benchmark is evaluating whether developmental constraints can be made operational at the level of interaction traces and intermediate behavior, rather than merely at the level of style transfer or persona simulation (Shen et al., 18 May 2026).

5. Experimental results and observed failure modes

The experiments compare Baseline prompting with age labels against Skill-guided prompting plus distilled age-specific cognitive constraints. The evaluated models include proprietary systems—GPT-5.4, Gemini-3.1-Pro, Gemini-3.1-Flash-Lite, and Qwen-3.6-Plus—and open-weight systems—Qwen3.5-27B and Gemma-4-31B-It. All runs use greedy decoding with temperature $0.0$ (Shen et al., 18 May 2026).

The principal empirical result is that standard age prompting does not reliably align cognition. Across strong proprietary models, baseline performance is often flat or non-monotonic across age. The paper gives GPT-5.4 as an example: its total score goes from 0.53 at age 7 down to 0.46 at age 13 and back up to 0.52 at age 16. Under skill-guided constraints, the stronger proprietary models show much better age differentiation, and the paper reports monotonically increasing total scores with target age for all evaluated proprietary models. For GPT-5.4 specifically, the total score rises from 0.41 in the 6–8 band to 0.50 in the 15–17 band, and the trajectory metric reaches Spearman $\rho = 1.00$ (Shen et al., 18 May 2026).

The improvement is not uniform across domains. The paper states that Gc improves most clearly under skill guidance, while WM is often saturated or difficult to calibrate, PSI remains weak and unstable, and Gf/Gv shows only limited improvement and often stays flat. The interpretation offered is architectural: vocabulary and explicit reasoning can be constrained by prompts and filters, whereas human-like memory decay, attentional limits, and speed constraints are much harder to simulate (Shen et al., 18 May 2026).

Model capability also conditions the effectiveness of the intervention. The skill-guided method works best on high-capability proprietary models. Open-weight models such as Qwen3.5-27B and Gemma-4-31B-It show much weaker age differentiation, and in some cases the added constraints cause task failure or compressed performance rather than better alignment. The paper therefore suggests a capability threshold below which developmental constraints degrade behavior rather than shape it (Shen et al., 18 May 2026).

The reported failure modes are correspondingly specific: prompt age labels do not change deep behavior; working memory remains too strong or structurally uncoupled from child-like limits; processing speed is confounded by system latency and inference overhead; visual-spatial reasoning is especially hard to calibrate; and open-weight models often lack the instruction-following capacity needed for skill guidance. The authors also stress that developmental alignment is not equivalent to making the model “sound younger.” Even improved child-like behavior may still be a surface imitation over adult-like internal structure (Shen et al., 18 May 2026).

6. Place within the broader agent-evaluation literature

ChildAgentEval belongs to a broader shift in agent evaluation away from single end-task success and toward richer behavioral assessment. In AgentEval, utility verification is motivated by the claim that end-to-end success is too crude, because utility may include clarity, completeness, efficiency, reasoning quality, instruction-following, and responsiveness to feedback; the framework therefore constructs application-specific multi-criterion utility profiles rather than a single scalar score (Arabzadeh et al., 2024). In An Empirical Study of Automating Agent Evaluation, the same general movement appears in a different form: final answers alone are described as inadequate because agent evaluation must inspect traces containing thoughts, tool calls, retries, state transitions, and recovery behavior, and the paper introduces an end-to-end pipeline for generating plans, metric code, and reports from real traces (Zhou et al., 12 May 2026).

Other recent work extends evaluation along complementary axes. JudgeAgent treats evaluation as an interviewer-style dynamic process with Benchmark Grading, Interactive Extension, and Evaluation Feedback, using target-adaptive difficulty adjustment and post hoc feedback validation to probe capability boundaries (Shi et al., 2 Sep 2025). The black-box AgentEval for conversational systems treats agents as stateful software with hidden workflow boundaries, mining a conversational workflow graph and replaying paths to those boundaries before applying perturbations (Lin et al., 8 Jul 2026). A2Eval addresses benchmark curation itself, arguing that static embodied VLM evaluation is redundant and imbalanced and replacing it with a two-agent workflow in which a Data Agent induces capability dimensions and an Eval Agent synthesizes and validates executable evaluation pipelines (Zhang et al., 2 Feb 2026).

Taken together, these papers suggest that agent evaluation is fragmenting into distinct but related objectives: utility verification, executable trace-based evaluation, adaptive probing, boundary testing, automated benchmark construction, and developmental calibration. ChildAgentEval is distinctive within that landscape because the evaluand is not primarily task success, safety auditability, or workflow correctness, but whether an agent can behave at the right developmental level for the user. In that sense, it extends the field’s movement from “can the agent solve the task?” to “can the agent solve—or fail to solve—the task in a cognitively age-appropriate way?” (Shen et al., 18 May 2026).

7. Limitations, interpretive scope, and implications

The paper is explicit that ChildAgentEval measures operational simulations, not biological cognition. Processing speed is only approximated because it is entangled with system latency and browser execution. Working memory is approximated externally rather than through architectural changes. The current setup is browser-based and therefore does not cover voice interaction or long-horizon tutoring. The reported results are mostly descriptive, and the paper does not provide inferential statistics such as confidence intervals or significance tests. Normative comparison uses age-normed scoring rather than a newly recruited child control sample, following psychometric convention (Shen et al., 18 May 2026).

These limitations define the benchmark’s proper scope. It does not claim that an MLLM agent possesses child cognition in a literal sense. Rather, it provides an evaluation infrastructure for whether an agent can simulate age-calibrated behavior under interactive conditions grounded in WISC/CHC-style constructs. This suggests a stringent distinction between correctness and appropriateness: an agent may be high-performing in the ordinary benchmark sense while still being developmentally misaligned (Shen et al., 18 May 2026).

The broader implication is that child-facing AI may require explicit constraints on perception, memory, reasoning, and communication, and possibly architectural changes or age-specific post-training, rather than prompt-only personas. ChildAgentEval therefore functions both as a benchmark and as a research agenda: it exposes where current systems can and cannot simulate age-specific cognitive behavior, and it frames developmental alignment as a first-class target for interactive agent evaluation (Shen et al., 18 May 2026).