WorksComputer Science
The Complexity of KKT Points of Quadratic Programs with Polynomially-Bounded Coefficients
Abstract
Computing a Karush-Kuhn-Tucker (KKT) point of a quadratic polynomial over the unit hypercube is known to be -complete, when the coefficients of the polynomial are encoded in binary. We prove that the problem remains -complete even when the coefficients are encoded in unary. This is the first problem proven to remain -hard when the input is given in unary.
Introduction
Quadratic programming is possibly one of the most fundamental problems in continuous optimization, with applications ranging from game theory and machine learning, to engineering and computational chemistry [12]. One of the most basic forms of the problem is defined via “box constraints” [1, 4]. There, the problem is defined via an symmetric coefficient-matrix , and an -dimensional coefficient-vector . The goal is to find a point in the -dimensional hypercube that minimizes the corresponding quadratic program, formally:
Since the complexity of the problem was proven to be NP-hard more than fifty years ago [17, 18], the relaxed solution of Karush-Kuhn-Tucker, a.k.a., KKT, points was considered. There, instead of a global optimum of , the solution can be any first-order stationary point, a.k.a. fixed point of gradient descent, of the polynomial. In a recent celebrated result, [11] showed that computing a KKT point of (1), even in the approximate sense, is CLS-complete. This serves as an indication that it is unlikely the problem to have a polynomial-time algorithm.
The complexity class CLS, standing for Continuous Local Search, was introduced by Daskalakis and Papadimitriou [9] to capture the complexity of search problems that belong to both PPAD [14] and PLS [13] yet no polynomial-time algorithm is known. In other words, the existence of the solution is guaranteed either via a fixed point argument or via a potential function argument. In their breakthrough result, [11] proved that CLS = PPAD PLS.
Since then, a series of works established CLS-hardness for congestion games [5] and equilibrium computation problems [10, 20], and very recently, in yet another breakthrough result, for KKT points in bilinear polynomials, i.e., polynomials without any square term [2]. However, in all of the above-mentioned results the input of the problem was given with binary encoding, hence the input could include coefficients of exponential precision. In fact, the correctness of the reduction of [11] heavily relies on the existence of exponential ratios between the coefficients and it explicitly asks as an open problem whether CLS-hardness still holds when the input is given in unary. Additionally, for the case of KKT points for bilinear polynomials, binary encoding is actually required in order to establish CLS-hardness [2]. If the input is given in unary the bilinear polynomials become polynomial-time solvable.
To see the latter, observe that a bilinear polynomial is affine in each coordinate separately. So it always has a global minimum, and hence a KKT point, at a Boolean vector, i.e., a solution where . Moreover, a Boolean point is a KKT point precisely when no single bit flip decreases the objective. If the coefficients are integers, each improving flip decreases the objective by at least one. With polynomially bounded coefficients, the range of the objective on Boolean points is also polynomially bounded, and this elementary local search procedure terminates in polynomial time.
Hence, the open question from [11] becomes even more intriguing: is binary encoding required to establish CLS-hardness in general, or can we prove hardness even when the input is given in unary?
Our result and its implications. We answer the open question of [11] and prove that finding a KKT point of (1) remains CLS-hard even when all coefficients, of and , are given in unary. This is the first problem where the input is given in unary and it is CLS-hard. As a consequence, our result rules out the existence of pseudopolynomial-time algorithm for KKT points. Our result provides an interesting dichotomy: allowing square terms produces a sharp change in complexity of KKT points. When the input is given in unary, bilinear KKT admits a polynomial time algorithm, whereas quadratic KKT remains CLS-complete; see Table 1. Thus, our result implies that the reduction of [2] is essentially “optimal”: the exponential blow-up in the coefficients of the produced bilinear polynomial is necessary. Finally, our result can be seen as a complement to the recent result of [2]. The equality gives two possible approaches to solving problems in CLS: one through fixed point computation and another through local search. A restriction that makes either approach efficient is enough to obtain a polynomial-time algorithm; for the latter, [19, 16] observed that many problems become tractable under unary encoding. Along with [2], we establish that hardness for KKT points persists when either square terms or large integer coefficients are excluded separately, but excluding both makes the problem tractable.
Table 1. The complexity of computing exact KKT points over the unit cube.
Technical overview
Our starting point is the CLS-complete problem QUADRATIC-KKTproblem from [11], where we are given a quadratic polynomial with variables and binary-encoded coefficients, i.e., coefficients can have exponential precision. We construct a quadratic polynomial that preserves all stationary points of and has polynomially-bounded integer coefficients. The high-level idea is to simulate the exponentially-precise coefficients of with polynomially-many auxiliary variables: we “push” the precision of the coefficients to the values of the auxiliary variables at any KKT point of the constructed polynomial. In other words, at any KKT point of the constructed polynomial, the auxiliary variables need exponential precision and thus they can simulate the coefficients or the original instance.
More concretely now, our reduction consists of three main steps.
Step 1: Normalization. As first step, we normalize the coefficients of the input polynomial to take values in , without changing its KKT points. This is done via a “standard” way. This step is crucial for our reduction in order to guarantee the solution space to be the unit hypercube; without normalization the next steps of the reduction would have required exponentially large numbers as solution for the auxiliary variables.
Step 2: Dyadic decomposition. The next step is to decompose the normalized polynomial as a sum of monomials with coefficients of the form , where is polynomially-bounded by the encoding length of the input. In fact, is bounded by the maximum number of bits required for any coefficient of . This decomposition of allows us to focus on each monomial independently and introduce independent auxiliary variables. Hence, every monomial is of the form , where may also be equal to , and its partial derivatives are
Observe, each variable contributes to the gradient of a value scaled by , i.e., a value obtained by halving the value times. Our goal then boils down to simulate this halving effect using integer coefficients by introducing a polynomial number of auxiliary variables.
Before we introduce the next step of the reduction, which simulates the halving effect, let us provide some high-level intuition for it. We can think of the variables and of a monomial as two gates that transmit signals to each other, using a chain of auxiliary variables as shown in Figure 1.
Figure 1. Two-way propagation of values through auxiliary variables.
A “blue” signal originates from and propagates towards , while getting halved each time it passes through an intermediate gate. Similarly, a “red” signal originates from and propagates towards , undergoing the same halving process. The goal is to create a gadget that: a) preserves the KKT points of and b) does not create any additional KKT points. The next step of the reduction formally achieves this.
Step 3: The quadratic chain gadget. For each monomial , where , we introduce auxiliary variables , denoted for brevity, each constrained in , and construct the following quadratic chain gadget:
Each chain gadget satisfies the following property which is crucial for our purposes: for every fixed values of and the function is strictly convex with respect to . This implies that, for fixed , admits a unique minimizer, denoted , which is the unique KKT point of as well. Crucially, the choice of the coefficients of ensure that always resides inside . In fact, we derive a closed form formula for the auxiliary variables where for we get that
Thus, at any KKT point of , the values of the auxiliary variables require exponential precision, which was our initial goal. The term represents the “blue” signal from Figure 1, which halves as increases from 1 to , while the term represents the “red” signal. The detailed analysis is presented in Lemma 4.4.
As a side note, let us highlight that our chain gadget guarantees more than just equality of the derivatives: it establishes the stronger identity that is equal to , which can be of independent interest (see Lemma 4.5).
Putting everything together. To complete our reduction, the final constructed polynomial is the summation of all chain gadgets of the dyadic monomials. The resulting polynomial has integer coefficients of polynomially bounded magnitude. To see this, observe the following: a) the coefficients in every chain gadget are constants; b) there are polynomially many chain gadgets; c) the length of each gadget is polynomial. Furthermore, since the auxiliary variables are disjoint across gadgets, the arguments from Step 3 for KKT points apply independently to them.
Preliminaries
For any positive integer , we denote . We consider quadratic polynomials over variables that are given as a sum of monomials:
where for all the coefficients are assumed to be rational numbers.
We follow the definition of box-constrained KKT point of from [11].
Definition 3.1 (Box-constrained KKT points). For a quadratic polynomial , as expressed in (2), consider the constrained optimization problem . A point is a Karush–Kuhn–Tucker (KKT) point of this optimization problem if for every coordinate the following conditions are satisfied:
As it was formally explained in [11], under the above mentioned definition a KKT point always does exist.
Binary Vs Unary encoding. A quadratic polynomial is represented by an explicit list of its monomials and their corresponding rational coefficients. Hence, the input size is given by the number of variables, which is , plus the size of encoding the coefficients. When the coefficients are encoded in binary, i.e., as bitstrings, they can represent numbers of exponential precision, with respect of the input size. For example, the ratio between two coefficients can be . On the other hand, when the coefficients are given in unary, they have only polynomially-bounded precision. In other words, can be represented using integer-coefficients with absolute values bounded by the unary encoding length of the input polynomial. Hence, in this case the ratio between any two coefficients is bounded by .
We proceed with the definition of Qᴜᴀᴅʀᴀᴛɪᴄ-KKT, the main computational problem of interest, and the current state of the art for its complexity.
Definition 3.2 (Qᴜᴀᴅʀᴀᴛɪᴄ-KKT). Given a quadratic polynomial over variables, compute a KKT point of the optimization problem .
Theorem 3.3 ([11]). Qᴜᴀᴅʀᴀᴛɪᴄ-KKT is CLS-complete, when the input polynomial is given in binary.
The CLS-hardness reduction
In this section we establish the main theorem of our paper.
Theorem 4.1. Quadratic-KKT is CLS-complete even when the input is given in unary.
We present a polynomial-time reduction from Quadratic-KKT given with binary-encoded coefficients. Specifically, given an instance defined by a quadratic polynomial , we construct a quadratic polynomial such that (i) the unary encoding length of is polynomially bounded in the encoding length of , and (ii) every KKT point of projects to a KKT point of .
We now outline the main steps of the reduction. In § 4.1, we construct a normalized quadratic polynomial whose coefficients admit a decomposition into signed dyadic terms of the form , with and being polynomially bounded in the encoding length of . Lemma 4.2 establishes that and have exactly the same KKT points.
In § 4.2, we introduce the quadratic chain gadget, which is the main technical tool of our reduction. For any dyadic monomial , the gadget constructs a quadratic polynomial with constant-bounded integer coefficients and auxiliary variables. We prove that minimizing this polynomial over its auxiliary variables reproduces exactly, as established in Lemma 4.5.
In Section 4.3, we combine the quadratic chain gadgets corresponding to all dyadic monomials of . We show that every KKT point of the resulting quadratic polynomial projects onto a KKT point of , and hence of . Finally, in § 4.4, we transform the auxiliary domain to the unit hypercube and establish that the resulting polynomial has integer coefficients with polynomially bounded unary encoding lengths.
Normalizing the Coefficients and Dyadic decomposition
Consider an instance of Quadratic-KKT and let be the corresponding quadratic polynomial over , as defined in Equation (2). Let denote the set of KKT points of . We construct a polynomial whose coefficients are polynomially bounded in the encoding length of , while preserving the set of KKT points, i.e., .
Let be the least common multiple of the denominators of the coefficients appearing in . Compute the polynomial . Notice that admits only integer coefficients. Let be the maximum value among and the absolute values of the coefficients in . Set and define . Finally compute the polynomial . For this transformation the following holds.
Lemma 4.2. The polynomials and have exactly the same KKT points, namely . Moreover every coefficient of admits a dyadic decomposition of the form
where and .
Proof. For every and every , we have . Since , the corresponding partial derivatives of and have the same sign and are equal to at the same points. Consequently, for every , the KKT condition associated with is satisfied by at if and only if it is satisfied by at and thus .
Since has integer coefficients and , every coefficient of can be expressed as , for some integer or which . If , the decomposition holds by
choosing and for every . Otherwise, let denote the sign of . Since , the binary representation of contains at most bits. Therefore, there exist such that . Dividing both sides by and multiplying by , we get .
Using the above lemma, we can express as a sum of monomials with signed dyadic coefficients. Let denote the total number of nonzero terms obtained from expanding each coefficient in according to (4). For each , let . Their products represent all types of monomials appearing in , namely, quadratic, linear, and constant monomials. Consequently, we can express as the following dyadic expansion.
Since each coefficient contributes at most terms to the dyadic expansion and contains monomials, we have . Moreover, because is polynomially bounded in the encoding length of , the dyadic expansion can be computed in polynomial time.
Quadratic chain gadget
Consider a dyadic monomial , where , and . Let and introduce auxiliary variables . Consider the following quadratic polynomial which we refer to as quadratic chain gadget.
For fixed values of and in , let denote the quadratic chain gadget, viewed as a function of . Notice that the quadratic polynomial is continuous over the compact set , and hence admits at least one global minimizer in this domain. Let denote a global minimizer of , that is,
The next lemma shows the relation between a global minimizer and a KKT point of .
Lemma 4.3. For every fixed , the quadratic polynomial admits a unique global minimizer over , which is also the unique KKT point of over $[-1,1]^m.
Proof. First we will show that is strictly convex in . Fix the values for the variables and let denote the quadratic part of , i.e.,
Since differs from only by linear terms in , to establish that is strictly convex, it suffices to show that is a positive-definite quadratic form, i.e., for every .1 Now observe that as a sum of square terms while the equality holds if and only if . Thus is a positive definite form and hence is strictly convex. As a consequence of strict convexity, admits at most one global minimizer in .2 Since admits at least one global minimizer in , then is the unique global minimizer of , namely
Finally since is differentiable and strictly convex over , a point satisfies the KKT conditions for if and only if it is a global minimizer.3 As is the unique global minimizer of , it follows that it is also the unique KKT point of .
However, uniqueness on its own does not suffice to establish the correctness of the gadget. Since the auxiliary variables are box-constrained a KKT point of , with and not fixed, does not necessarily satisfy the stationarity equations of the original polynomial. We must therefore ensure that the auxiliary variables satisfy the unconstrained stationarity conditions, at every KKT of . This is satisfied if we prove that the unique minimizer for every fixed lies in the interior of the feasible domain.
Lemma 4.4. For every it holds that .
Proof. Fix and consider the unconstrained stationarity equations of as follows
Since is strictly convex, as established in Lemma 4.3, its unconstrained stationarity system LS admits at most one solution. As a consequence if LS admits a solution, then this is unique.
We express the solution of LS as a linear combination of two auxiliary solutions. Specifically let LS1 denote the linear system obtained from LS by setting and and let LS2 denote the linear system obtained from LS by setting and . Observe that all three linear systems share the same coefficient matrix, which is equal to the Hessian of multiplied by . Since is a strictly convex quadratic form, its Hessian, and as a consequence the coefficient matrix of the linear systems, is positive definite and hence invertible. Consequently, each system admits a unique solution.
Consider and to be the solutions of LS1 and LS2, respectively. By linearity we get that
We now obtain the explicit solutions of LS1 and LS2. Observe that both systems share the following linear homogeneous recurrence relation
where we use to denote either or . We solve this recurrence relation using the roots of its characteristic polynomial. Assume that for . Substituting with in the recurrence relations and dividing by we get the characteristic quadratic polynomial
Solving this polynomial we get the characteristic roots and . Finally a solution to the homogeneous recurrence relation is given by the formula
We proceed to determine the parameters and using the boundary conditions in LS1 and LS2. Consider LS1 whose boundary conditions are and and let for be a solution. Setting as and substituting (12) in the first boundary condition we get . Similarly substituting (12) in the second boundary condition we get and thus
Next, consider LS2, whose boundary conditions are and and let for be a solution. Setting as and substituting (12) in the first boundary condition we get . Similarly substituting (12) in the second boundary condition we get and therefore
Substituting (13) and (14) in (11) we get that the unique solution of LS is
Finally since , applying the triangle inequality for every ,
which proves that .
Having established the interiority of the unique minimizer , we proceed to show that substituting into the quadratic chain gadget yields exactly the original dyadic monomial .
Lemma 4.5. For every and in , it holds that . Proof. We substitute the explicit form of as given in (15), in the quadratic chain gadget expression (6). For it holds that
Therefore,
Adding the explicit expression and we get that
since after calculations we get that .
Computing a KKT point of
Recall that the dyadic decomposition of , given in (5), consists of monomials of the form , with and . For each monomial consider a set of disjoint auxiliary variables and construct the quadratic chain gadget as given in (6). Letting and , we add the quadratic chain gadgets and get the induced combined quadratic chain polynomial
Notice that because each quadratic chain gadget admits a disjoint set of auxiliary variables, then for fixed , minimizing over is equivalent to minimizing each over . Since for fixed by the notation we defined earlier
then combining the above expression with Lemma 4.5 we get that
For any , let , where denotes the unique auxiliary minimizer of the -th quadratic chain gadget. Then for a fixed vector ,
where from equations (17) and (5) we get that . Finally we establish that every KKT point of maps back to an exact KKT point of .
Proposition 4.6. If is a KKT point of then is a KKT point of .
Proof. Let be a KKT point of . Since the auxiliary variables of the quadratic chain gadgets are disjoint, for every , the vector satisfies the KKT conditions of with respect to . From Lemma 4.3, each quadratic chain gadget admits a unique KKT point with respect to its auxiliary variables which coincides with its unique interior minimizer. Consequently, for every , and hence . From the previous discussion it holds that for every , . Differentiating with respect to , for , we get that
Because in Lemma 4.4, we showed that the auxiliary minimizers are interior therefore for every and . Hence for every it follows that
Since is a KKT point of , it follows from Equation (18) that satisfies the KKT conditions of over .
Transforming the auxiliary domain
To finalize the reduction we transform the domain of the auxiliary variables from to . Let denote the total number of auxiliary variables in . For each auxiliary variable , we introduce a new variable such that . Let denote the vector of all new auxiliary variables and let be the resulting quadratic polynomial defined over . For every and every auxiliary coordinate ,
Also the substitution maps the boundary points of the two intervals as follows: if and only if and if and only if , while every interior point is mapped again to an interior point. Consequently, the KKT conditions for the original variables remain unchanged, while the KKT conditions for the auxiliary variables are preserved because their partial derivatives are multiplied by 2. Hence,
Proof of Theorem 4.1
Consider an arbitrary instance of Quadratic-KKT defined by a quadratic polynomial over , and let denote its binary encoding length. By Lemma 4.2, the polynomial constructed in § 4.1 has exactly the same KKT points as and admits a dyadic decomposition consisting of monomials.
We first show that . Recall from § 4.1 that , where let’s assume that are the denominators of the coefficients of , and that is the maximum of 1 and the absolute values of the coefficients of , while . Since , we have . As the numerators of also have binary encoding lengths at most , every coefficient of has absolute value at most . Consequently, , and hence .
Next, we construct the combined quadratic chain polynomial as defined in Equation (16), that is summing over all quadratic chain gadgets (see Equation (6)) that correspond to the dyadic monomials of . Let denote the total number of auxiliary variables that emerge from the construction of . By Proposition 4.6, every KKT point of projects onto a KKT point of , and hence, by Lemma 4.2, to a KKT of . Applying the affine transformation of § 4.4, we obtain the quadratic polynomial over the domain . By Equation (19), every KKT point of corresponds to a KKT point of , and therefore yields a KKT point of .
It remains to bound the encoding length of the constructed instance . Since each gadget introduces auxiliary variables, their total number satisfies
Moreover, each gadget contains terms with constant-bounded integer coefficients (see Equation (6)), which also holds after the affine transformation. Thus, every coefficient of the resulting polynomial is an integer of absolute value at most
Since we established that , both the number of variables and the unary encoding lengths of the coefficients of are polynomially bounded in , i.e, the encoding length of . Since is quadratic and has polynomially many variables, then it contains at most quadratically many monomials in the number of variables. Furthermore, its construction can be performed in polynomial time, which concludes our proof.
Discussion
One obvious and important question that remains open is whether Quadratic-KKT remains hard, either for CLS, or for a different complexity class/under a different assumption, when the coefficients are constant. Such a result would not contradict anything of our current understanding of CLS, or TFNP in general. As a side note, let us highlight two things. Firstly, quadratic optimization is NP-hard even when all coefficients are constants (think of Max-Cut in unweighted graphs). Second, there exist several problems that are PPAD-hard even when all coefficients are small constants [7] or even binary [15, 8, 3].
One approach to attack the above mentioned problem would be to use our technique when the starting point is a quadratic polynomial with polynomially-bounded coefficients and hope that after a number of iterations we can end up with a polynomial with constant coefficients. Unfortunately, this cannot work under the current construction of the chain gadget. The reason is that after the dyadic decomposition, for every occurrence of a variable it creates one copy of in every chain. One way to bypass this obstacle could be by proving that Quadratic-KKT remains hard even on “sparse” instances, i.e., instances where each variable appears only a constant number of times. Another direction could be to create a different chain gadget that does not preserve a one-to-one correspondence of stationary points, but works only the approximate sense.
Acknowledgements
We sincerely thank Ioannis Panageas for the insightful discussions on CLS and for his substantial contribution to the development of the final form of the paper. Argyrios Deligkas and Stavros D. Ioannidis were supported by the EPSRC grant EP/X039862/1. We used various versions of GPT Pro to explore proof strategies during the development of this work.
References
- [1]Amar Andjouh and Mohand Ouamer Bibi. Adaptive global algorithm for solving box-constrained non-convex quadratic minimization problems. Journal of Optimization Theory and Applications, 192(1):360–378, 2022.
- [2]Ioannis Anagnostides, Ioannis Panageas, and Jingming Yan. The complexity of Nash equilibrium in network congestion and coordination games. arXiv preprint arXiv:2609.33225, 2026.arxiv.org/abs/2609.33225
- [3]Eleni Batziou, John Fearnley, Abheek Ghosh, and Rahul Savani. The complexity of sparse win-lose bimatrix games. arXiv preprint arXiv:2602.18380, 2026.arxiv.org/abs/2602.18380
- [4]Samuel Burer and Adam N Letchford. On nonconvex quadratic programming with box constraints. SIAM Journal on Optimization, 20(2):1073–1089, 2009.DOI
- [5]Yakov Babichenko and Aviad Rubinstein. Settling the complexity of Nash equilibrium in congestion games. In Proceedings of the 53rd annual ACM SIGACT symposium on theory of computing, pages 1426–1437, 2021.arxiv.org/abs/2012.04327
- [6]Stephen Boyd and Lieven Vandenberghe. Convex Optimization. Cambridge University Press, 2004.
- [7]Argyrios Deligkas, John Fearnley, Alexandros Hollender, and Themistoklis Melissourgos. Pure-circuit: Tight inapproximability for PPAD. Journal of the ACM, 71(5):1–48, 2024.DOI
- [8]Jérémi Do Dinh and Alexandros Hollender. Tight inapproximability of Nash equilibria in public goods games. Inf. Process. Lett., 186:106486, 2024.DOI
- [9]Constantinos Daskalakis and Christos Papadimitriou. Continuous local search. In Proceedings of the twenty-second annual ACM-SIAM symposium on Discrete Algorithms, pages 790–804. SIAM, 2011.
- [10]Edith Elkind, Abheek Ghosh, and Paul W. Goldberg. Simultaneous contests with equal sharing allocation of prizes: Computational complexity and price of anarchy. In International Symposium on Algorithmic Game Theory, pages 133–150. Springer, 2022.
- [11]John Fearnley, Paul W. Goldberg, Alexandros Hollender, and Rahul Savani. The complexity of computing KKT solutions of quadratic programs. J. ACM, 72(5):31:1–31:48, 2025.
- [12]Vladimir V. Gusev, Duncan Adamson, Argyrios Deligkas, Dmytro Antypov, Christopher M. Collins, Piotr Krysta, Igor Potapov, George R. Darling, Matthew S. Dyer, Paul G. Spirakis, and Matthew J. Rosseinsky. Optimality guarantees for crystal structure prediction. Nat., 619(7968):68–72, 2023.DOI
- [13]David S Johnson, Christos H Papadimitriou, and Mihalis Yannakakis. How easy is local search? Journal of computer and system sciences, 37(1):79–100, 1988.DOI
- [14]Christos H Papadimitriou. On the complexity of the parity argument and other inefficient proofs of existence. Journal of Computer and system Sciences, 48(3):498–532, 1994.
- [15]Christos Papadimitriou and Binghui Peng. Public goods games in directed networks. In Proceedings of the 22nd ACM Conference on Economics and Computation (EC), pages 745–762, 2021.arxiv.org/abs/2106.00718
- [16]C. H. Papadimitriou, A. A. Schäffer, and M. Yannakakis. On the complexity of local search. In Proceedings of the Twenty-Second Annual ACM Symposium on Theory of Computing, STOC ’90, page 438–445, New York, NY, USA, 1990. Association for Computing Machinery.
- [19]Alejandro A. Schäffer and Mihalis Yannakakis. Simple local search problems that are hard to solve. SIAM J. Comput., 20(1):56–87, 1991.DOI
- [20]Emanuel Tewolde, Caspar Oesterheld, Vincent Conitzer, and Paul W Goldberg. The computational complexity of single-player imperfect-recall games. arXiv preprint arXiv:2305.17805, 2023.arxiv.org/abs/2305.17805
