Stable Slot Allocation: Methods & Applications

• Stable slot allocation is a family of assignment models that match finite slots to entities under feasibility, capacity, and exclusivity constraints.
• It employs mechanisms such as identity-slot alignment in 3D generation, orthogonal scheduling in communication systems, and lattice-based approaches in bipartite matching.
• Applications span diverse domains—from enhancing part-aware 3D synthesis and cognitive radio scheduling to optimizing online ad delivery—offering quantifiable performance improvements.

Stable slot allocation is a family of constrained assignment problems in which a finite set of slots, bands, seats, or allocation units must be matched to entities under feasibility constraints together with a domain-specific notion of stability. In the cited arXiv literature, the relevant instability takes different forms: identity–slot permutation freedom in part-aware 3D generation, queue instability and collisions in slotted communication systems, blocking edges in capacitated bipartite allocations, blocking contracts in matching with contracts, and redundancy or concentration across multiple advertising slots. The common structural pattern is that stable allocation requires an explicit admissible assignment model, capacity or exclusivity constraints, and a mechanism that rules out inconsistent or locally improvable reallocations.

1. Core meanings of stability across domains

The phrase is not tied to a single mathematical definition. In the current literature, “stable” may refer to semantic consistency of slot identities, stochastic queue stability, the absence of blocking edges or contracts, or robustness of long-term delivery under operational constraints.

Setting	Slot object	Stability criterion
Part-aware 3D generation (Hao et al., 10 Jun 2026)	generation slots / learnable queries	break identity–slot permutation freedom by anchoring each semantic identity to exactly one slot
Cognitive radio and slotted access (Shafie et al., 2014, 0809.5023)	licensed bands or transmission opportunities	queues remain stable when long-term arrival rates stay below service rates
Capacitated matching and allocation (Cseh et al., 2012, Karzanov, 2024, Cseh et al., 2014)	edges, seats, or machine assignments	no blocking edge; in some models the stable set forms a lattice
Matching with contracts (Avataneo et al., 2020)	original and shadow slots	COM yields a unique stable matching under SSPwCT
Multi-slot advertising (Zhang et al., 20 May 2026)	page-view ad slots	slot exclusivity and Page View constraints prevent duplication and concentration

This distribution of meanings suggests that “stable slot allocation” is best understood as a class of formally specified allocation regimes rather than a single theory. A common misconception is to treat the stability notion as uniform across these settings; the papers instead define stability relative to the operative failure mode of each domain.

2. Identity-slot alignment in part-aware 3D generation

In "ISAP-3D: Identity-Slot Aligned Part-Aware 3D Generation" (Hao et al., 10 Jun 2026), stable slot allocation is introduced to address structural ambiguity in part-aware 3D generation. The paper attributes instability to identity–slot permutation freedom. Let $I=\{ID_1,\dots,ID_K\}$ denote semantic identities and $S=\{Slot_1,\dots,Slot_K\}$ denote generation slots. The model must learn a bijection $\pi:I\leftrightarrow S$, but if supervision only requires that some $K$ boxes and some $K$ geometries match the ground-truth parts, then any permutation produces the same loss. The paper identifies two resulting inference failures: slot swapping and part merging or collapse.

ISAP-3D resolves this by enforcing identity-aligned one-to-one slot modelling in three stages. First, each semantic part $k$ is associated with a text prompt $T_k$, encoded by a text encoder such as BERT or CLIP-Text into an identity token $y_k=\mathrm{TextEnc}(T_k)\in\mathbb{R}^d$. Slot queries are initialized as $slot_i^{(0)}=W_0\cdot y_i$, so each slot is explicitly conditioned on a unique identity token. Second, the model performs identity-conditioned layout prediction. Local condition $LC_k$ combines $S=\{Slot_1,\dots,Slot_K\}$0 with any mask features, while global context tokens $S=\{Slot_1,\dots,Slot_K\}$1 encode the reference image and a coarse voxel prior. Slot states are updated by a transformer stack,

$S=\{Slot_1,\dots,Slot_K\}$2

and decoded into axis-aligned boxes $S=\{Slot_1,\dots,Slot_K\}$3. Third, geometry synthesis is conditioned on the predicted layout by cropping the coarse voxel prior into layout voxels $S=\{Slot_1,\dots,Slot_K\}$4, encoding them to features $S=\{Slot_1,\dots,Slot_K\}$5, and applying a geometry flow transformer with local attention conditioned on $S=\{Slot_1,\dots,Slot_K\}$6 and global attention over all parts and $S=\{Slot_1,\dots,Slot_K\}$7. Training uses a layout loss with classification and regression terms, followed by a conditional flow-matching loss plus Chamfer distance.

The framework is supported by a part-level dataset with a unified semantic protocol. The construction starts from PartVerseXL ($S=\{Slot_1,\dots,Slot_K\}$8K shapes), defines a closed vocabulary of $S=\{Slot_1,\dots,Slot_K\}$9 part tokens per category, forces unique mappings by adding directional or cardinal modifiers such as “left_hand” and “right_hand,” uses a vision–LLM constrained to this vocabulary for initial labels, and then manually refines them. The final split contains 8K shapes for training and 100 held out for evaluation.

On the held-out 100 shapes, ISAP-3D reports $\pi:I\leftrightarrow S$0, $\pi:I\leftrightarrow S$1, $\pi:I\leftrightarrow S$2, and $\pi:I\leftrightarrow S$3, compared with OmniPart’s $\pi:I\leftrightarrow S$4, $\pi:I\leftrightarrow S$5, $\pi:I\leftrightarrow S$6, and $\pi:I\leftrightarrow S$7. The paper highlights a $\pi:I\leftrightarrow S$8 NMI gain over OmniPart and states that $\pi:I\leftrightarrow S$9 corresponds to near-perfect consistency of slot IDs across multiple views of the same shape. In this setting, stable slot allocation means preserving one-to-one semantic identity throughout semantic, spatial, and geometric stages.

3. Queueing-theoretic and communication-system formulations

In cognitive radio networks, stable slot allocation appears as orthogonal band allocation. El Shafie and Khattab study $K$0 buffered secondary users and $K$1 primary bands in slotted time, with the requirement that each secondary user is assigned to exactly one licensed band and each band hosts at most one secondary user; neither band sharing nor multi-band allocations are permitted (Shafie et al., 2014). Primary and secondary arrivals are independent Bernoulli processes under the late-arrival model. Stability is defined by

$K$2

and, by Loynes’ theorem, a queue is stable exactly when its long-term arrival rate is strictly below its long-term service rate.

Let $K$3 denote the fraction of slots in which secondary user $K$4 is assigned to band $K$5. The orthogonality constraints are

$K$6

With $K$7, the secondary service rate is

$K$8

and each secondary queue is stable iff $K$9. The stability-region envelope is a convex polyhedron and can be traced by a standard linear program in the $K$0 variables $K$1. Once an optimal $K$2 is obtained, Birkhoff’s theorem yields a decomposition into randomized permutation schedules, so that each slot implements a collision-free matching. The paper proves that the stability region of orthogonal allocation contains those of both random band selection and fixed deterministic assignment.

A different communication meaning appears in slotted-Aloha, where all users contend for a shared channel. Bordenave, McDonald and Proutiere define the stability region $K$3 as the set of arrival-rate vectors for which the global Markov chain is positive recurrent (0809.5023). Exact results are given for $K$4 and $K$5, while for $K$6 the exact region is generally unknown in closed form. The paper introduces an approximate region $K$7 based on an independence ansatz and proves asymptotic exactness: for every $K$8, sufficiently large $K$9 implies that $k$0 guarantees stability and $k$1 guarantees instability. Mean-field analysis yields fixed-point equations

$k$2

which characterize global stability of the limiting nonlinear dynamical system.

Taken together, these papers use “stable” in the queueing sense. The central object is not absence of blocking pairs but persistence of bounded queues under stochastic arrivals and service.

4. Capacitated bipartite allocation, local dynamics, and lattice structure

In the stable allocation problem of Baïou–Balinski type, summarized in "Paths to stable allocations" (Cseh et al., 2012), one considers a bipartite graph $k$3 with vertex quotas $k$4 and edge capacities $k$5. An allocation $k$6 must satisfy $k$7 and $k$8. Each vertex ranks incident edges strictly. An edge $k$9 blocks a feasible allocation $T_k$0 if it is not saturated, job $T_k$1 has free quota or prefers $T_k$2 to its worst positive-$T_k$3 edge, and machine $T_k$4 has free quota or prefers $T_k$5 to its worst positive-$T_k$6 edge. Stability means that no edge blocks.

The paper studies uncoordinated dynamics. In a better-response step, an agent chooses any blocking edge and increases allocation on it while refusing worse edges as needed. In a best-response step, a job chooses its best blocking edge and redistributes as much allocation as possible. For rational input, Theorem A and Theorem B show that there exist finite sequences of better- and best-response steps leading to a stable allocation, and random dynamics converge with probability 1. Theorem C states that both better-response and best-response paths may be exponentially long in $T_k$7 in the worst case. Theorem D gives a strongly polynomial $T_k$8 accelerated two-phase better-response algorithm for arbitrary real input. Theorem E shows that in correlated markets, where a global order $T_k$9 induces every local preference order, the stable allocation is unique and random best-response reaches it in expected time $y_k=\mathrm{TextEnc}(T_k)\in\mathbb{R}^d$0.

Karzanov’s mixed model extends this line by replacing strict preferences with weak orders and using choice functions of mixed type (Karzanov, 2024). Here $y_k=\mathrm{TextEnc}(T_k)\in\mathbb{R}^d$1 is bipartite, capacities and quotas are integral, and each vertex has a weak order on incident edges. The choice function $y_k=\mathrm{TextEnc}(T_k)\in\mathbb{R}^d$2 keeps all allocation from ties strictly above a critical tie, cuts within the critical tie to fill exactly $y_k=\mathrm{TextEnc}(T_k)\in\mathbb{R}^d$3, and rejects strictly worse ties. An assignment is fully stationary if $y_k=\mathrm{TextEnc}(T_k)\in\mathbb{R}^d$4 for all vertices. In this model, an edge $y_k=\mathrm{TextEnc}(T_k)\in\mathbb{R}^d$5 blocks $y_k=\mathrm{TextEnc}(T_k)\in\mathbb{R}^d$6 iff $y_k=\mathrm{TextEnc}(T_k)\in\mathbb{R}^d$7 and $y_k=\mathrm{TextEnc}(T_k)\in\mathbb{R}^d$8, where $y_k=\mathrm{TextEnc}(T_k)\in\mathbb{R}^d$9 is the tail no better than the critical tie. The set of stable assignments is nonempty and, under the firms’ preference order, forms a distributive lattice.

The mixed model also develops a rotation theory. From a stable assignment $slot_i^{(0)}=W_0\cdot y_i$0, one constructs an active graph $slot_i^{(0)}=W_0\cdot y_i$1; maximal strongly connected components define rotations $slot_i^{(0)}=W_0\cdot y_i$2, and every stable assignment can be represented as

$slot_i^{(0)}=W_0\cdot y_i$3

where $slot_i^{(0)}=W_0\cdot y_i$4 is the $slot_i^{(0)}=W_0\cdot y_i$5-optimal stable assignment, $slot_i^{(0)}=W_0\cdot y_i$6 is the matrix of rotation columns, and $slot_i^{(0)}=W_0\cdot y_i$7 is a closed function on the rotation poset. This gives a compact affine representation of the lattice of stable assignments and reduces minimum-cost stable assignment to an ideal-minimization problem solvable by a single $slot_i^{(0)}=W_0\cdot y_i$8-$slot_i^{(0)}=W_0\cdot y_i$9 cut in $LC_k$0 time.

5. Slot-specific priorities and capacity transfers

"Slot-specific Priorities with Capacity Transfers" formalizes stable slot allocation in matching with contracts by endowing each branch with original slots and shadow slots (Avataneo et al., 2020). Agents $LC_k$1, branches $LC_k$2, and contracts $LC_k$3 define the market. Each branch $LC_k$4 has physical capacity $LC_k$5, original slots $LC_k$6, shadow slots $LC_k$7, slot-specific priorities $LC_k$8, precedence orders over originals and shadows, a location vector $LC_k$9 interleaving originals and shadows in the fill order, and a capacity-transfer vector $S=\{Slot_1,\dots,Slot_K\}$00. If original slot $S=\{Slot_1,\dots,Slot_K\}$01 remains empty, then its associated shadow slot $S=\{Slot_1,\dots,Slot_K\}$02 becomes active exactly when $S=\{Slot_1,\dots,Slot_K\}$03.

For an offer set $S=\{Slot_1,\dots,Slot_K\}$04, the branch choice rule $S=\{Slot_1,\dots,Slot_K\}$05 processes the renumbered fill sequence $S=\{Slot_1,\dots,Slot_K\}$06 slot by slot. Original slots choose the highest-priority feasible contract not already assigned to another slot for the same agent; shadow slots are active only if their associated original slot is empty and transferability allows activation. The final branch choice is the set of all non-$S=\{Slot_1,\dots,Slot_K\}$07 contracts chosen after $S=\{Slot_1,\dots,Slot_K\}$08 steps.

The market is cleared by the Cumulative Offer Mechanism (COM). At each round, an unmatched agent proposes its favorite unrejected contract; the relevant branch recomputes its held set by applying its choice rule to its current holds plus the new proposal. Stability requires individual rationality and the absence of a blocking branch-plus-contract set. Under SSPwCT choice rules, Theorem 1 states that the COM outcome is the unique stable matching, is strategy-proof for all agents, and follows from a substitutable, LAD- and IRC-satisfying completion argument together with Hatfield–Milgrom or Hatfield–Kominers results.

The framework also has strong comparative statics. If $S=\{Slot_1,\dots,Slot_K\}$09 coordinate-wise and strictly exceeds $S=\{Slot_1,\dots,Slot_K\}$10 in at least one coordinate, then the COM outcome is a weak Pareto improvement for all agents. Adding one more original slot and its shadow slot likewise yields an outcome that weakly Pareto-dominates the original one. The paper further states that adding a batch of new bottom-ranked contracts or adding a new contract for a single agent weakly improves the COM outcome for every agent. In applications such as school choice or affirmative action, this means that reserved seats and spillover capacity can be represented directly at the slot level while preserving stability and strategy-proofness.

6. Unsplittable assignments and multi-slot operational optimization

In the unsplittable stable allocation problem, jobs have sizes $S=\{Slot_1,\dots,Slot_K\}$11, machines have capacities $S=\{Slot_1,\dots,Slot_K\}$12, and each job must be assigned in its entirety to a single machine or remain unassigned. The assignment variable satisfies $S=\{Slot_1,\dots,Slot_K\}$13 (Cseh et al., 2014). Under the paper’s stricter blocking rule, an edge $S=\{Slot_1,\dots,Slot_K\}$14 blocks if $S=\{Slot_1,\dots,Slot_K\}$15 prefers $S=\{Slot_1,\dots,Slot_K\}$16 to its current partner and machine $S=\{Slot_1,\dots,Slot_K\}$17 is either under capacity or prefers $S=\{Slot_1,\dots,Slot_K\}$18 to at least one currently assigned job. A reversed Gale–Shapley algorithm in which machines propose computes a machine-optimal relaxed unsplit stable allocation in $S=\{Slot_1,\dots,Slot_K\}$19 time. If a fully feasible unsplittable stable assignment exists, it coincides with the machine-optimal solution. When no feasible solution exists, the machine-optimal relaxed stable allocation minimizes total congestion

$S=\{Slot_1,\dots,Slot_K\}$20

The same paper gives a rounding method that transforms any fractional stable allocation into a relaxed unsplit stable assignment by at most $S=\{Slot_1,\dots,Slot_K\}$21 augmentations, while ensuring that each machine exceeds capacity by at most one job-sized chunk.

A distinct large-scale operational use appears in guaranteed display advertising. "Beyond Single Slot: Joint Optimization for Multi-Slot Guaranteed Display Advertising" formulates multi-slot allocation as an offline bipartite matching between supply nodes $S=\{Slot_1,\dots,Slot_K\}$22 and contract nodes $S=\{Slot_1,\dots,Slot_K\}$23, with decision variables $S=\{Slot_1,\dots,Slot_K\}$24, contract demand constraints, one-allocation-per-slot constraints, and per-page-view exposure constraints $S=\{Slot_1,\dots,Slot_K\}$25 (Zhang et al., 20 May 2026). The objective combines a quadratic smoothness term penalizing deviation from the normalized target delivery ratio $S=\{Slot_1,\dots,Slot_K\}$26, a linear priority-reward term $S=\{Slot_1,\dots,Slot_K\}$27, and a linear interest-reward term $S=\{Slot_1,\dots,Slot_K\}$28.

The online realization adds a contract roulette mechanism. For each slot, candidates are roulette-sampled with probability proportional to a delivery-urgency weight; low-score candidates are filtered by $S=\{Slot_1,\dots,Slot_K\}$29; if a contract appears in multiple slot candidate sets, it is kept only in the slot where it ranks best; and adaptive bidword control reuses or resamples bidwords to preserve contextual consistency and diversity. The optimization algorithm derives a KKT closed form,

$S=\{Slot_1,\dots,Slot_K\}$30

and updates dual variables by projected gradients. Each iteration costs $S=\{Slot_1,\dots,Slot_K\}$31 work, can be fully parallelized across edges, and, according to the paper, convergence in a few tens of iterations suffices in practice.

The online evaluation reports two kinds of numerical evidence. The abstract states a $S=\{Slot_1,\dots,Slot_K\}$32 increase in Average Revenue Per User under $S=\{Slot_1,\dots,Slot_K\}$33 traffic. The detailed summary gives Difference-in-Differences effects under $S=\{Slot_1,\dots,Slot_K\}$34 gray scale of Merchant ROI $S=\{Slot_1,\dots,Slot_K\}$35, Payment ROI $S=\{Slot_1,\dots,Slot_K\}$36, CTR $S=\{Slot_1,\dots,Slot_K\}$37, Payment CVR $S=\{Slot_1,\dots,Slot_K\}$38, ARPU $S=\{Slot_1,\dots,Slot_K\}$39, and Fulfillment $S=\{Slot_1,\dots,Slot_K\}$40. In this setting, stability is operational rather than lattice-theoretic: Page View constraints and contract roulette prevent within-page duplication and head-slot monopolization, while the global matching keeps each contract close to its demand target.

Across these formulations, stable slot allocation is not a single theorem but a recurring design principle. The literature shows that stability can be enforced by identity anchoring, orthogonal scheduling, blocking-edge elimination, substitutable slot choice rules, lattice representations of stable assignments, or large-scale constrained optimization. This suggests that the unifying question is not what a slot is, but which formal mechanism makes its allocation persistent, nonconflicting, and robust under the constraints of the domain.