PolyFact: Multifaceted Factorization Research

• PolyFact is a polysemous term defining diverse factorization frameworks in NLP, numerical algebra, and machine learning.
• It encompasses a multilingual factual QA benchmark using reinforcement learning to improve cross-lingual consistency and factual recall in language models.
• Other approaches include numerical, symbolic, and convex-dense algorithms that leverage geometric and algebraic innovations for efficient polynomial and matrix factorization.

PolyFact is a polysemous term in contemporary research usage rather than the name of a single unified theory or system. In recent arXiv literature, it denotes at least three major families of work: a multilingual factual question-answering dataset and reinforcement-learning framework for cross-lingual factual recall in LLMs; a geometric-optimization formulation of numerical polynomial factorization; and several factorization-oriented methods in symbolic computation, algebraic complexity, convex-dense bivariate factorization, multilinear factorization over $\mathbb{F}_2$, multi-output polynomial modeling, and polytopic matrix factorization (Rad et al., 4 Jun 2026, Wu et al., 2021). The common lexical element is therefore not a shared technical core, but the reuse of “PolyFact” as shorthand for “polynomial factorization,” “polytopic factorization,” or a named factual-recall benchmark.

1. PolyFact as a multilingual factual-recall benchmark and training framework

In natural language processing, PolyFact refers to a fully parallel multilingual multiple-choice factual QA dataset together with a consistency-driven reinforcement-learning recipe for improving cross-lingual factual recall in LLMs. The dataset contains 100,000 Wikidata-grounded facts, each rendered as aligned MCQs across 12 typologically diverse languages: English, German, Spanish, French, Portuguese, Indonesian, Russian, Arabic, Chinese, Japanese, Swahili, and Bengali. The split is 95,000 training facts, 2,500 validation facts, and 2,500 test facts, with fact-level partitioning so that all 12 language realizations of one fact remain in the same split (Rad et al., 4 Jun 2026).

The motivating problem is cross-lingual factual inconsistency: a model may encode a fact but fail to express it consistently across languages. The work defines per-language factual accuracy as

$$Acc_\ell = \frac{1}{N}\sum_{i=1}^N \mathbf{1}[\hat y_{i,\ell}=y_i],$$

and identifies a bundle-level consistency rate aligned with training:

$$Cons_{\mathrm{all}} = \frac{1}{N}\sum_{i=1}^N \mathbf{1}[\forall \ell,\hat y_{i,\ell}=y_i].$$

PolyFact operationalizes the problem through strict parallelism, four-option MCQs, and Wikidata grounding via subject, property, and object IDs, relation metadata, per-language questions and options, and verification flags (Rad et al., 4 Jun 2026).

The dataset construction pipeline begins from the Wikidata truthy-triples dump and selects triples whose property lies in a curated set of 22 relations spanning geography, biography, creative works, media, and organizational or cultural ties. Labels are extracted across the 12 target languages, with missing labels backfilled via batched wbgetentities. Distractors are sampled from other objects of the same property, constrained by shared instance of (P31) type, length-matched English labels within a tolerance, and distinctness under case/whitespace normalization. A fact is retained only if question generation succeeds in all 12 languages, enforcing an all-or-nothing parallel bundle criterion (Rad et al., 4 Jun 2026).

Quality control is explicit. An LLM-as-judge labeled 3,600 items, while independent human review labeled 600 items across six languages. Inter-judge agreement is reported as 91.0% overall, with language-dependent ambiguity concentrated in “country of origin” (35%), “place of birth” (25%), and “genre” (25%). A recommended “PolyFact-Clean” filter excludes these higher-ambiguity relations, removing about 17% of items (Rad et al., 4 Jun 2026).

2. Reinforcement learning on PolyFact and mechanistic findings

The associated training methodology compares continual pretraining, supervised finetuning, and Group Relative Policy Optimization. In this setting, GRPO uses grouped on-policy rollouts over parallel bundles: for each fact, $G=8$ rollouts are sampled, and each rollout generates 12 answers, one per language. The reward is

$$R=\sum_{\ell=1}^L r_\ell + \mathbf{1}[\forall \ell, r_\ell=1],$$

where $r_\ell=+1$ for a correct option match, $r_\ell=0$ for an incorrect option match, and $r_\ell=-0.5$ for a hallucination not matching any option. The all-correct bonus is a shaping term for cross-lingual consistency. Optimization uses AdamW with learning rate $10^{-5}$, cosine schedule with 3% warmup, LoRA with $r=64$, $$Acc_\ell = \frac{1}{N}\sum_{i=1}^N \mathbf{1}[\hat y_{i,\ell}=y_i],$$0, dropout $$Acc_\ell = \frac{1}{N}\sum_{i=1}^N \mathbf{1}[\hat y_{i,\ell}=y_i],$$1, batch size 1 fact, and no KL regularization ($$Acc_\ell = \frac{1}{N}\sum_{i=1}^N \mathbf{1}[\hat y_{i,\ell}=y_i],$$2) (Rad et al., 4 Jun 2026).

The supervised objective combines per-language classification with a consistency term:

$$Acc_\ell = \frac{1}{N}\sum_{i=1}^N \mathbf{1}[\hat y_{i,\ell}=y_i],$$3

with $$Acc_\ell = \frac{1}{N}\sum_{i=1}^N \mathbf{1}[\hat y_{i,\ell}=y_i],$$4, where $$Acc_\ell = \frac{1}{N}\sum_{i=1}^N \mathbf{1}[\hat y_{i,\ell}=y_i],$$5 is the group mean over the 12 parallel copies in a single forward pass. A pure-SFT ablation with $$Acc_\ell = \frac{1}{N}\sum_{i=1}^N \mathbf{1}[\hat y_{i,\ell}=y_i],$$6 is also reported (Rad et al., 4 Jun 2026).

The headline empirical finding is that GRPO consistently outperforms SFT on PolyFact and transfers better to free-form factual recall. For OLMo-2-7B on PolyFact, the base model reports high-resource accuracy 57.93 and low-resource accuracy 51.80, while GRPO reaches 64.21 and 54.48. For Qwen-2.5-7B, the base model reports 66.69 and 52.26, while GRPO reaches 73.15 and 56.93. On KLAR-CLC, GRPO also improves seen and OOD language performance for both model families, whereas SFT can regress and often outputs “1” or “2” rather than answers. On Global-MMLU, gains are modest, indicating stronger effects for direct factual recall than for broader multilingual reasoning (Rad et al., 4 Jun 2026).

Mechanistic analysis attributes these gains to changes in multilingual routing and specialization. Using language-specific neuron analysis based on Shannon entropy of MLP activations, GRPO increases English-specific neurons by 38.2% while reducing specialization for several non-English languages; ECDF analysis indicates that specialization is deferred to later layers under GRPO. Using attention-head importance estimates, GRPO reduces early-layer language routing, spreads routing across layers 0–10, and increases overlap even for typologically distant pairs such as Japanese–Chinese. This suggests, in the paper’s interpretation, that GRPO promotes more shared cross-lingual representations before later-stage language specialization (Rad et al., 4 Jun 2026).

3. PolyFact as numerical polynomial factorization

In numerical algebraic geometry, PolyFact denotes the framework developed in “The numerical factorization of polynomials,” where polynomial factorization is recast as a regularized geometric projection problem. The central claim is that classical exact factorization is ill-posed under coefficient perturbations because factorization structure is discontinuous with respect to small numerical noise. The paper formalizes numerical factorization by stratifying polynomial space into factorization manifolds and defining the desired output as the projection of a perturbed polynomial onto the manifold with the intended factorization structure (Wu et al., 2021).

The formal setting uses polynomial spaces $$Acc_\ell = \frac{1}{N}\sum_{i=1}^N \mathbf{1}[\hat y_{i,\ell}=y_i],$$7 and the Euclidean norm on coefficient vectors. Because factorization is defined only up to nonzero scalar multiples, a scaling-invariant “sine” distance is introduced:

$$Acc_\ell = \frac{1}{N}\sum_{i=1}^N \mathbf{1}[\hat y_{i,\ell}=y_i],$$8

For a factorization structure

$$Acc_\ell = \frac{1}{N}\sum_{i=1}^N \mathbf{1}[\hat y_{i,\ell}=y_i],$$9

the associated manifold $$Cons_{\mathrm{all}} = \frac{1}{N}\sum_{i=1}^N \mathbf{1}[\forall \ell,\hat y_{i,\ell}=y_i].$$0 is a complex analytic manifold of codimension

$$Cons_{\mathrm{all}} = \frac{1}{N}\sum_{i=1}^N \mathbf{1}[\forall \ell,\hat y_{i,\ell}=y_i].$$1

The numerical factorization is then defined as projection onto the manifold of highest singularity within tolerance $$Cons_{\mathrm{all}} = \frac{1}{N}\sum_{i=1}^N \mathbf{1}[\forall \ell,\hat y_{i,\ell}=y_i].$$2 (Wu et al., 2021).

The paper establishes existence and uniqueness of this projection in a tubular neighborhood, local Lipschitz continuity, a condition-number bound governed by the Moore–Penrose pseudoinverse of the Jacobian, and convergence to the underlying exact factorization. Sensitivity is bounded by

$$Cons_{\mathrm{all}} = \frac{1}{N}\sum_{i=1}^N \mathbf{1}[\forall \ell,\hat y_{i,\ell}=y_i].$$3

and the condition number is defined through an infimum of pseudoinverse norms over equivalent parameterizations (Wu et al., 2021).

Algorithmically, the method has two stages: identification of factorization structure using approximate GCDs induced by directional derivatives, reducibility tests such as Ruppert matrices, rank-revealing methods, and generalized eigenvalue analysis; and refinement via nonlinear least squares solved by Gauss–Newton:

$$Cons_{\mathrm{all}} = \frac{1}{N}\sum_{i=1}^N \mathbf{1}[\forall \ell,\hat y_{i,\ell}=y_i].$$4

The implementation described as NAClab’s PolynomialFactor uses coefficient scaling, orthonormalization, and robust linear algebra via SVD and pseudoinverse. The reported univariate example with rounded coefficients and multiple roots shows recovery of the intended multiplicity structure with residual on the order of coefficient rounding, e.g. $$Cons_{\mathrm{all}} = \frac{1}{N}\sum_{i=1}^N \mathbf{1}[\forall \ell,\hat y_{i,\ell}=y_i].$$5 (Wu et al., 2021).

4. PolyFact in symbolic and algebraic factorization theory

A second major family of uses treats PolyFact as shorthand for polynomial factorization in exact or approximative complexity theory. In “The Complexity of Factors of Multivariate Polynomials,” the core result is a multiplicity-free upper bound on the approximative complexity of an irreducible factor $$Cons_{\mathrm{all}} = \frac{1}{N}\sum_{i=1}^N \mathbf{1}[\forall \ell,\hat y_{i,\ell}=y_i].$$6 of a polynomial $$Cons_{\mathrm{all}} = \frac{1}{N}\sum_{i=1}^N \mathbf{1}[\forall \ell,\hat y_{i,\ell}=y_i].$$7 over characteristic-zero fields. With straight-line program complexity $$Cons_{\mathrm{all}} = \frac{1}{N}\sum_{i=1}^N \mathbf{1}[\forall \ell,\hat y_{i,\ell}=y_i].$$8, approximative complexity $$Cons_{\mathrm{all}} = \frac{1}{N}\sum_{i=1}^N \mathbf{1}[\forall \ell,\hat y_{i,\ell}=y_i].$$9, exclusion complexity $G=8$0, and univariate multiplication cost $G=8$1, the paper proves

$G=8$2

removing factor multiplicity from the bound that appears in Kaltofen’s 1986 result. Special cases include

$G=8$3

A main implication is that over characteristic zero there are no algebraic analogues of one-way functions among polynomial mappings of polynomially bounded degree whose graphs are checkable in polynomial time, because graph-checkability implies polynomially bounded approximative computability (Bürgisser, 2018).

In sparse exact factorization, another PolyFact usage names an algorithm that reduces sparse multivariate factorization over $G=8$4 or $G=8$5 to $G=8$6 bivariate factorizations by diagonal slicing. One chooses a main variable $G=8$7, substitutes the remaining variables by powers of a fresh indeterminate $G=8$8, and compares paired slices differing only in one exponent parameter. If a monomial has total $G=8$9-exponents differing by $$R=\sum_{\ell=1}^L r_\ell + \mathbf{1}[\forall \ell, r_\ell=1],$$0 between two such slices, the corresponding original exponent is recovered by a difference quotient. The method relies on content normalization via leading coefficients in $$R=\sum_{\ell=1}^L r_\ell + \mathbf{1}[\forall \ell, r_\ell=1],$$1, monomial-count stability checks, and random dilations by $$R=\sum_{\ell=1}^L r_\ell + \mathbf{1}[\forall \ell, r_\ell=1],$$2 or $$R=\sum_{\ell=1}^L r_\ell + \mathbf{1}[\forall \ell, r_\ell=1],$$3 to disambiguate coefficients when multiple monomials share the same $$R=\sum_{\ell=1}^L r_\ell + \mathbf{1}[\forall \ell, r_\ell=1],$$4-degree. It is implemented in Giac/Xcas and is reported to factor the paper’s five-variable test input in under 2 seconds (Parisse, 2016).

A different exact setting appears in factorization of multilinear polynomials over $$R=\sum_{\ell=1}^L r_\ell + \mathbf{1}[\forall \ell, r_\ell=1],$$5. There, the structural theorem is variable-disjointness: if a multilinear polynomial factors nontrivially, its factors depend on disjoint variable sets. The baseline derivative-based algorithm partitions variables using

$$R=\sum_{\ell=1}^L r_\ell + \mathbf{1}[\forall \ell, r_\ell=1],$$6

with $$R=\sum_{\ell=1}^L r_\ell + \mathbf{1}[\forall \ell, r_\ell=1],$$7 and $$R=\sum_{\ell=1}^L r_\ell + \mathbf{1}[\forall \ell, r_\ell=1],$$8. An improved recursive identity test IsEqual(A,D,B,C) avoids explicitly forming products and yields a conservative complexity bound $$R=\sum_{\ell=1}^L r_\ell + \mathbf{1}[\forall \ell, r_\ell=1],$$9 via Akra–Bazzi analysis, while empirically outperforming Maple’s built-in factorization on large sparse inputs (Emelyanov et al., 2018).

These strands share the term “PolyFact” only nominally. A plausible implication is that, across algebraic-computation papers, the label functions as a compact name for factorization systems rather than as a standardized framework.

5. Geometry-driven and convex-dense PolyFact algorithms

For bivariate polynomial factorization, “Improvements of convex-dense factorization of bivariate polynomials” introduces a deterministic algorithm whose complexity is driven by Newton-polygon geometry rather than solely by total degree. The polynomial is represented in convex-dense form, and the key geometric quantities are the Newton polygon volume $r_\ell=+1$0 and the minimal lower lattice length $r_\ell=+1$1. Under a non-degeneracy hypothesis, the main bound is

$r_\ell=+1$2

improving classical $r_\ell=+1$3 bounds when $r_\ell=+1$4. The method combines lower-hull combinatorics, local analytic factorization with respect to slope valuations in $r_\ell=+1$5, and global recombination by residues (Weimann, 10 Jan 2025).

The local stage uses the slope valuation

$r_\ell=+1$6

and a fast $r_\ell=+1$7-analytic factorization theorem that computes local factors to relative precision $r_\ell=+1$8 in soft-linear time $r_\ell=+1$9, modulo univariate factorizations of total degree at most $r_\ell=0$0. Recombination is then posed as a linear-algebra problem over a sparse ring using residues

$r_\ell=0$1

with constancy enforced through divisibility by a differential operator $r_\ell=0$2. Choosing the average slope $r_\ell=0$3 yields a recombination matrix of dimension $r_\ell=0$4 (Weimann, 10 Jan 2025).

The paper emphasizes that $r_\ell=0$5, so lower lattice length controls the number of positive-volume Minkowski summands and hence the number of irreducible factors of positive $r_\ell=0$6-degree. This makes $r_\ell=0$7 both a complexity parameter and a combinatorial bound on factor structure. The approach is presented as suitable for a modular “PolyFact system” with geometry, local factorization, and recombination modules (Weimann, 10 Jan 2025).

In a different mathematical direction, “Polynomial factorization statistics and point configurations in $r_\ell=0$8” uses PolyFact to denote any statistic on monic degree-$r_\ell=0$9 polynomials over $r_\ell=-0.5$0 that depends only on factorization type. Such a statistic is equivalently a class function on $r_\ell=-0.5$1 via cycle type. The paper proves that expected values of these statistics over all monic polynomials are governed by the $r_\ell=-0.5$2-representation on

$r_\ell=-0.5$3

with

$r_\ell=-0.5$4

and establishes an analogous squarefree formula via $r_\ell=-0.5$5. This is a representation-theoretic use of “PolyFact,” not an algorithmic one (Hyde, 2018).

6. PolyFact in machine learning models and matrix factorization

A further cluster of meanings uses “PolyFact” for factorization-based machine-learning models rather than for polynomial decomposition itself. In “Multi-output Polynomial Networks and Factorization Machines,” PolyFact names a shared-basis multi-output framework extending polynomial networks and factorization machines to vector-valued prediction. For degree 2, output slice $r_\ell=-0.5$6 is parameterized as

$r_\ell=-0.5$7

or, in infinite-dictionary form,

$r_\ell=-0.5$8

Learning is cast as the convex program

$r_\ell=-0.5$9

with row-sparsity-inducing norms such as $10^{-5}$0, $10^{-5}$1, and $10^{-5}$2. A fully corrective conditional-gradient algorithm has a global convergence guarantee: if $10^{-5}$3 is $10^{-5}$4-smooth and the atom-selection oracle has multiplicative factor $10^{-5}$5, then

$10^{-5}$6

for

$10^{-5}$7

Empirically, the framework achieves strong classification accuracy with sparser models than several baselines and improves ranking accuracy on recommendation tasks when paired with an ordinal-to-classification reduction (Blondel et al., 2017).

In matrix factorization, PolyFact also refers to Polytopic Matrix Factorization. Here the model is

$10^{-5}$8

with full-column-rank $10^{-5}$9 and latent columns constrained to a polytope $r=64$0. The determinant-maximization criterion is

$r=64$1

equivalently minimizing $r=64$2. Identifiability is analyzed through the maximum-volume inscribed ellipsoid $r=64$3 and a sufficiently scattered condition:

$r=64$4

For canonical polytopes such as $r=64$5, $r=64$6, $r=64$7, and $r=64$8, the paper proves recovery up to signed permutations or permutations only, depending on the geometry (Tatli et al., 2022).

A companion work studies identifiable polytope characterization via automorphism groups. The central criterion is that PMF is identifiable up to signed permutations exactly when the linear automorphism group of the polytope is contained in the signed permutation group. The work proposes checking this through graph automorphism on an edge-colored complete graph built from polytope combinatorics, then lifting graph symmetries back to linear maps. This replaces factorial search over permutations with a graph-automorphism-based procedure (Bozkurt et al., 2022).

Taken together, these uses extend “PolyFact” well beyond classical factorization. This suggests that the term has become a flexible naming convention for structured factor models in which algebraic or geometric constraints provide identifiability, sparsity, or cross-output sharing.

7. Terminological scope and disambiguation

The term “PolyFact” therefore has no single canonical meaning across arXiv literature. In current usage, at least four distinct senses are technically established. First, it names a multilingual factual-recall corpus and GRPO-based training protocol for LLMs (Rad et al., 4 Jun 2026). Second, it denotes numerical polynomial factorization framed as projection onto factorization manifolds (Wu et al., 2021). Third, it appears as a shorthand for several exact or approximative polynomial-factorization algorithms and statistics, including sparse slicing methods, multilinear factorization over $r=64$9, convex-dense bivariate factorization, and class-function statistics over $$Acc_\ell = \frac{1}{N}\sum_{i=1}^N \mathbf{1}[\hat y_{i,\ell}=y_i],$$00 (Parisse, 2016, Emelyanov et al., 2018, Weimann, 10 Jan 2025, Hyde, 2018). Fourth, it labels shared-factor machine-learning methods such as multi-output polynomial networks and polytopic matrix factorization (Blondel et al., 2017, Conradi et al., 2022).

Because these usages are domain-specific and largely unrelated, disambiguation depends on context. In NLP, PolyFact almost always refers to the 2026 multilingual factual QA benchmark and consistency-driven RL setup. In numerical algebra and symbolic computation, it usually abbreviates some polynomial factorization framework. In matrix-factorization literature, it can instead refer to polytopic latent-variable models with determinant-based identifiability criteria. The term is thus best treated as a context-sensitive research label rather than a single established concept.