Agent Humanization Benchmark (AHB)

• Agent Humanization Benchmark (AHB) is a framework that evaluates mobile GUI agents by measuring both task success and human-like touch dynamics to avoid detection.
• It differentiates between utility, robustness, and humanization using behavioral biometrics and formal interaction models to assess anti-detection strategies.
• The benchmark employs an adversarial game formulation and detailed feature extraction from motion events to quantify agent imitability and evaluate humanization methods.

to=arxiv_search.search 福利彩票天天彩json {"^^^^^^^^3query3 OR \3^^^^^^^^"Turing Test on Screen\"^^^^^^^^3 OR \3^^^^^^^^"Agent Humanization Benchmark\"","max_results":^^^^3(Zhu et al., 24 Feb 2026) OR \3query3^^^^,"sort_by":"submittedDate"} to=arxiv_search.search 】【。】【”】【json {"^^^^^^^^3query3 OR (&&&^^^^3(Zhu et al., 24 Feb 2026) OR \3^^^^&&&)^^^^^^^^3 OR \3^^^^^^^^"JobBench\" OR (&&&^^^^3 OR \3^^^^&&&)^^^^^^^^3 OR \3^^^^^^^^"COMMA\" (Ossowski et al., 2024)","max_results":^^^^3(Zhu et al., 24 Feb 2026) OR \3query3^^^^,"sort_by":"submittedDate"} Agent Humanization Benchmark (AHB) is a benchmark for mobile GUI agents that evaluates not just whether an agent can complete a task, but whether it can do so with human-like touch behavior that avoids detection by platform-side classifiers. Introduced in “Turing Test on Screen: A Benchmark for Mobile GUI Agent Humanization” (&&&^^^^3query3^^^^&&&), it treats humanization as a first-class requirement in human-centric, adversarial digital ecosystems. The acronym is ambiguous in current literature: “Adversarial Humanities Benchmark” uses AHB for a distinct frontier-model safety benchmark rather than mobile GUI anti-detection (Galisai et al., 20 Apr 2026).

^^^^3(Zhu et al., 24 Feb 2026) OR \3^^^^. Problem setting and scope

AHB is motivated by the claim that mobile GUI agents need not only task-solving ability, but also behavioral camouflage. The benchmark distinguishes utility / robustness from humanization / anti-detection. Utility asks whether the agent finishes the task; robustness asks whether it survives adversarial perturbations from the platform; humanization asks whether it performs the task in a way that does not expose it as non-human. The paper frames the relevant threat model as detect vs. anti-detect, arguing that a platform can deploy a detector before stronger countermeasures are needed (&&&^^^^3query3^^^^&&&).

Within this framing, “humanization” means making the agent’s observable touch behavior sufficiently similar to real users that platform-side detectors cannot reliably distinguish it. The paper’s “Turing Test on Screen” formulation shifts imitation from language to GUI behavior: given the sequence of touch and sensor events generated while operating a mobile UI, the central question is whether a detector can tell whether the operator was a human or a GUI agent. AHB therefore does not benchmark social warmth, empathy, or anthropomorphic dialogue; it benchmarks behavioral biometrics.

The benchmark focuses on mobile GUI agents, Android smartphone interactions, multiple application domains, and touch-driven interaction. It spans ^^^^3 OR \3(Zhu et al., 24 Feb 2026) OR \3^^^^ applications grouped into five clusters: Social Media (Toutiao, Weibo, Xiaohongshu, Zhihu), Shopping (JD, Taobao, Cainiao, Meituan, Eleme), Video Streaming (iQIYI, Bilibili, QQ Music), Trip Planning (Ctrip, Amap (Gaode), Umetrip, Qunar), and Office & Learning (Tencent Docs, Tencent Meeting, Youdao Dictionary, Haodafu). Its action modalities include taps, swipes, scrolls, drags, temporal gaps between actions, tap durations, and optionally sensor streams in the dataset, although the paper explicitly says it prioritizes humanization of MotionEvents and treats sensor simulation as future work (&&&^^^^3query3^^^^&&&).

^^^^3 OR \3^^^^. Formal interaction model and detectability theory

AHB formalizes GUI interaction at two levels. At the agent level, the agent chooses a high-level UI action:

PRESERVED_PLACEHOLDER_^^^^3query3^^^^

where PRESERVED_PLACEHOLDER_^^^^3(Zhu et al., 24 Feb 2026) OR \3^^^^ is the GUI agent, PRESERVED_PLACEHOLDER_^^^^3 OR \3^^^^ is the current state, and $a_t$ is a logical action such as tap or swipe. At the event level, each logical action expands into fine-grained events:

$$E_t = \{e_{t,1}, e_{t,2}, \dots, e_{t,k}\} = f(a_t),$$

with $e \in M \cup S$, where $M$ denotes motion events and $S$ denotes sensor events. The complete observed behavior is

$\mathcal{E}_{1:T} = \bigcup_{t=1}^{T} E_t.$

This decomposition matters because the detector operates on the event stream rather than on high-level action labels alone (&&&^^^^3query3^^^^&&&).

The detector–agent interaction is posed as a MinMax adversarial game. In the appendix, the detector objective is written as

$$\mathcal{L}_D(D_\Theta; P, G'_\Phi) = \mathbb{E}_{x \sim P}[\log D_\Theta(x)] + \mathbb{E}_{x \sim G'_\Phi}[\log(1 - D_\Theta(x))],$$

where PRESERVED_PLACEHOLDER_^^^^3(Zhu et al., 24 Feb 2026) OR \3query3^^^^ is the true human trajectory distribution, PRESERVED_PLACEHOLDER_^^^^3(Zhu et al., 24 Feb 2026) OR \3(Zhu et al., 24 Feb 2026) OR \3^^^^ is the raw agent distribution, PRESERVED_PLACEHOLDER_^^^^3(Zhu et al., 24 Feb 2026) OR \3 OR \3^^^^ is a humanization wrapper as a Markov kernel, and PRESERVED_PLACEHOLDER_^^^^3(Zhu et al., 24 Feb 2026) OR \33^^^^ is the humanized agent distribution. The paper’s main theoretical quantity is the Jensen–Shannon divergence between PRESERVED_PLACEHOLDER_^^^^3(Zhu et al., 24 Feb 2026) OR \3^^^^4 and PRESERVED_PLACEHOLDER_^^^^3(Zhu et al., 24 Feb 2026) OR \3^^^^5. Theorem ^^^^3(Zhu et al., 24 Feb 2026) OR \3^^^^ states:

PRESERVED_PLACEHOLDER_^^^^3(Zhu et al., 24 Feb 2026) OR \3^^^^6

The intended interpretation is explicit: detector power is controlled by PRESERVED_PLACEHOLDER_^^^^3(Zhu et al., 24 Feb 2026) OR \3^^^^7, and as the humanized-agent distribution approaches the human distribution, the best detector approaches random guessing (&&&^^^^3query3^^^^&&&).

Two further theoretical claims structure the benchmark’s design. First, smoothing strictly reduces detectability in principle: if PRESERVED_PLACEHOLDER_^^^^3(Zhu et al., 24 Feb 2026) OR \3^^^^8, then

PRESERVED_PLACEHOLDER_^^^^3(Zhu et al., 24 Feb 2026) OR \3^^^^9

Second, history matching is asymptotically superior: with enough human exemplars, the history-matched empirical distribution approaches the human one in feature space, and detector performance approaches the random-guessing lower bound. These results support the benchmark’s dual emphasis on imitability and utility, but the empirical study still reports a gap between theoretical high imitability and full indistinguishability in practice (&&&^^^^3query3^^^^&&&).

3. Dataset construction and behavioral feature space

AHB introduces a new high-fidelity mobile touch dynamics dataset containing both motion and sensor event streams. Human traces come from four sub-populations—Young Man, Young Woman, Middle-aged, and Elderly—intended to capture age- and physiology-related variation. All experiments were run on a single phone, the Xiaomi Mi Max ^^^^3 OR \3^^^^ with MIUI ^^^^3(Zhu et al., 24 Feb 2026) OR \3(Zhu et al., 24 Feb 2026) OR \3^^^^.^^^^3query3^^^^. OR \3^^^^.^^^^3query3^^^^. The overall human event count is 37,768,698; human motion-event counts sum to 8,^^^^3(Zhu et al., 24 Feb 2026) OR \3\3 OR \3^^^^ taps and 3,34^^^^3(Zhu et al., 24 Feb 2026) OR \3^^^^ swipes. Raw-agent traces total ^^^^3 OR \3^^^^43,^^^^3query3 OR (Sturgeon et al., 10 Sep 2025) OR \3query3^^^^,9^^^^3 OR \3^^^^9 events and are collected from UI-TARS, Mobile-Agent-E (GPT-4o), Mobile-Agent-E (Claude-3.5-Sonnet), AgentCPM, and AutoGLM. The paper also reports dedicated humanized-agent datasets, with ^^^^3 OR \3\3query3^^^^,3^^^^3 OR \3query3^^^^,486 events without fake actions and ^^^^3(Zhu et al., 24 Feb 2026) OR \3^^^^54,7^^^^3query3^^^^4, events with fake actions (&&&^^^^3query3^^^^&&&).

The benchmark defines each finger event as PRESERVED_PLACEHOLDER_^^^^3 OR \3query3^^^^, and classifies an action as a tap if PRESERVED_PLACEHOLDER_^^^^3 OR \3(Zhu et al., 24 Feb 2026) OR \3^^^^ and a swipe if PRESERVED_PLACEHOLDER_^^^^3 OR \3 OR \3^^^^. Sensor streams are recorded as

PRESERVED_PLACEHOLDER_^^^^3 OR \33^^^^

However, the benchmark’s detection and humanization focus is overwhelmingly on MotionEvents rather than full sensor realism (&&&^^^^3query3^^^^&&&).

Feature extraction uses a ^^^^3 OR \3^^^^4-dimensional touch-dynamics representation inspired by Touchalytics-style behavioral biometrics. The feature families are kinematic and velocity features, acceleration features, deviation and linearity features, geometric and spatial features, and coordinate/time features. The paper identifies linearly and path geometry as especially revealing for raw agents. The highest-information features include maxDev with information gain PRESERVED_PLACEHOLDER_^^^^3 OR \3^^^^4, ratio_end_to_length with PRESERVED_PLACEHOLDER_^^^^3 OR \3^^^^5, meanResultantLength with PRESERVED_PLACEHOLDER_^^^^3 OR \3^^^^6, direction with PRESERVED_PLACEHOLDER_^^^^3 OR \3^^^^7, and avgDirection with PRESERVED_PLACEHOLDER_^^^^3 OR \3^^^^8. This supports the empirical claim that raw-agent swipes are unusually rigid and straight relative to human traces (&&&^^^^3query3^^^^&&&).

4. Evaluation protocol and benchmark metrics

AHB evaluates three trace sources—human users, raw GUI agents, and humanized GUI agents—and outputs detector accuracies, task success rates, and per-feature detectability analyses. The detector hierarchy has two layers. Rule-based detectors perform simple threshold-based anomaly checks using Swipe Accuracy, Time Interval Accuracy, and Tap Duration Accuracy. Learning-based detectors use SVM and XGBoost trained on the ^^^^3 OR \3^^^^4-dimensional touch feature vector. Under the paper’s operational criterion, a perfectly humanized agent should push detector performance toward random guessing, with accuracy near PRESERVED_PLACEHOLDER_^^^^3 OR \3^^^^9; this is the benchmark’s concrete interpretation of passing the “Turing Test on Screen” (&&&^^^^3query3^^^^&&&).

Utility is measured as Task Success Rate or Task Accuracy. AHB is therefore a bi-objective benchmark: imitability improves as detector accuracy approaches $a_t$^^^^3query3^^^^, while utility improves as task completion remains high. The paper repeatedly emphasizes the imitability–utility Pareto frontier, because some anti-detection techniques reduce detectability at the cost of breaking task execution. This is also why the benchmark distinguishes online mode, where humanization is applied during real-time execution and utility is measured jointly with detectability, from offline mode, where traces are modified post hoc to isolate detectability effects (&&&^^^^3query3^^^^&&&).

The benchmark release includes code and data through a public GitHub repository and a Hugging Face dataset release. At the same time, several evaluation details remain underspecified in the paper: it does not provide explicit train/validation/test splits, does not report formal significance tests or confidence intervals for the main experimental tables, and does not specify detailed normalization procedures for SVM or XGBoost input (&&&^^^^3query3^^^^&&&).

5. Humanization strategies and empirical findings

The benchmark evaluates several “vanilla” LMM-based mobile agents—UI-TARS, Mobile-Agent-E variants, AgentCPM, and AutoGLM—and identifies three main reasons they are detectable: overly linear swipes, non-human inter-action delays caused by inference latency, and near-zero tap durations. The reported latency ranges are especially stark: UI-TARS shows intervals around 5–^^^^3(Zhu et al., 24 Feb 2026) OR \3query3^^^^ seconds, while Mobile-Agent-E shows around 5^^^^3query3^^^^–8^^^^3query3^^^^ seconds. Human tap durations, by contrast, occupy a roughly Gaussian band around ^^^^3query3^^^^.^^^^3query3 OR \3query3^^^^ seconds (&&&^^^^3query3^^^^&&&).

The paper distinguishes two humanization paradigms, Internal Injection and External Wrapper, but studies the External Wrapper approach. It evaluates four concrete strategies: Heuristic Noise Injection using B-spline smoothing,

$a_t$^^^^3(Zhu et al., 24 Feb 2026) OR \3^^^^

Data-Driven History Matching or Rotation & Match, Fake Actions, and Longer Presses sampled from a Gaussian fit to human tap data. Fake actions are implemented as circular gestures of radius 5^^^^3query3^^^^ px emitted according to a Poisson process with rate $a_t$^^^^3 OR \3^^^^ Hz (&&&^^^^3query3^^^^&&&).

Raw agents are nearly perfectly separable. In Social Media, the RAW condition yields max single $a_t$3, SVM accuracy $a_t$4, XGBoost accuracy $a_t$5, interval accuracy $a_t$6, tap accuracy $a_t$7, and task accuracy $a_t$8. In Shopping, the RAW condition yields SVM accuracy $a_t$9, XGBoost accuracy $$E_t = \{e_{t,1}, e_{t,2}, \dots, e_{t,k}\} = f(a_t),$$^^^^3query3^^^^, interval accuracy $$E_t = \{e_{t,1}, e_{t,2}, \dots, e_{t,k}\} = f(a_t),$$^^^^3(Zhu et al., 24 Feb 2026) OR \3^^^^, tap accuracy $$E_t = \{e_{t,1}, e_{t,2}, \dots, e_{t,k}\} = f(a_t),$$^^^^3 OR \3^^^^, and task accuracy $$E_t = \{e_{t,1}, e_{t,2}, \dots, e_{t,k}\} = f(a_t),$$3. The paper’s headline empirical claim is therefore that vanilla LMM agents fail the Turing Test on Screen badly (&&&^^^^3query3^^^^&&&).

Among swipe-humanization methods, History Matching consistently outperforms B-Spline smoothing. In Social Media offline swipe-only experiments, History yields max single $$E_t = \{e_{t,1}, e_{t,2}, \dots, e_{t,k}\} = f(a_t),$$4, SVM $$E_t = \{e_{t,1}, e_{t,2}, \dots, e_{t,k}\} = f(a_t),$$5, and XGBoost $$E_t = \{e_{t,1}, e_{t,2}, \dots, e_{t,k}\} = f(a_t),$$6, whereas B-spline yields max single $$E_t = \{e_{t,1}, e_{t,2}, \dots, e_{t,k}\} = f(a_t),$$7, SVM $$E_t = \{e_{t,1}, e_{t,2}, \dots, e_{t,k}\} = f(a_t),$$8, and XGBoost $$E_t = \{e_{t,1}, e_{t,2}, \dots, e_{t,k}\} = f(a_t),$$9. Long Press is an effective low-cost intervention for tap-duration realism: in Shopping, tap accuracy drops from $e \in M \cup S$^^^^3query3^^^^ in RAW to $e \in M \cup S$^^^^3(Zhu et al., 24 Feb 2026) OR \3^^^^ with Long Press. Fake Actions can drive interval-based detection close to random guessing—examples include $e \in M \cup S$^^^^3 OR \3^^^^ in Social Media, $e \in M \cup S$3 in Shopping, $e \in M \cup S$4 in Video Streaming, and $e \in M \cup S$5 in Office & Learning—but online utility can collapse. The most dramatic example is Trip Planning, where task accuracy falls from $e \in M \cup S$6 in RAW to $e \in M \cup S$7 under the online history+fake+long combination (&&&^^^^3query3^^^^&&&).

Humanization therefore helps, but full indistinguishability is not reached under the strongest learned-detector settings. In Social Media, RAW yields SVM $e \in M \cup S$8 and XGBoost $e \in M \cup S$9, while the best online history+long condition still yields SVM $M$^^^^3query3^^^^ and XGBoost $M$^^^^3(Zhu et al., 24 Feb 2026) OR \3^^^^. The easiest features to humanize are path-shape and linearity measures such as maxDev, which moves from $M$^^^^3 OR \3^^^^ in RAW to $M$3 under On.RM, and ratio_end_to_len, which moves from $M$4 to $M$5. The hardest features are endpoint precision and action interval, because both are tightly coupled to task utility and inference architecture (&&&^^^^3query3^^^^&&&).

6. Adjacent benchmark families, conceptual boundaries, and limitations

The term “humanization” is used differently across adjacent benchmark families. COMMA evaluates communicative multimodal multi-agent collaboration under asymmetric information and is explicitly framed as a benchmark for a specific slice of agent humanization rather than for social warmth or anthropomorphism (Ossowski et al., 2024). HumanAgencyBench measures six dimensions of human agency support—Ask Clarifying Questions, Avoid Value Manipulation, Correct Misinformation, Defer Important Decisions, Encourage Learning, and Maintain Social Boundaries—making it a benchmark for autonomy-preserving assistant behavior rather than for touch-dynamics camouflage (&&&^^^^3(Zhu et al., 24 Feb 2026) OR \3^^^^&&&). JobBench selects occupational tasks from duties workers identify as high-priority for delegation, thereby centering benchmark scope on human will rather than economic value alone (&&&^^^^3 OR \3^^^^&&&). AgentA/B evaluates whether LLM-based web agents can emulate human-like behavior patterns on real webpages in live browser environments (&&&^^^^3 OR \3^^^^5&&&). This suggests that current “humanization” research is multi-axis: AHB in the strict sense of (&&&^^^^3query3^^^^&&&) isolates execution-level human-likeness in mobile touch behavior, while adjacent benchmarks probe communicative collaboration, agency support, worker-prioritized delegation, or web-behavior realism.

AHB itself has explicit and implicit limits. All data are collected on a single device, so cross-device generalization is untested. Current detectors are feature-based and interpretable rather than deep sequence models. Sensor streams are collected but not seriously humanized. Some strategies, especially fake actions, sharply reduce success. The paper also states that current benchmark realism is concentrated at the execution layer and does not yet model richer “intent-layer” human traits such as indecision or curiosity. Additional omissions include the absence of explicit train/validation/test splits, the lack of a reported MotionEvent sampling frequency, and the absence of formal statistical testing in the reported tables (&&&^^^^3query3^^^^&&&).

In that sense, AHB is best understood as a benchmark for a specific operational definition of humanization: whether a mobile GUI agent can finish tasks while exhibiting sufficiently human-like touch dynamics to resist platform-side detection. Its main contribution is to recast GUI-agent evaluation from pure task success to a joint problem of imitability and utility, and to do so with both a formal detector–agent game and a large empirical study over real app interactions (&&&^^^^3query3^^^^&&&).