Slot Guessing: Inference & Decision Strategies

• Slot guessing is a family of inference problems that predict the content or value of a ‘slot’ under partial observability across domains like card games, object-centric learning, and dialogue state tracking.
• The approaches span sequential guessing with binary feedback, one-shelf shuffles with position matrices, adaptive latent slot selection using differentiable sampling, and slot-value updating in dialogue state tracking.
• Techniques leverage probabilistic bounds, coupling, supermartingale analysis, and gating mechanisms to optimize prediction accuracy in settings with constrained feedback and uncertainty.

Searching arXiv for the cited papers to ground the article in current records. Verifying metadata via the arXiv API. $S_i$45 Slot guessing is a family of inference and decision problems in which an agent predicts the content, identity, number, or value associated with a “slot” under limited information. In recent arXiv literature, the term is used for several technically distinct settings: sequential guessing of cards drawn from a multiset deck with only correct/incorrect feedback (Nie, 2022); prediction of card identities at positions generated by a one-time single-shelf shuffle (Clay, 14 Jul 2025); dynamic selection of active latent slots in object-centric learning (Fan et al., 2024); and per-turn slot-value decisions in dialogue state tracking (DST) via inheritance, re-extraction, or memory-based selection (Sun et al., 2022). The common structure is partial observability, but the mathematical objects called “slots,” the available feedback, and the optimality criteria vary substantially.

1. Terminological scope and problem classes

In the cited literature, “slot” denotes different primitives. In sequential card-guessing, a slot is a position in a deck or shuffled output sequence. In object-centric learning, a slot is a latent vector intended to represent an object. In DST, a slot is a schema field such as a domain attribute whose value must be updated turn by turn.

Setting	Meaning of slot	Objective
Multiset deck guessing	One of $mn$ sequential card positions	Maximize expected total correct guesses
One-shelf shuffle	Position $j\in\{1,\dots,n\}$ in the shuffled deck	Maximize expected reward under no feedback or complete feedback
Adaptive Slot Attention	Candidate latent object slot $S_i$	Dynamically determine how many slots to keep
Dialogue state tracking	Schema slot $s$ with value $y_{s,t}$	Decide inheritance, re-extraction, or memory selection

This polysemy matters technically. In the card and shuffle settings, slot guessing is a stochastic control problem over permutations or multiset permutations. In Adaptive Slot Attention, it is a discrete latent-selection mechanism embedded in an auto-encoder. In DST, it is a structured prediction problem over dialogue turns with slot-specific memory and gating.

2. Sequential guessing with partial yes/no feedback

A canonical formalization considers a deck with $m$ copies of each of $n$ distinct cards, shuffled uniformly at random. At time $t=1,\dots,mn$, the Dealer draws uniformly at random from the remaining deck, the Guesser chooses $g_t\in\{1,\dots,n\}$, receives only $y_t\in\{0,1\}$ indicating whether the guess is correct, and the card is discarded. The objective is to maximize

$j\in\{1,\dots,n\}$0

A useful notation is

$j\in\{1,\dots,n\}$1

$j\in\{1,\dots,n\}$2

Whenever the denominator is positive, one has

$j\in\{1,\dots,n\}$3

The main upper bound proves that if $j\in\{1,\dots,n\}$4, then for any guessing strategy $j\in\{1,\dots,n\}$5,

$j\in\{1,\dots,n\}$6

In particular, for $j\in\{1,\dots,n\}$7, the maximal expected number of correct slots is $j\in\{1,\dots,n\}$8. This matches the Diaconis–Graham–Spiro lower bound in the regime $j\in\{1,\dots,n\}$9: if $S_i$0, there exists a strategy $S_i$1 such that

$S_i$2

(Nie, 2022).

The upper-bound proof is organized around truncation, coupling, and supermartingale control. With $S_i$3, a concentration lemma implies that with high probability $S_i$4, so the regime $S_i$5 contributes only an $S_i$6 tail. A coupled Boolean process $S_i$7 is then constructed with parameters $S_i$8 such that $S_i$9 almost surely whenever the one-step bound above is dominated by the corresponding $s$0-process mean, and

$s$1

whenever $s$2. Here $s$3 counts earlier $s$4 with $s$5, and satisfies

$s$6

The fluctuation analysis uses, for each $s$7 and $s$8,

$s$9

where $y_{s,t}$0 for $y_{s,t}$1, $y_{s,t}$2 for $y_{s,t}$3, and $y_{s,t}$4 for $y_{s,t}$5. The process $y_{s,t}$6 is a supermartingale under the $y_{s,t}$7-process. Optional stopping together with Azuma-type bounds yields

$y_{s,t}$8

Recovering the original process $y_{s,t}$9 from $m$0 gives the final estimate $m$1.

The lower-bound strategy is a “balanced-guess” rule: at each slot $m$2, guess the symbol $m$3 with the smallest gap

$m$4

equivalently the label with the largest posterior-style estimate

$m$5

A second-moment argument yields the $m$6 lower bound when $m$7. The accompanying information-theoretic interpretation is that under partial yes/no feedback one cannot identify which of the $m$8 labels was drawn except with probability $m$9, so the gain above the baseline $n$0 comes only from exploiting small count fluctuations. If $n$1, the method does not directly apply; it is conjectured that the same $n$2 bound holds for all $n$3, but new ideas would be needed when $n$4 is very small.

3. Slot guessing for one-time single-shelf shuffles

A distinct slot-guessing model studies a deck of $n$5 distinct cards labeled $n$6 under a one-time single-shelf shuffle. Cards are drawn from the bottom of the deck one at a time and placed on the top or bottom of a new pile with probability $n$7 each, producing a random permutation $n$8 of $n$9. Slots are positions $t=1,\dots,mn$0 in the shuffled deck, with $t=1,\dots,mn$1 the top and $t=1,\dots,mn$2 the bottom. The central object is the position matrix

$t=1,\dots,mn$3

which has the explicit form

$t=1,\dots,mn$4

The matrix is symmetric under $t=1,\dots,mn$5 and doubly stochastic (Clay, 14 Jul 2025).

In the no-feedback case, an observer pre-commits to guesses $t=1,\dots,mn$6 and sees no feedback until all cards are revealed. By linearity of expectation,

$t=1,\dots,mn$7

Thus each slot is optimized independently by choosing from the column-argmax set

$t=1,\dots,mn$8

Proven optimal choices in special positions are as follows: at slots $t=1,\dots,mn$9 and $g_t\in\{1,\dots,n\}$0, the unique best guess is card $g_t\in\{1,\dots,n\}$1; at slots $g_t\in\{1,\dots,n\}$2 and $g_t\in\{1,\dots,n\}$3, the unique best guesses are cards $g_t\in\{1,\dots,n\}$4 or $g_t\in\{1,\dots,n\}$5; if $g_t\in\{1,\dots,n\}$6 is even, at the two middle slots $g_t\in\{1,\dots,n\}$7 and $g_t\in\{1,\dots,n\}$8, the best guesses are cards $g_t\in\{1,\dots,n\}$9 or $y_t\in\{0,1\}$0; if $y_t\in\{0,1\}$1 is odd, at the unique middle slot $y_t\in\{0,1\}$2, the best guesses are $y_t\in\{0,1\}$3. The combinatorial basis includes the fact that cards $y_t\in\{0,1\}$4 and $y_t\in\{0,1\}$5 move always as an equally-likely adjacent block.

A conjectured general no-feedback rule for $y_t\in\{0,1\}$6 is specified by a piecewise prescription involving $y_t\in\{0,1\}$7 and an integer $y_t\in\{0,1\}$8 determined by the nearest square of the form $y_t\in\{0,1\}$9 or $j\in\{1,\dots,n\}$00. The rule is: guess $j\in\{1,\dots,n\}$01 at $j\in\{1,\dots,n\}$02; guess card $j\in\{1,\dots,n\}$03 or $j\in\{1,\dots,n\}$04 for $j\in\{1,\dots,n\}$05; guess card $j\in\{1,\dots,n\}$06 for $j\in\{1,\dots,n\}$07; and guess card $j\in\{1,\dots,n\}$08 or $j\in\{1,\dots,n\}$09 for $j\in\{1,\dots,n\}$10. Symmetry completes the sequence for $j\in\{1,\dots,n\}$11. The conjecture agrees with all computed instances up to several hundred.

The complete-feedback case is qualitatively different. Now the player guesses slot by slot, sees each card face up immediately, and may adapt future guesses. Defining $j\in\{1,\dots,n\}$12 as the revealed-cards process and $j\in\{1,\dots,n\}$13, an optimal strategy is $j\in\{1,\dots,n\}$14-predictable and maximizes conditional transition probability at each step. The optimal and unique strategy is: set $j\in\{1,\dots,n\}$15; as long as neither $j\in\{1,\dots,n\}$16 nor $j\in\{1,\dots,n\}$17 has appeared, if the last revealed card is $j\in\{1,\dots,n\}$18, guess $j\in\{1,\dots,n\}$19; if $j\in\{1,\dots,n\}$20, guess $j\in\{1,\dots,n\}$21 or $j\in\{1,\dots,n\}$22; once the first of $j\in\{1,\dots,n\}$23 appears at time $j\in\{1,\dots,n\}$24, the remaining unseen cards are in descending order almost surely, and guessing that descending order is then perfect. In the heuristic phrasing of the paper, one should guess “$j\in\{1,\dots,n\}$25” until seeing $j\in\{1,\dots,n\}$26 or $j\in\{1,\dots,n\}$27, then reverse to the descending order.

The proof uses a stopping time

$j\in\{1,\dots,n\}$28

Before $j\in\{1,\dots,n\}$29, the process $j\in\{1,\dots,n\}$30 is a time-homogeneous Markov chain with

$j\in\{1,\dots,n\}$31

while all competing transitions have probability at most $j\in\{1,\dots,n\}$32, so $j\in\{1,\dots,n\}$33 is the unique best guess. Moreover, $j\in\{1,\dots,n\}$34, and after $j\in\{1,\dots,n\}$35 the remaining cards are deterministically decreasing. The exact expected reward is

$j\in\{1,\dots,n\}$36

The decomposition is explicit: $j\in\{1,\dots,n\}$37; the pre-$j\in\{1,\dots,n\}$38 contribution is $j\in\{1,\dots,n\}$39, since each guess before $j\in\{1,\dots,n\}$40 is correct with conditional probability $j\in\{1,\dots,n\}$41; the post-$j\in\{1,\dots,n\}$42 contribution is $j\in\{1,\dots,n\}$43, since the rest is guessed with probability $j\in\{1,\dots,n\}$44. These results rigorously establish, in the one-shelf special case, the Diaconis–Fulman–Holmes conjecture that the “monotone-until-descent-then-reverse” strategy is optimal and yields about $j\in\{1,\dots,n\}$45 correct guesses.

4. Dynamic slot-number selection in object-centric learning

In object-centric learning, “slot guessing” can denote the mechanism that decides how many latent slots should remain active for a given input. Adaptive Slot Attention (“AdaSlot”) begins with a standard Slot-Attention bottleneck that produces candidate slots

$j\in\{1,\dots,n\}$46

For each slot $j\in\{1,\dots,n\}$47, a Bernoulli keep/drop distribution is predicted by a small MLP $j\in\{1,\dots,n\}$48: $j\in\{1,\dots,n\}$49 where $j\in\{1,\dots,n\}$50 is the keep probability and $j\in\{1,\dots,n\}$51 the drop probability. A hard, differentiable binary decision is then obtained through a Gumbel-Softmax with straight-through estimator: $j\in\{1,\dots,n\}$52 so that the chosen slot count is

$j\in\{1,\dots,n\}$53

Because $j\in\{1,\dots,n\}$54 is factorized in mean-field form, sampling $j\in\{1,\dots,n\}$55 independent Bernoulli variables is efficient and end-to-end differentiable (Fan et al., 2024).

Each slot is decoded into an object reconstruction and raw mask logits,

$j\in\{1,\dots,n\}$56

with spatial soft masks

$j\in\{1,\dots,n\}$57

Dropped slots are suppressed by zeroing their masks and renormalizing: $j\in\{1,\dots,n\}$58 This “zero-mask” strategy ensures that dropped slots play no role in the mixture.

The per-example loss is

$j\in\{1,\dots,n\}$59

with reconstruction term

$j\in\{1,\dots,n\}$60

and complexity regularizer

$j\in\{1,\dots,n\}$61

Larger $j\in\{1,\dots,n\}$62 encourages fewer slots; $j\in\{1,\dots,n\}$63 in practice. The end-to-end algorithm encodes the image, runs slot attention, predicts keep probabilities, samples the binary mask, decodes each slot, masks and mixes the outputs, and backpropagates through all modules via the Gumbel-Softmax straight-through estimator.

The experimental setup uses frozen DINO-ViT/B-16 features; $j\in\{1,\dots,n\}$64 for CLEVR10 and MOVi-C, $j\in\{1,\dots,n\}$65 for MOVi-E, and $j\in\{1,\dots,n\}$66 for COCO; slot dimension $j\in\{1,\dots,n\}$67; $j\in\{1,\dots,n\}$68 attention iterations; a decoder MLP of $j\in\{1,\dots,n\}$69 layers; Adam with learning rate $j\in\{1,\dots,n\}$70, $j\in\{1,\dots,n\}$71k warm-up steps, exponential decay, and $j\in\{1,\dots,n\}$72k total steps; $j\in\{1,\dots,n\}$73 on MOVi and $j\in\{1,\dots,n\}$74 on COCO. Evaluation uses ARI, Precision, Recall, F1, mBO, CorLoc, AMI, and NMI. On foreground ARI / F1 / mBO / CorLoc, the reported comparisons are: MOVi-C, AdaSlot $j\in\{1,\dots,n\}$75 versus fixed-6-slot DINOSAUR $j\in\{1,\dots,n\}$76; MOVi-E, AdaSlot $j\in\{1,\dots,n\}$77 versus best fixed-9-slot DINOSAUR $j\in\{1,\dots,n\}$78; COCO, AdaSlot $j\in\{1,\dots,n\}$79 versus best fixed-6-slot DINOSAUR $j\in\{1,\dots,n\}$80. Qualitatively, predicted slot counts correlate with true object counts across datasets, and on CLEVR10 the paper reports ARI $j\in\{1,\dots,n\}$81, compared with fixed-11-slot ARI $j\in\{1,\dots,n\}$82, while noting that the fixed-slot model over-segments background.

The limitations are explicit. In very cluttered real-world scenes such as COCO, AdaSlot sometimes merges small objects or splits semantic parts, limiting precise IoU. The mean-field assumption over $j\in\{1,\dots,n\}$83 ignores higher-order slot correlations. Complex part–whole hierarchies remain challenging, and performance on tiny dense objects is weaker because reconstruction is performed at feature resolution. Future work may explore structured priors over $j\in\{1,\dots,n\}$84 or joint discovery of part–whole hierarchies.

5. Slot-value guessing in dialogue state tracking

In DST, slot guessing refers to the per-turn decision process that determines, for each slot $j\in\{1,\dots,n\}$85, whether to inherit its previous value $j\in\{1,\dots,n\}$86, reject inheritance and re-extract a value from the current dialogue context $j\in\{1,\dots,n\}$87, or pick its value indirectly from a Mentioned Slot Pool (MSP). Formally, at turn $j\in\{1,\dots,n\}$88, one predicts

$j\in\{1,\dots,n\}$89

by combining a gating or hit-type decision with either inheritance, re-extraction, or selection from the MSP. The motivation is twofold: naive changed-state models propagate a wrong value across many turns, and indirectly mentioned slots such as restaurant-day $j\in\{1,\dots,n\}$90 train-day may not be explicitly recoverable from the current context alone (Sun et al., 2022).

The model encodes the dialogue context

$j\in\{1,\dots,n\}$91

with a pretrained neural LLM after adding a $j\in\{1,\dots,n\}$92 token: $j\in\{1,\dots,n\}$93 For each slot $j\in\{1,\dots,n\}$94, the MSP

$j\in\{1,\dots,n\}$95

contains up to $j\in\{1,\dots,n\}$96 candidate values drawn from the previous value $j\in\{1,\dots,n\}$97 and from up to three relevant slots $j\in\{1,\dots,n\}$98 whose values $j\in\{1,\dots,n\}$99 may inherit. These candidates are embedded into $S_i$00, then fused with context by low-rank bilinear soft attention: $S_i$01

Hit-type prediction produces

$S_i$02

with four classes: none, dontcare, mentioned, and hit. If the model predicts “mentioned,” it selects from the MSP using

$S_i$03

and sets $S_i$04 to the argmax candidate. If it predicts “hit,” then categorical slots are updated by classification,

$S_i$05

whereas span slots use start/end prediction,

$S_i$06

After predicting $S_i$07, the MSP for turn $S_i$08 is updated by appending $S_i$09 when it is not none, together with relevant-slot entries.

Training supervises hit-type, mentioned-value, and hit-value prediction: $S_i$10

$S_i$11

$S_i$12

with joint loss

$S_i$13

where $S_i$14, $S_i$15, and $S_i$16. Optimization uses BERT-base or BERT-large, Adam with learning rate $S_i$17, $S_i$18 epochs, $S_i$19 warmup, linear decay, and early stopping on development-set JGA.

The evaluation datasets are MultiWOZ 2.1, MultiWOZ 2.2, and WOZ 2.0. The primary metric is Joint Goal Accuracy (JGA), with slot-level Precision/Recall/F1 also reported. On MultiWOZ 2.1 / 2.2 / WOZ 2.0, the reported JGA values are: TRADE $S_i$20; SUMBT $S_i$21; DS-DST $S_i$22; TripPy $S_i$23; MSP-B $S_i$24; Seq2Seq-DU $S_i$25; AG-DST-S $S_i$26; and MSP-L $S_i$27. The paper states that MSP-B improves over TripPy on MultiWOZ 2.1 by $S_i$28 JGA, and over DS-DST on MultiWOZ 2.2 by $S_i$29, while MSP-L sets new SOTA among models without external data. In the update-strategy comparison on MultiWOZ, pure context gives $S_i$30, changed state $S_i$31, full state (AG-DST) $S_i$32, and MSP $S_i$33. In $S_i$34 manually identified indirectly mentioned slot cases, MSP-full corrects $S_i$35 ($S_i$36) by selecting from relevant-slot entries.

6. Comparative interpretation and open directions

Across these literatures, slot guessing is unified less by a shared ontology than by a shared decision pattern: a model must commit to a slot-level prediction while only partially observing the latent state. In the multiset deck problem, the uncertainty is combinatorial and the feedback is binary. In the one-shelf shuffle model, the uncertainty is determined by an explicit position matrix and by whether feedback is absent or complete. In AdaSlot, the uncertainty lies in how many latent object slots an input should activate. In DST, uncertainty concerns whether a slot should preserve prior state, retrieve from a memory of mentions, or be re-extracted from current evidence.

Several contrasts are especially sharp. First, the effect of feedback differs radically by process class. For partial-feedback multiset guessing, the maximal expected payoff is $S_i$37, and the proven explicit bound is $S_i$38 when $S_i$39 (Nie, 2022). By contrast, in the complete-feedback one-shelf shuffle, the optimal and unique strategy achieves exactly $S_i$40 expected correct guesses (Clay, 14 Jul 2025). This suggests that the information structure, not only the state space, determines the attainable improvement over a naive baseline.

Second, “more slots” is not uniformly beneficial. AdaSlot introduces a complexity regularizer $S_i$41 precisely to discourage unnecessary latent slots, and its reported performance matches or exceeds top fixed-slot models on MOVi-C and MOVi-E while trailing the best fixed-6-slot DINOSAUR on COCO for several reported metrics (Fan et al., 2024). In DST, the MSP results similarly indicate that richer slot-specific memory can improve JGA and handling of indirect mentions, but only when coupled to an explicit gating decision that can reject wrong inheritance (Sun et al., 2022).

Third, multiple open problems remain. The multiset deck paper conjectures that the $S_i$42 upper bound should hold for all $S_i$43, including $S_i$44, but notes that new ideas are required in that regime. The one-shelf shuffle paper proves optimality for the complete-feedback strategy and special no-feedback positions, while a general no-feedback strategy remains conjectural. AdaSlot identifies the mean-field assumption over slot-selection variables and the absence of part–whole supervision as current limitations. The DST formulation points to failures in deep semantic disambiguation when multiple candidate values occur in the same dialogue context.

A common misconception is that “slot guessing” denotes a single technical problem. The literature instead uses the term for several families of inference problems: sequential guessing on shuffled structures, latent-slot selection in object-centric representation learning, and slot-value updating in sequential language understanding. The shared vocabulary is real, but the operative mathematics ranges from coupling and supermartingales to Markov-chain analysis, differentiable discrete sampling, and bilinear memory-attention architectures.