GuessingGame: Sequential Inference Overview

• GuessingGame is a family of sequential inference games where agents use strategic queries and controlled feedback to reveal hidden states.
• The literature spans coding theory, combinatorics, and multimodal dialogue, influencing applications from card games to LLM evaluations.
• Methodologies include optimal query policies, cyclic-shift strategies, and NP-hard partitioning challenges, highlighting both theoretical and practical complexities.

GuessingGame denotes a broad family of sequential inference games in which an agent, or a set of agents, attempts to identify a hidden object, symbol, word, permutation, card, or real number from constrained feedback. Across the literature, the shared structure is a hidden state, a policy for asking questions or making guesses, a feedback channel that may range from exact revelation to noisy or partial supervision, and an objective such as minimizing the expected number of questions, maximizing the number of correct guesses, or characterizing the distribution of solving times. The term appears in information-theoretic source coding, combinatorics, probability, card and word games, visual dialogue, and recent evaluation protocols for LLMs (0906.2864, Ardimanov et al., 2018, Vries et al., 2016, Hutson et al., 23 Sep 2025).

1. Canonical structure and feedback regimes

A GuessingGame typically consists of a hidden target, a sequence of guesses or questions, and a feedback rule that restricts what is learned after each move. In the twenty-question formulation, a ball is hidden in one of $20$ boxes, only “Yes” or “No” answers are allowed, and each question is treated as spending one bit; the goal is to minimize the expected number of questions (0906.2864). In "Permutation Wordle," the hidden state is a permutation $\sigma \in S_n$, the guesser proposes a permutation $\gamma_r$ at round $r$, and the feedback is the set of positions

$U_r=\{\,i\in[n]\mid \gamma_r(i)=\sigma(i)\,\},$

that is, the exact indices where the guess is correct (Kutin et al., 2024). In "Minimum Guesswork with an Unreliable Oracle," Bob may ask one binary question $A\subseteq[N]$ about a random variable $X$, but the oracle’s answer is corrupted through a binary symmetric channel with crossover probability $\epsilon$ (Ardimanov et al., 2018).

In card-guessing games, the hidden state is a shuffled deck and the feedback model is decisive. With complete feedback after a one-time Gilbert–Shannon–Reeds riffle shuffle, the card is revealed after every guess, so the player can adapt to the evolving structure of the deck (Liu, 2019). With partial feedback, by contrast, the Guesser is told only whether the current guess is correct, not the actual label of the card, and the resulting information constraint changes the asymptotic payoff from the complete-feedback regime (Nie, 2022). In limited-memory variants, the Guesser additionally has only $m<n$ memory bits and must update a bounded internal state from the transcript (Menuhin et al., 2021).

Visual and multimodal variants instantiate the same pattern in richer observation spaces. GuessWhat?! is a two-player visual guessing game in which one object in an image is the hidden target, the Questioner asks a sequence of natural-language yes/no/N/A questions, and the Oracle answers until the Questioner is ready to guess the object (Vries et al., 2016). Sketch-QA fixes the question to “What object is being drawn?” and requires guesses from a temporally accumulating sketch stroke sequence, producing a sequence of open-ended guess-words rather than a classification from a fixed label set (Sarvadevabhatla et al., 2018). Recent LLM-oriented protocols generalize the setting further: in "GuessingGame: Measuring the Informativeness of Open-Ended Questions in LLMs," a Guesser LLM identifies a hidden object by asking free-form questions to an Oracle, and a Checker enforces question-type constraints (Hutson et al., 23 Sep 2025).

Representative game	Hidden state	Feedback
Twenty Questions	Ball in one of $20$ boxes	Yes/No answers
Permutation Wordle	$\sigma \in S_n$0	Correct positions $\sigma \in S_n$1
Unreliable oracle guesswork	$\sigma \in S_n$2	Noisy binary or $\sigma \in S_n$3-ary answer
GuessWhat?!	Target object in an image	Yes/No/N/A dialogue
Card guessing	Order of a shuffled deck	Complete revelation or correctness only
Open-ended LLM GuessingGame	Hidden object $\sigma \in S_n$4	Free-form answers plus Checker constraints

The differences among these regimes are not superficial. They determine whether the game is equivalent to source coding, to a decision tree, to a constrained search problem, or to a sequential estimation problem under uncertainty.

2. Information-theoretic and coding-theoretic formulations

The twenty-question problem is the canonical coding-theoretic GuessingGame. For $\sigma \in S_n$5 equiprobable boxes, the position of the ball is a source symbol with entropy

$\sigma \in S_n$6

The paper compares three strategies: one-by-one asking, top-down division, and down-top merging. One-by-one asking consumes an expected $\sigma \in S_n$7 bits by the calculation

$\sigma \in S_n$8

although the paper also reports essentially the same quantity as around $\sigma \in S_n$9 bits. Top-down division, which repeatedly splits the candidate set as evenly as possible, yields

$\gamma_r$0

and down-top merging, which is explicitly Huffman-like, also yields

$\gamma_r$1

All three strategies extract the same total information, $\gamma_r$2, but differ sharply in expected effort. The paper’s conclusion is that Huffman-style down-top merging is optimal, whereas top-down division is only sub-optimal in general, even if the two coincide at $\gamma_r$3 bits in the $\gamma_r$4-box example (0906.2864).

The source-coding interpretation is explicit. A strategy of questions is a prefix-free binary code for the $\gamma_r$5 possible positions, with average length

$\gamma_r$6

subject to the entropy lower bound

$\gamma_r$7

and Huffman’s guarantee

$\gamma_r$8

The same logic reappears in oracle-assisted guesswork with noisy side information. When Bob asks Carole a binary question $\gamma_r$9, the answer is

$r$0

where $r$1. The central result is that, for every distribution $r$2, every $r$3, and every nondecreasing $r$4, the optimal question is the zigzag partition

$r$5

that is, “odd or even” in descending probability order. This partition simultaneously minimizes all guessing moments, settling a conjecture of Burin and Shayevitz. In the $r$6-ary generalization, when the noisy channel is completely symmetric, the optimal question is the residue class of the ordered index modulo $r$7; for more general symmetric channels, deciding whether a given question is optimal becomes NP-hard through a reduction to modular difference disequations (Ardimanov et al., 2018).

These results formalize a central principle of GuessingGame design: the amount of information eventually obtained may be fixed by the hidden-state space, but efficiency depends on whether the query policy aligns the feedback channel with the probability structure of the target.

3. Combinatorial and algebraic guesswork

In permutation-based GuessingGame models, solving time itself becomes a classical combinatorial statistic. In "Permutation Wordle," the hidden object is a permutation $r$8, and the cyclic-shift strategy begins with the identity permutation, preserves entries already known to be correct, and cyclically shifts the values in the currently incorrect positions. The main theorem states that the number of guesses required is exactly

$r$9

where $U_r=\{\,i\in[n]\mid \gamma_r(i)=\sigma(i)\,\},$0 is the number of excedances of $U_r=\{\,i\in[n]\mid \gamma_r(i)=\sigma(i)\,\},$1. Consequently, the number of permutations solved in $U_r=\{\,i\in[n]\mid \gamma_r(i)=\sigma(i)\,\},$2 guesses is the Eulerian number $U_r=\{\,i\in[n]\mid \gamma_r(i)=\sigma(i)\,\},$3, and the expected number of guesses is

$U_r=\{\,i\in[n]\mid \gamma_r(i)=\sigma(i)\,\},$4

The strategy always solves the game in at most $U_r=\{\,i\in[n]\mid \gamma_r(i)=\sigma(i)\,\},$5 guesses, and the authors conjecture that no strategy can have a smaller average number of rounds; for $U_r=\{\,i\in[n]\mid \gamma_r(i)=\sigma(i)\,\},$6, they checked experimentally that the cyclic-shift strategy dominates all others in the sense of maximizing the probability of success within any fixed number of rounds (Kutin et al., 2024).

The multicolor extension replaces each value by a colored card. For an $U_r=\{\,i\in[n]\mid \gamma_r(i)=\sigma(i)\,\},$7-colored hidden object $U_r=\{\,i\in[n]\mid \gamma_r(i)=\sigma(i)\,\},$8, the generalized cyclic-shift strategy requires

$U_r=\{\,i\in[n]\mid \gamma_r(i)=\sigma(i)\,\},$9

guesses. The corresponding counting sequence $A\subseteq[N]$0 has generating function

$A\subseteq[N]$1

and satisfies the recurrence

$A\subseteq[N]$2

When $A\subseteq[N]$3, this reduces to the Eulerian recurrence; when $A\subseteq[N]$4, grouped solving-time counts recover type $A\subseteq[N]$5 Eulerian numbers, linking the game to signed permutations (Kutin et al., 2024).

A different algebraic line appears in continuous guessing games. When players guess a random real number and the winner is either closest in magnitude or closest without exceeding the target, optimal strategies become mixed strategies on $A\subseteq[N]$6. In the two-player Price Is Right variant, an optimal strategy is

$A\subseteq[N]$7

with support bounded above by $A\subseteq[N]$8. For the $A\subseteq[N]$9-player Price Is Right game, if a differentiable optimal cdf exists, the least upper bound of the support is

$X$0

In the three-player Closest Wins game, the uniform distribution on $X$1 is optimal, while the four-player Closest Wins game admits an approximate symmetric power-series solution centered at $X$2 (Mendes et al., 2014).

These combinatorial and algebraic models are structurally different from source coding, but they preserve the same core question: how does one encode uncertainty reduction into a sequential policy under a specified payoff rule?

4. Card guessing, shuffle models, and asymptotic laws

Card-guessing GuessingGames study how far adaptive guessing can exploit structure in the deck distribution. For a deck labeled $X$3 and subjected to exactly one Gilbert–Shannon–Reeds riffle shuffle, the optimal complete-feedback strategy is explicit. The player guesses $X$4 first; if the revealed sequence remains consecutive $X$5, the next guess is $X$6; when the consecutive pattern breaks at a revealed card $X$7, the remaining cards are separated into

$X$8

and the player guesses the first card in the longer sequence. This "Algorithm 1.1" is optimal, and the maximum expected reward is

$X$9

for one time riffle shuffle with complete feedback (Liu, 2019). A generating-function treatment later provided an expression for the expected number of correct guesses with arbitrary number of terms and a unified framework for higher-order moments, and also showed that the optimal guessing strategy for $\epsilon$0 riffle shuffle does not necessarily apply to $\epsilon$1 shuffles, settling McGrath’s conjectured strategy in the negative (Krityakierne et al., 2021).

The distributional theory of the one-time riffle game is more precise. If $\epsilon$2 is the number of correct guesses under the optimal strategy and

$\epsilon$3

then

$\epsilon$4

where $\epsilon$5 has density

$\epsilon$6

Moreover,

$\epsilon$7

The first occurrence of a pure luck guess has its own limit law: if $\epsilon$8 is the number of remaining cards divided by two when the first pure luck guess occurs, then

$\epsilon$9

the arcsine distribution (Kuba et al., 2023).

Partial-feedback models alter the picture fundamentally. When a deck contains $m<n$0 copies of each of $m<n$1 labels and the Guesser is told only whether each guess is correct, the maximal expected payoff satisfies

$m<n$2

for every strategy $m<n$3. Combined with the earlier lower bound

$m<n$4

this pins down the regime $m<n$5 for large $m<n$6 and sufficiently large $m<n$7 (Nie, 2022).

The continuous-number and card-shuffle literatures therefore converge on a common theme: when the hidden state comes from a structured distribution rather than an arbitrary adversary, optimal GuessingGame play is governed by exact recurrences, generating functions, and non-Gaussian asymptotic laws.

5. Memory bounds, adversarial dealers, and computational hardness

GuessingGame complexity is often determined less by the visible rules than by what the guesser can remember and what the environment is allowed to do. In "Keep That Card in Mind: Card Guessing with Limited Memory," a Guesser with perfect memory can keep track of all cards that have appeared and guess uniformly from the unseen cards, obtaining $m<n$8 correct guesses in expectation. With no memory, the best achievable expectation is a single correct guess. For the intermediate regime, the paper proves three sharp distinctions. First, there is a Guesser with $m<n$9 memory bits that attains a near-optimal result against any static Dealer. Second, no Guesser with $20$0 bits of memory can score better than $20$1 correct guesses, so no Guesser can do better than

$20$2

Third, there is an efficient adaptive Dealer against which no Guesser with $20$3 memory bits can make more than

$20$4

correct guesses in expectation. The proofs use compression arguments that encode information through the guessing strategy itself (Menuhin et al., 2021).

Oracle-assisted guessing also exposes sharp computational boundaries. In the binary noisy-oracle model, the odd/even zigzag partition is globally optimal, but when the side-information query becomes $20$5-ary and the symmetric channel is no longer completely uniform, testing whether a given question is optimal becomes NP-hard. The hardness reduction proceeds through systems of modular difference disequations

$20$6

showing that even the feasibility of a posterior-respecting partition can be intractable for $20$7 (Ardimanov et al., 2018).

A related design problem appears in “Reversing The Twenty Questions Game,” where the computer chooses the hidden entity and the human asks arbitrary natural-language questions. The implementation combines BM25 retrieval, Sentence-BERT similarity, and a BoolQ-style transformer-based boolean QA model to answer with yes, no, probably yes, probably no, or idk. Restricting the domain to animals, the improved pipeline reaches $20$8 accuracy against hand annotations on $20$9 questions and $\sigma \in S_n$00 on an automatically labeled set of $\sigma \in S_n$01 questions, while also revealing persistent failure modes for comparison questions, missing facts, and cultural references that dominate retrieved evidence (Parikh et al., 2023).

These results show that GuessingGame difficulty can arise from at least three distinct sources: state compression, adversarial adaptation, and the algorithmic complexity of selecting or validating informative queries.

6. Visual, multimodal, and grounded dialogue games

In visual GuessingGame research, the hidden state is not an abstract symbol but an object embedded in a scene. GuessWhat?! established the modern benchmark form: a cooperative, two-player, turn-based visual guessing game on MS COCO images. The dataset contains $\sigma \in S_n$02 dialogues, $\sigma \in S_n$03 question-answer pairs, $\sigma \in S_n$04 unique images, and $\sigma \in S_n$05 unique objects. The average number of questions per dialogue is $\sigma \in S_n$06, and answer proportions are $\sigma \in S_n$07 No, $\sigma \in S_n$08 Yes, and $\sigma \in S_n$09 N/A. The Oracle receives a question, the image, and the target object; the Questioner sees the image, conducts the dialogue, and eventually guesses the target object (Vries et al., 2016).

Later work replaced task-specific encoders with pretrained vision-language representations and explicit state tracking. VilBERT-based Oracle, Guesser, and Questioner architectures in GuessWhat?! introduced two-way background/target fusion for the Oracle and a state-estimator for the Guesser and Questioner. Experimentally, the proposed models outperform state-of-the-art models significantly by $\sigma \in S_n$10, $\sigma \in S_n$11, and $\sigma \in S_n$12 for Oracle, Guesser, and End-to-End Questioner respectively; in the reported tables, VilBERT-Oracle reaches $\sigma \in S_n$13, VilBERT-Guesser $\sigma \in S_n$14, and the full VilBERT-Oracle + VilBERT-Guesser + VilBERT-Questioner system attains $\sigma \in S_n$15 end-to-end success rate (Tu et al., 2021).

A complementary line asks whether the representations learned by playing guessing games transfer beyond the original game. In “An Empirical Study on the Generalization Power of Neural Representations Learned via Visual Guessing Games,” the authors consider supervised learning on successful GuessWhat?! games and a self-play regime called Self-play via Iterated Experience Learning (SPIEL). They report an increased accuracy of $\sigma \in S_n$16 on the evaluation suite CompGuessWhat?! and improved VQA performance on TDIUC in terms of harmonic average accuracy of $\sigma \in S_n$17, attributing the gains to more fine-grained object representations learned via SPIEL (Suglia et al., 2021).

Sketch-based variants move the visual GuessingGame into the temporal regime. “Game of Sketches” introduces Sketch-QA, in which the fixed question is “What object is being drawn?” and the visual observation is an incrementally accumulated sketch sequence rather than a completed image. The resulting WordGuess-160 dataset contains $\sigma \in S_n$18 guess sequences from $\sigma \in S_n$19 human subjects over $\sigma \in S_n$20 object categories. The proposed recurrent model, which predicts guess-word embeddings as strokes accumulate, reaches $\sigma \in S_n$21 top-1 sequence-level accuracy, compared with $\sigma \in S_n$22 for the two-phase baseline and $\sigma \in S_n$23 for the CNN-only baseline, and the qualitative evaluation emphasizes that the model makes human-like mistakes while guessing (Sarvadevabhatla et al., 2018).

Taken together, these multimodal systems show that GuessingGame can serve not only as a task in its own right but also as a training interface for grounded state estimation, object-centric representation learning, and sequential language generation.

7. Word games, LLM evaluation, and strategic question-asking

Word-based GuessingGame settings provide a bridge from combinatorial puzzle play to interactive evaluation of LLMs. In “Strategic Insights in Human and LLM Tactics at Word Guessing Games,” the game is a five-letter Wordle-style daily puzzle with at most six guesses and per-letter feedback indicating correct position, different position, or absence from the word. A survey on a Latvian adaptation collected $\sigma \in S_n$24 responses, estimated as about $\sigma \in S_n$25 of frequent players. Reported motivations are $\sigma \in S_n$26 entertainment and pastime, $\sigma \in S_n$27 mental exercise and language development, $\sigma \in S_n$28 routine and habit, and $\sigma \in S_n$29 social interaction and competition. The Latvian corpus also shows stable opening heuristics: “SAULE” was used nearly $\sigma \in S_n$30 times as the first word, while “SIENA” and “TIESA” appeared about $\sigma \in S_n$31 and $\sigma \in S_n$32 times respectively. LLM comparison reveals recurrent failures to maintain the five-letter constraint, avoid repetitions, and prevent hallucinated non-words or inflectional forms, especially in Latvian; among the tested systems, Gemini Advanced is the strongest overall performer (Rikters et al., 2024).

Spy-style language games emphasize a different capacity: strategic communication under asymmetric information. In “Leveraging Word Guessing Games to Assess the Intelligence of LLMs,” DEEP evaluates aggressive and conservative word descriptions, while SpyGame places a guest spy model into a four-player board-game-style setting with three GPT-3.5-Turbo hosts. For GPT-4 as the guest spy, the reported SpyGame metrics are Win $\sigma \in S_n$33, Round $\sigma \in S_n$34, and Voted $\sigma \in S_n$35, outperforming Text-Davinci-002, Text-Davinci-003, and GPT-3.5-Turbo. The same study also introduces Theory of Mind metrics such as Self-Identity, 1-Word, 1-Identity, 2-Word, and 2-Identity, making GuessingGame a probe of disguising ability and multi-agent reasoning rather than simple lexical search (Liang et al., 2023).

The most explicit LLM-centric formalization appears in “GuessingGame: Measuring the Informativeness of Open-Ended Questions in LLMs.” Here the Guesser LLM identifies a hidden object $\sigma \in S_n$36 by asking free-form questions, the Oracle answers, and a Checker classifies each question as Attribute, Function, Location, Category, or Direct while enforcing prompting constraints. The game runs for up to $\sigma \in S_n$37 turns. Two information-gain metrics are proposed. The Bayesian metric tracks a belief state over concepts using an Interpreter LLM and defines per-turn information gain as

$\sigma \in S_n$38

The entropy-based metric tracks a candidate set $\sigma \in S_n$39 using ConceptNet and defines

$\sigma \in S_n$40

Across $\sigma \in S_n$41 games, higher IG strongly predicts shorter game length: a one-standard-deviation increase in Bayesian IG reduces expected game length by $\sigma \in S_n$42, and prompting constraints based on IG, such as enforcing question diversity or open-ended questions, substantially improve success rates even for weaker models (Hutson et al., 23 Sep 2025).

This suggests a broader reinterpretation of GuessingGame. In older formulations the game measured efficient coding or combinatorial structure; in current LLM work it also measures whether a model can identify what it does not know, formulate discriminative queries, and use feedback strategically in open-domain interaction.

8. Open problems and recurring themes

Several open problems recur across otherwise distant GuessingGame literatures. Optimality often remains conjectural outside small or highly structured cases. The cyclic-shift strategy in permutation Wordle is proved to solve any permutation in at most $\sigma \in S_n$43 guesses, but its full optimality is still conjectural (Kutin et al., 2024). In card guessing after more than one riffle shuffle, the $\sigma \in S_n$44 optimal policy no longer extends directly, and the extension to $\sigma \in S_n$45 remains technically difficult (Krityakierne et al., 2021). In noisy-oracle guesswork, binary and completely symmetric $\sigma \in S_n$46-ary cases admit clean optimal questions, but more general symmetric channels lead to NP-hard feasibility questions (Ardimanov et al., 2018).

A second recurring theme is the decisive role of feedback granularity. Complete revelation in card guessing yields $\sigma \in S_n$47 corrections on top of a linear $\sigma \in S_n$48 main term after one riffle shuffle, whereas partial correctness-only feedback yields $\sigma \in S_n$49 in the repeated-card model (Kuba et al., 2023, Nie, 2022). In GuessWhat?!, richer multimodal context and state tracking make end-to-end success rates substantially higher than earlier baselines (Tu et al., 2021). In LLM evaluation, moving from yes/no-only word guessing to open-ended question-asking changes the object of measurement from mere answer accuracy to strategic informativeness (Hutson et al., 23 Sep 2025).

A plausible implication is that GuessingGame functions less as a single benchmark than as a common formal idiom for studying sequential information acquisition. The hidden state may be a source symbol, a deck permutation, an image region, a word, or an open-domain concept, but the central questions remain stable: what feedback is available, how much uncertainty does each move remove, how much memory is required, and when is the induced policy actually optimal.