Hide and Seek Game (HSG) Models

• Hide and Seek Game (HSG) is a family of adversarial models where one agent conceals targets and another strategically searches under explicit resource and imperfect detection conditions.
• HSG formulations range from classical box-search games with sequential inspections to Bayesian Stackelberg, networked, and embodied models that adapt to diverse application settings.
• These models underpin practical applications in security, inspection, privacy protection, and even in training LLMs by generating and diagnosing deceptive errors.

Searching arXiv for recent hide-and-seek game papers and related search-game literature. Hide and Seek Game (HSG) denotes, in the cited literature, a family of adversarial models in which one agent conceals an object, state, route, trajectory, source, or error, while another agent searches, infers, inspects, or diagnoses under explicit resource, information, and payoff constraints. The simplest instances reduce to a hider choosing one location and a searcher checking locations sequentially; more elaborate variants introduce imperfect detection, multiple targets, hypergraph action spaces, network design, route revelation, Bayesian inference, embodied reinforcement learning, virtual-reality intervention, and adversarial large-language-model diagnosis (Lidbetter, 14 Feb 2025, Theodorakopoulos et al., 2014, Chen et al., 2019, Weihs et al., 2019, Zou et al., 5 Aug 2025).

1. Scope and conceptual range

Across the literature, HSG is not a single universally fixed game. In one line of work it is a zero-sum search game over boxes or discrete locations, with search times $t_i$, detection probabilities $q_i$, and expected time to detection as payoff (Lidbetter, 14 Feb 2025, Clarkson et al., 2021). In another, it is a Bayesian Stackelberg game in which a user hides a trajectory behind pseudolocations and an adversary infers it (Theodorakopoulos et al., 2014). Other formulations place the game on hypergraphs with booby traps, networks with strategic inspection, Brownian motion on surfaces, embodied 3D environments, or adversarial LLM pipelines (Lidbetter et al., 2019, Bloch et al., 2020, Doyle et al., 2017, Weihs et al., 2019, Zou et al., 5 Aug 2025).

This suggests that HSG is best understood as a modeling template rather than a single canonical mathematical object. The common structure is adversarial concealment plus selective revelation: the hider allocates a scarce concealment resource, and the seeker allocates a scarce search, inspection, inference, or diagnostic resource.

Formulation	Hider action	Seeker action
Minimum-cost box search	Hide one or more targets in boxes	Search boxes sequentially
Trajectory privacy	Choose an LPPM $$f(o_{\text{post}} \mid a_{\text{trg}}, o_{\text{pre}})$$	Choose an inference attack $$h(\hat a_{\text{trg}} \mid o_{\text{pre}}, o_{\text{post}})$$
Hypergraph booby traps	Booby-trap $k$ boxes	Open one hyperedge
Network hide-and-seek	Design a graph and choose hiding node	Inspect one node
Embodied visual HSG	Hide, flee, or manipulate objects in 3D scenes	Pursue, search, or locate
LLM HSG	Generate stealthy reasoning errors	Diagnose and support correction

The historical lineage of the box-search branch is explicitly traced to Isaacs, Gal, Bram, Blackwell / Black, Bellman, Ruckle, Condon et al., Lidbetter, and Alpern, situating HSG within the broader field of search games (Lidbetter, 14 Feb 2025).

2. Canonical discrete search formulations

A central formalization is the minimum-cost box-searching game. There is a finite set of boxes

$\mathcal{B} = \{1,2,\ldots,n\},$

with search time $t_j>0$ and detection probability $q_j\in(0,1]$ for box $j$. The hider’s mixed strategy is $\mathbf p=(p_1,\dots,p_n)$, while the searcher’s pure strategy is an infinite sequence $q_i$0, because imperfect detection may require repeated visits. The value satisfies

$q_i$1

and the searcher’s best response against a fixed $q_i$2 is an index policy that orders looks by

$q_i$3

or equivalently, after posterior updating, chooses the box maximizing

$q_i$4

In the perfectly symmetric case $q_i$5 and $q_i$6, the value is

$q_i$7

and when $q_i$8 and $q_i$9, the value is

$$f(o_{\text{post}} \mid a_{\text{trg}}, o_{\text{pre}})$$0

For multiple targets under perfect detection, if the hider chooses $$f(o_{\text{post}} \mid a_{\text{trg}}, o_{\text{pre}})$$1 with $$f(o_{\text{post}} \mid a_{\text{trg}}, o_{\text{pre}})$$2, the equilibrium hiding distribution is

$$f(o_{\text{post}} \mid a_{\text{trg}}, o_{\text{pre}})$$3

and the value is

$$f(o_{\text{post}} \mid a_{\text{trg}}, o_{\text{pre}})$$4

These results make explicit how asymmetry in $$f(o_{\text{post}} \mid a_{\text{trg}}, o_{\text{pre}})$$5 and $$f(o_{\text{post}} \mid a_{\text{trg}}, o_{\text{pre}})$$6 breaks the simple “random permutation” solution and replaces it with priority-index scheduling (Lidbetter, 14 Feb 2025).

A related discrete formulation studies the same classical search game over $$f(o_{\text{post}} \mid a_{\text{trg}}, o_{\text{pre}})$$7 locations with imperfect detection and proves the existence of an optimal strategy for each player. In that treatment, the hider’s optimal mixed strategy hides in each location with a nonzero probability, and the searcher’s optimal mixed strategy can be constructed with up to $$f(o_{\text{post}} \mid a_{\text{trg}}, o_{\text{pre}})$$8 simple search sequences. A key equilibrium property is equalization: $$f(o_{\text{post}} \mid a_{\text{trg}}, o_{\text{pre}})$$9 so the searcher makes every hiding location yield the same expected detection time (Clarkson et al., 2021).

Other discrete HSG variants change the action geometry. In the caching game, the hider buries $$h(\hat a_{\text{trg}} \mid o_{\text{pre}}, o_{\text{post}})$$0 objects among $$h(\hat a_{\text{trg}} \mid o_{\text{pre}}, o_{\text{post}})$$1 locations with total deepest-depth sum at most $$h(\hat a_{\text{trg}} \mid o_{\text{pre}}, o_{\text{post}})$$2, while the searcher has digging capacity $$h(\hat a_{\text{trg}} \mid o_{\text{pre}}, o_{\text{post}})$$3 and wins only by finding all objects. For $$h(\hat a_{\text{trg}} \mid o_{\text{pre}}, o_{\text{post}})$$4 and $$h(\hat a_{\text{trg}} \mid o_{\text{pre}}, o_{\text{post}})$$5, the value is asymptotically

$$h(\hat a_{\text{trg}} \mid o_{\text{pre}}, o_{\text{post}})$$6

for $$h(\hat a_{\text{trg}} \mid o_{\text{pre}}, o_{\text{post}})$$7 (Csóka et al., 2015). In the hypergraph game with booby traps, the searcher opens a hyperedge $$h(\hat a_{\text{trg}} \mid o_{\text{pre}}, o_{\text{post}})$$8 and receives $$h(\hat a_{\text{trg}} \mid o_{\text{pre}}, o_{\text{post}})$$9 only if $k$0; the 1-uniform case, the complete hypergraph with equal rewards, the case $k$1, and the case $k$2 are solved, but the authors conclude that a general simple, closed form solution appears unlikely (Lidbetter et al., 2019). In the capacitated-location model, seeker pure actions are $k$3, hider pure actions are $k$4, and the expected number of undetected items is

$k$5

There, mixed-strategy Nash equilibria are characterized by unidimensional marginals, then lifted to equilibrium mixed strategies in quadratic time with linear support (Bahamondes et al., 2023).

3. Information, dynamics, and network structure

Several HSG formulations turn on what information is revealed during search. In trajectory privacy, the hider is the user or device running an LPPM

$k$6

the seeker is a strategic adversary using

$k$7

and the prior is

$k$8

The payoff is expected adversarial estimation error, and mechanism design is cast as

$k$9

subject to

$\mathcal{B} = \{1,2,\ldots,n\},$0

This makes the HSG explicitly Bayesian and Stackelberg rather than simultaneous-move zero-sum in the narrow matrix-game sense (Theodorakopoulos et al., 2014).

A route-revelation variant gives the hider partial information about the seeker’s future path. The seeker chooses a permutation $\mathcal{B} = \{1,2,\ldots,n\},$1 of locations, the hider observes a prefix

$\mathcal{B} = \{1,2,\ldots,n\},$2

and may relocate the treasure once to an unvisited node at switching cost $\mathcal{B} = \{1,2,\ldots,n\},$3. The restricted model yields a payoff matrix $\mathcal{B} = \{1,2,\ldots,n\},$4; the seeker-aware feedback model yields $\mathcal{B} = \{1,2,\ldots,n\},$5. The value-of-information satisfies

$\mathcal{B} = \{1,2,\ldots,n\},$6

and the informational benefit decreases with reveal time: $\mathcal{B} = \{1,2,\ldots,n\},$7 Seeker awareness reduces the hider’s value: $\mathcal{B} = \{1,2,\ldots,n\},$8 This is a distinct HSG theme: partial route revelation converts a static hiding decision into a dynamic relocation problem (Surve et al., 27 Mar 2026).

Networked HSGs use graph structure itself as a strategic object. In infection spreading and source identification, the hider is the infection source on a tree $\mathcal{B} = \{1,2,\ldots,n\},$9, the seeker is a network administrator, and the administrator probes nodes within radius $t_j>0$0 of a Jordan center of the infection graph. The source controls infection rates subject to upper bounds, and the relevant concealment statistic is the safety margin

$t_j>0$1

The best response of the source is either the maximum-spread strategy $t_j>0$2 or the just-outside-the-probe strategy $t_j>0$3; pure-strategy Nash equilibria are characterized by $t_j>0$4 or $t_j>0$5 under explicit gain–cost inequalities (Luo et al., 2015). In a different network formulation, the hider chooses both the graph $t_j>0$6 and hiding node $t_j>0$7, the seeker observes $t_j>0$8 and attacks one node, and optimal networks are either equivalent to cycles or variants of core–periphery networks (Bloch et al., 2020).

At the opposite end of abstraction, HSG can be continuous and spectral. On a compact Riemannian surface $t_j>0$9, the seeker moves according to Brownian motion and the hiding point is random with respect to normalized volume. The finite part of the expected duration of Game I is a spectral invariant, identified with the regularized trace $q_j\in(0,1]$0 of the Laplacian; the Markov-chain analogue yields Kemeny’s constant

$q_j\in(0,1]$1

In that setting, the hide-and-seek duration is not merely a performance measure but a spectral quantity (Doyle et al., 2017).

4. Embodied, perceptual, and clinical HSGs

In embodied reinforcement-learning formulations, HSG becomes a partially observable control problem in a 3D world. In “Visual Hide and Seek,” the hider is a prey policy in a Unity environment, the seeker is a hand-crafted predator, observations are egocentric RGB frames, and the hider’s reward includes $q_j\in(0,1]$2 per step before capture and $q_j\in(0,1]$3 on collision. The hider is trained with PPO, no explicit memory is used, and post hoc logistic regression on frozen mid-level features shows that the agent learns to predict its own visibility: self-visibility awareness reaches $q_j\in(0,1]$4 in the visibilityreward variant and $q_j\in(0,1]$5 in stochasticmaps + stochasticseeker, far above random-initialized baselines (Chen et al., 2019).

A richer object-hiding formulation is “Cache,” played in AI2-THOR. The game is decomposed into five RL subtasks: Exploration & Mapping, Perspective Simulation, Object Hiding, Object Manipulation, and Seeking. The agent architecture combines a 13-layer U-Net–style encoder, a Convolutional GRU, an LSTM, a metric map, and stage-specific actor–critic heads trained with A3C and GAE. The resulting dynamic image representations encode developmental-psychology-like regularities: containment-versus-behind classification reaches approximately $q_j\in(0,1]$6, occluded-object tracking approximately $q_j\in(0,1]$7, and free-space seriation $q_j\in(0,1]$8. The same work reports that SIR features learned only by hide-and-seek gameplay are comparable to ImageNet-pretrained SIRs on 6 tasks—statistically indistinguishable in 4 and significantly better in 2 (Weihs et al., 2019).

In a clinical VR formulation, the Hide and Seek Virtual Reality System uses a virtual family room, a parent-like avatar, integrated 7Invensun eye tracking at $q_j\in(0,1]$9 Hz, and a gaze-based success rule: when the user’s gaze remains within the avatar’s body box for $j$0 ms, the system determines that the avatar has been found and moves to the next round. A pilot study at the Third Affiliated Hospital of Sun Yat-sen University involved $j$1 children with ASD. The customized-avatar group showed Face Fixation Proportion $j$2 versus $j$3 in the uncustomized group, and Background Fixation Proportion $j$4 versus $j$5, alongside better questionnaire and game-performance outcomes. Here HSG functions as an auxiliary intervention for gaze fixation and related social-communication behaviors (Yu et al., 2023).

These embodied and clinical variants depart from the canonical zero-sum matrix-game setting, but they preserve the central HSG motif: useful behavior emerges from selective concealment, partial observability, and adversarial or game-like search pressure.

5. Adversarial reasoning and language-model HSGs

A recent LLM formulation makes the hide-and-seek metaphor explicit at the level of reasoning traces. “Sneaky” is the hider: it generates mathematically incorrect but subtle chain-of-thought solutions. “Diagnosis” is the seeker: it must decide whether the solution is correct, identify the error, and support correction. Both roles are instantiated as prompted heads over the same base model $j$6, while a fixed correction model $j$7 supplies feedback. The individual rewards are

$j$8

for Sneaky and

$j$9

for Diagnosis, with collaborative and adversarial feedback defined through correction success: $\mathbf p=(p_1,\dots,p_n)$0

$\mathbf p=(p_1,\dots,p_n)$1

Training alternates between improving Sneaky and training Diagnosis on the hardest currently generated examples (Zou et al., 5 Aug 2025).

On GSM8K, MATH, and NuminaMATH-TIR, this HSG significantly boosts correction accuracy. The reported gain relative to baseline diagnostics is $\mathbf p=(p_1,\dots,p_n)$2–$\mathbf p=(p_1,\dots,p_n)$3 across Qwen3-4B, Qwen3-8B, Qwen3-14B, DeepSeek, and GPT-4o. In the ablation on NuminaMATH-TIR, HSG-produced errors have average correction failure rate $\mathbf p=(p_1,\dots,p_n)$4, compared with $\mathbf p=(p_1,\dots,p_n)$5 for LLM-rater Adv+RL and $\mathbf p=(p_1,\dots,p_n)$6 for RL only; moreover, $\mathbf p=(p_1,\dots,p_n)$7 of HSG errors avoid both Type A and Type B shortcuts. This makes HSG not just a metaphor for adversarial training but a concrete mechanism for generating deceptive errors and self-improving diagnosis (Zou et al., 5 Aug 2025).

A plausible implication is that HSG has become a reusable abstraction for adversarial co-evolution: in classical search games the hidden object is spatial, while in LLM HSG it is a latent logical defect embedded in an otherwise coherent reasoning trace.

6. Interpretation, applications, and recurrent misconceptions

A recurrent misconception is that HSG always means a childlike pursuit game or a single-target box search. The cited literature contradicts that narrow reading. HSG may involve a stationary treasure, multiple targets, booby traps, hidden trajectories, dynamically spreading infections, network locations, object manipulation in 3D scenes, or deceptive reasoning errors (Lidbetter, 14 Feb 2025, Lidbetter et al., 2019, Theodorakopoulos et al., 2014, Luo et al., 2015, Zou et al., 5 Aug 2025). Likewise, the seeker may be a sequential inspector, a Bayesian adversary, a network administrator, an embodied agent, or a diagnostic critic.

A second misconception is that detection is necessarily perfect. Imperfect detection is central in several formulations: classical discrete-location search assumes box-specific $\mathbf p=(p_1,\dots,p_n)$8, the minimum-cost box-search review allows repeated visits because detection can fail, and the capacitated inspection game defines location-specific probabilities $\mathbf p=(p_1,\dots,p_n)$9 that hidden items remain undetected with probability $q_i$00 (Clarkson et al., 2021, Lidbetter, 14 Feb 2025, Bahamondes et al., 2023). In these models, revisitation, posterior updating, and index policies are not embellishments but equilibrium-defining structure.

A third misconception is that HSG is always static. Some models are static simultaneous-move zero-sum games, but others are sequential or feedback-rich. The trajectory-privacy model is explicitly Bayesian Stackelberg; the route-revelation model allows one relocation after observing a route prefix; embodied RL versions are POMDPs; the clinical VR version uses least-to-most prompting and reinforcement across repeated trials (Theodorakopoulos et al., 2014, Surve et al., 27 Mar 2026, Weihs et al., 2019, Yu et al., 2023).

The application range is correspondingly broad. The box-search review explicitly lists search and rescue, fault detection / maintenance, security / intrusion detection, quality control / inspection, and cyber “hide-and-seek” as natural interpretations (Lidbetter, 14 Feb 2025). The hypergraph model maps to military drone search and machine scheduling (Lidbetter et al., 2019). The trajectory-privacy model treats location-based services (Theodorakopoulos et al., 2014). The clinical VR system targets gaze fixation in ASD (Yu et al., 2023). The LLM formulation targets mathematical error generation and diagnosis (Zou et al., 5 Aug 2025).

Taken together, these works present HSG as a general research program on adversarial concealment and selective revelation. Its unifying questions are consistent across domains: what can the hider conceal under a limited resource budget, what can the seeker infer or recover under partial observability and risk, and what equilibrium structure emerges when both sides optimize against each other?