Quadratic Ranking Methods

• Quadratic Ranking is a family of methods defined by quadratic interactions in computation, objectives, or constraints across various applications such as self-attention and recommendation systems.
• It includes strategies like least-squares ranking on graphs and pairwise prompting, where quadratic costs arise from exhaustive comparisons or dense attention maps.
• Structured reductions like BlockRank, active sampling, and QCQP relaxations are employed to mitigate the quadratic computational burden while maintaining ranking accuracy.

Searching arXiv for the cited papers to ground the article in current records. arXiv search: "Scalable In-context Ranking with Generative Models" Quadratic Ranking denotes a family of ranking formulations in which the dominant structure is quadratic in either computation, objective design, sample complexity, or constraint geometry. In contemporary usage, the expression covers dense self-attention over long in-context ranking prompts with $O(L^2)$ cost (Gupta et al., 6 Oct 2025), pairwise ranking procedures that enumerate $N^+N^-$ or $n(n-1)$ comparisons (Shen et al., 2017, Wu et al., 7 Jul 2025), least-squares ranking on graphs with a weighted quadratic objective (Hirani et al., 2010), online rank aggregation under Spearman squared distance (Ailon, 2013), multi-slot recommendation as a quadratically constrained quadratic program (Basu et al., 2016), and a recent matching algorithm explicitly named Quadratic Ranking because its key expressions are quadratic forms in two parameter functions (Huang et al., 25 Jul 2025). In the cited literature, the term is therefore not a single standardized method but a cluster of technically distinct constructions.

1. Terminological scope

Across the literature, “quadratic” refers to different objects: the number of interactions that must be considered, the algebraic form of the loss, or the optimization class induced by interactions and constraints.

Domain	Quadratic object	Representative paper
In-context IR with LLMs	Dense self-attention over prompt length	(Gupta et al., 6 Oct 2025)
Bipartite ranking and PRP	Enumeration of positive–negative or document pairs	(Shen et al., 2017, Wu et al., 7 Jul 2025)
Graph ranking	Weighted least-squares objective	(Hirani et al., 2010)
Online rank aggregation	Spearman squared distance	(1308.07336)
Recommendation	QCQP with interaction matrices	(Basu et al., 2016)
Matching	Quadratic forms in priority and gain functions	(Huang et al., 25 Jul 2025)
Evolutionary selection	Degree-2 polynomial ranking	(0803.2925)

This multiplicity matters because the source of the quadratic phenomenon determines the remedy. Dense attention invites structured sparsity; pairwise enumeration invites sampling or distillation; least-squares formulations invite Laplacian solvers; QCQPs invite convex relaxation; and quadratic rank losses invite specialized differentiable operators or online-regret analyses.

2. Computational quadraticity in modern ranking systems

In in-context ranking (ICR), a prompt concatenates instructions, the query, and a shortlist of candidate document blocks. Standard self-attention in an LLM layer computes

$A=\operatorname{softmax}(QK^\top/\sqrt d)V,$

with per-head compute dominated by $O(L^2 d)$ terms and attention-memory $O(L^2)$, where $L$ is total context length. Because $L=|Inst|+|Q|+\sum_i |B_i|$, and each additional document contributes tokens, the attention cost becomes $O(N^2)$ in the number of documents for fixed per-document size (Gupta et al., 6 Oct 2025). That paper identifies two empirical regularities in a Mistral-7B-v0.3 model fine-tuned for ICR: inter-document block sparsity, and query-document block relevance concentrated in middle layers. These observations motivate BlockRank, which masks cross-document interactions except through the query and reduces per-layer attention from standard $O(N^2 L_{\text{chunk}}^2 d)$ to $N^+N^-$0 under fixed chunking assumptions (Gupta et al., 6 Oct 2025).

Pairwise ranking prompting with LLMs exhibits a different quadratic source. For a query $N^+N^-$1 and candidate set $N^+N^-$2, Pairwise Ranking Prompting evaluates ordered pairs $N^+N^-$3 with $N^+N^-$4, so the number of LLM calls is

$N^+N^-$5

The same paper defines teacher probabilities $N^+N^-$6, symmetrizes them to reduce order sensitivity, and shows that the outstanding ranking performance of PRP can be transferred to a pointwise student through Pairwise Ranking Distillation, with teacher labeling reduced to about $N^+N^-$7 calls at $N^+N^-$8 when the sampling ratio is $N^+N^-$9 (Wu et al., 7 Jul 2025).

Classical bipartite ranking has the same combinatorial growth. If there are $n(n-1)$0 positives and $n(n-1)$1 negatives, the pair set

$n(n-1)$2

has size $n(n-1)$3 when class proportions are not extreme. RankSVM-style objectives therefore accumulate a quadratic number of difference vectors $n(n-1)$4, which is prohibitive on large datasets (Shen et al., 2017). A closely related annotation problem appears in human pairwise labeling: exhaustive comparisons over $n(n-1)$5 items require $n(n-1)$6, which motivates uncertainty-aware automation and active pair selection in Dodgersort (Park et al., 21 Mar 2026).

Quadratic behavior can also arise in sample complexity rather than runtime. In ranking from discrete ratings with unknown user thresholds, items are partitioned into ordered bins, and achieving constant expected Maximum Spearman Footrule requires $n(n-1)$7 users and therefore $n(n-1)$8 queries. The same paper proves that the expected number of users needed to achieve a total order is infinite, because very small consecutive score gaps induce heavy-tailed waiting times for informative thresholds (Villemaud et al., 2 Oct 2025).

3. Quadratic objectives and algebraic formulations

A canonical quadratic ranking formulation is least-squares ranking on graphs. Given an undirected comparison graph $n(n-1)$9, oriented incidence matrix $A=\operatorname{softmax}(QK^\top/\sqrt d)V,$0, edge weights $A=\operatorname{softmax}(QK^\top/\sqrt d)V,$1, and observed pairwise differences $A=\operatorname{softmax}(QK^\top/\sqrt d)V,$2, the score vector $A=\operatorname{softmax}(QK^\top/\sqrt d)V,$3 is obtained from

$A=\operatorname{softmax}(QK^\top/\sqrt d)V,$4

The normal equations are

$A=\operatorname{softmax}(QK^\top/\sqrt d)V,$5

where $A=\operatorname{softmax}(QK^\top/\sqrt d)V,$6 is the weighted graph Laplacian (Hirani et al., 2010). Because $A=\operatorname{softmax}(QK^\top/\sqrt d)V,$7 is singular with nullspace $A=\operatorname{softmax}(QK^\top/\sqrt d)V,$8 on a connected graph, a gauge must be fixed or the Moore–Penrose pseudoinverse $A=\operatorname{softmax}(QK^\top/\sqrt d)V,$9 used. The residual $O(L^2 d)$0 admits a discrete Hodge decomposition

$O(L^2 d)$1

which separates globally consistent gradient structure from curl on triangles and harmonic inconsistency on larger cycles (Hirani et al., 2010). This is the key reason the paper connects ranking to spectral graph theory, multilevel methods for graph Laplacian systems, Hodge decomposition, and arbitrage detection.

In bipartite ranking, the widely adopted pairwise squared surrogate is also quadratic, but in the parameter vector. With a linear scorer $O(L^2 d)$2 and pairwise feature $O(L^2 d)$3, the loss is

$O(L^2 d)$4

Summing over all positive–negative pairs yields an empirical risk of the form

$O(L^2 d)$5

so the all-pairs ERM is a quadratic program over a norm ball (Gultekin et al., 2019). The Low Cost Bipartite Ranking method replaces the all-pairs statistics $O(L^2 d)$6 with subsampled $O(L^2 d)$7, obtaining a stochastic, gradient-free estimator whose required subsample size $O(L^2 d)$8 does not scale like $O(L^2 d)$9 (Gultekin et al., 2019).

Quadratic ranking losses also appear in rank aggregation. For two rankings $O(L^2)$0 and $O(L^2)$1, Spearman squared distance is

$O(L^2)$2

In online ranking with full permutation feedback, minimizing the linear dot-product loss $O(L^2)$3 is exactly equivalent to minimizing Spearman squared distance up to an additive constant, yielding an expected regret bound $O(L^2)$4 under the paper’s framework (1308.07336). The same metric motivates differentiable rank operators based on projections onto the permutahedron, where soft ranks are defined by

$O(L^2)$5

and exact differentiation is obtained through a reduction to isotonic optimization with $O(L^2)$6 forward time and $O(L^2)$7 space (Blondel et al., 2020).

4. Structured reductions of the quadratic burden

Because quadratic growth is often the central obstacle, much of the recent literature is devoted to preserving ranking quality while removing dense interactions.

BlockRank is the most explicit architectural intervention in LLM ranking. It partitions the context into instructions, query, and document blocks, then enforces a mask in which document tokens attend only to their own block and instructions, while query tokens retain global visibility. The same model adds an auxiliary InfoNCE-style loss at a middle layer $O(L^2)$8 with temperature $O(L^2)$9 and coefficient $L$0, training query-to-document attention to serve as a relevance signal. On BEIR, MSMarco, and NQ with Mistral-7B, the paper reports BEIR zero-shot reranking average nDCG@10 of $L$1, MSMarco Precision@1 of $L$2 and MRR@10 of $L$3, NQ Precision@1 of $L$4, and a $L$5 speed-up at $L$6 with latency about $L$7 at $L$8 and roughly $L$9K tokens (Gupta et al., 6 Oct 2025).

In pairwise bipartite ranking, active sampling and the Combined Ranking and Classification framework reduce the effective training set from $L=|Inst|+|Q|+\sum_i |B_i|$0 to $L=|Inst|+|Q|+\sum_i |B_i|$1. The sampling rules favor either margin closeness,

$L=|Inst|+|Q|+\sum_i |B_i|$2

or correctness,

$L=|Inst|+|Q|+\sum_i |B_i|$3

with inverse-probability correction in the loss (Shen et al., 2017). The paper reports that, with budget $L=|Inst|+|Q|+\sum_i |B_i|$4, Active Sampling within CRC coupled with a linear SVM usually outperforms state-of-the-art point-wise and pair-wise ranking approaches in both accuracy and efficiency (Shen et al., 2017).

Pairwise Ranking Distillation attacks the same bottleneck at teacher-labeling time. Its central empirical claim is that with only $L=|Inst|+|Q|+\sum_i |B_i|$5 of pairs, the distilled pointwise student obtains the same performance as using all pairs for teacher labels (Wu et al., 7 Jul 2025). Dodgersort applies a human-in-the-loop variant of the same principle: CLIP-based hierarchical pre-ordering prunes trivial comparisons, a neural ranking head and an Elo–BTL–GP ensemble estimate probabilities and uncertainty, and information-theoretic pair selection reduces human comparisons by $L=|Inst|+|Q|+\sum_i |B_i|$6–$L=|Inst|+|Q|+\sum_i |B_i|$7 while improving inter-rater reliability (Park et al., 21 Mar 2026).

A different linearization strategy appears in uncertain databases. Probabilistic top-$L=|Inst|+|Q|+\sum_i |B_i|$8 similarity ranking traditionally computes instance-rank distributions by dynamic programming with $L=|Inst|+|Q|+\sum_i |B_i|$9 time. By incrementally accessing instances in increasing distance order and maintaining the Active Object List, the proposed framework reduces this to $O(N^2)$0, or $O(N^2)$1 when instances are already ordered, with the same $O(N^2)$2 space and the same exact rank distributions (0907.2868).

5. QCQP-based ranking, matching, and evolutionary selection

In recommendation, quadratic ranking arises when interactions across display slots are modeled directly. The multi-slot optimization problem uses decision variables $O(N^2)$3 for placing item $O(N^2)$4 in slot $O(N^2)$5 for user $O(N^2)$6, with engagement and cost vectors induced by

$O(N^2)$7

The resulting convex formulation is

$O(N^2)$8

which is a QCQP (Basu et al., 2016). Because generic SDP and RLT relaxations are not scalable, the paper proposes a minimal relaxation: replace the ellipsoidal quadratic constraint with tangent hyperplanes sampled from an equidistributed cover of the boundary, yielding a QP whose solution converges to the QCQP optimum as the number of tangent planes grows (Basu et al., 2016).

In oblivious bipartite matching, “Quadratic Ranking” is the name of a specific randomized primal–dual algorithm. Each vertex receives a random rank $O(N^2)$9, edges are prioritized by the perturbed weight

$O(N^2 L_{\text{chunk}}^2 d)$0

and dual gains satisfy

$O(N^2 L_{\text{chunk}}^2 d)$1

The functions $O(N^2 L_{\text{chunk}}^2 d)$2 and $O(N^2 L_{\text{chunk}}^2 d)$3 are constrained by $O(N^2 L_{\text{chunk}}^2 d)$4 for all $O(N^2 L_{\text{chunk}}^2 d)$5. This quadratic-form design yields a polynomial-time $O(N^2 L_{\text{chunk}}^2 d)$6-competitive algorithm for Oblivious Bipartite Matching and, by robustness, for Query-Commit Bipartite Matching even with correlated edge realizations, thereby breaking the $O(N^2 L_{\text{chunk}}^2 d)$7 barrier posed as an open question by Tang, Wu, and Zhang (Huang et al., 25 Jul 2025).

In evolutionary computation, quadratic ranking means polynomial rank selection of degree $O(N^2 L_{\text{chunk}}^2 d)$8:

$O(N^2 L_{\text{chunk}}^2 d)$9

for rank $N^+N^-$00. Hingee and Hutter show that every probabilistic tournament is equivalent to a unique polynomial rank scheme, and that most linear and most practical quadratic rank schemes are probabilistic tournaments (0803.2925). This establishes an exact bridge between degree-2 rank-based selection and size-$N^+N^-$01 tournament selection.

6. Regimes, limitations, and recurring misconceptions

A recurring misconception is that quadratic ranking always means pairwise ranking. The cited literature contradicts this. In some settings the quadratic object is the number of comparisons; in others it is an $N^+N^-$02 attention map, a least-squares residual, a squared rank distance, or a quadratically constrained feasible region. This suggests that comparisons between methods are meaningful only within a shared quadratic regime.

The limitations are correspondingly regime-specific. BlockRank is most beneficial for large shortlists in which relevance is largely determined by query-to-document interactions, but tasks requiring deep cross-document fusion among document tokens are constrained by blocked document-to-document attention (Gupta et al., 6 Oct 2025). Active sampling in bipartite ranking can oversample noisy pairs under soft-correct selection at large budgets (Shen et al., 2017). Ranking from discrete ratings has a fundamental lower bound: the quadratic cost is tied to unknown thresholds, and the near-perfect regime requires $N^+N^-$03 ratings even under universal agreement and noiseless feedback (Villemaud et al., 2 Oct 2025). In Dodgersort, the GP preference learner is disabled beyond $N^+N^-$04 because of $N^+N^-$05 cost (Park et al., 21 Mar 2026). In multi-slot QCQP ranking, convexity depends on $N^+N^-$06 and $N^+N^-$07, and approximation quality depends on the tangent-plane cover (Basu et al., 2016).

The long-term significance of the topic lies in this common pattern: ranking systems repeatedly expose quadratic structure because ranking is relational. Documents interact through attention, candidates interact through pairwise supervision, alternatives interact through graph edges, slots interact through exposure externalities, and rankings interact through squared permutation distances. The most successful lines of work do not remove that structure conceptually; they exploit sparsity, decomposition, sampling, or convex reformulation so that the quadratic object remains analyzable without remaining the computational bottleneck (Gupta et al., 6 Oct 2025, Hirani et al., 2010, Gultekin et al., 2019).