arXiv subjects

Ankit Garg

Publications and source records attributed to Ankit Garg.

At least 19 recordsLinked to original sources

arXiv2026

EigenLI: Spectral Approximations to Late Interaction

Late-interaction models such as ColBERT achieve strong effectiveness by representing each document with many token-level vectors, but this expressivity leads to large indexing cost, storage footprints and expensive MaxSim scoring. We show that late-interaction representations exhibit an intrinsic low-rank structure: document token embeddings concentrate in a low-dimensional subspace that preserves most of the retrieval signal. Leveraging this observation, we introduce EigenLI, a spectral approximation framework that compresses late-interaction representations via document-specific low-dimensional subspaces. Unlike clustering or pooling methods, EigenLI identifies the dominant eigendirections of each document and uses them to construct reduced interaction representations. Empirically, $k$-EigenLI with $k \le 32$ outperforms k-means and Ward clustering based pooling methods on ColBERTv2 and AnswerAI-ColBERT-small; GTE-ModernColBERT exhibits a different tradeoff at $k=32$, where clustering methods perform better. The same spectral construction also yields EigenLI-SV, an ANN-compatible single-vector representation derived from the second-order summary of the reduced structure. Across multiple datasets and all three text models, EigenLI-SV consistently outperforms comparable single-vector surrogates such as MUVERA.

cs.IR↗

arXiv2026

Retrieval Needs Multivectors: An Exponential Separation

Recent works have highlighted the expressive limitations of embedding based retrieval models through both theoretical analyses and challenging benchmarks such as LIMIT. While multi-vector embeddings consistently outperform single-vector embeddings, the precise representational gap between them remains poorly understood. In this work, following Jayaram's work, we provide the first explicit family of query and document sets, together with their relevance matrices, for which single-vector embeddings that rank all relevant documents above irrelevant ones require exponential size, whereas polynomial-size multi-vector embeddings suffice. Our result establishes an exponential separation between the expressive power of single-vector and multi-vector embeddings for the task of ranking of documents as opposed to approximating numerical scores as in the work of Jayaram. Motivated by our theoretical construction, we introduce ANDOR, a new retrieval benchmark that naturally instantiates these hard examples. We show that state-of-the-art single-vector embedding models perform poorly on ANDOR in the zero-shot setting and exhibit only marginal improvements after fine-tuning, highlighting the inherent difficulty of the benchmark compared to prior work. In contrast, multi-vector models consistently outperform their single-vector counterparts and improve substantially with fine-tuning, closely aligning with our theoretical predictions.

cs.IR↗

arXiv2026

Reading Between the Pixels: Linking Text-Image Embedding Alignment to Typographic Attack Success on Vision-Language Models

We study typographic prompt injection attacks on vision-language models (VLMs), where adversarial text is rendered as images to bypass safety mechanisms, posing a growing threat as VLMs serve as the perceptual backbone of autonomous agents, from browser automation and computer-use systems to camera-equipped embodied agents. In practice, the attack surface is heterogeneous: adversarial text appears at varying font sizes and under diverse visual conditions, while the growing ecosystem of VLMs exhibits substantial variation in vulnerability, complicating defensive approaches. Evaluating 1,000 prompts from SALAD-Bench across four VLMs, namely, GPT-4o, Claude Sonnet 4.5, Mistral-Large-3, and Qwen3-VL-4B-Instruct under varying font sizes (6--28px) and visual transformations (rotation, blur, noise, contrast changes), we find: (1) font size significantly affects attack success rate (ASR), with very small fonts (6px) yielding near-zero ASR while mid-range fonts achieve peak effectiveness; (2) text attacks are more effective than image attacks for GPT-4o (36% vs 8%) and Claude (47% vs 22%), while Qwen3-VL and Mistral show comparable ASR across modalities; (3) text-image embedding distance from two multimodal embedding models (JinaCLIP and Qwen3-VL-Embedding) shows strong negative correlation with ASR across all four models (r = -0.71 to -0.93, p < 0.01); (4) heavy degradations increase embedding distance by 10--12% and reduce ASR by 34--96%, while rotation asymmetrically affects models (Mistral drops 50%, GPT-4o unchanged). These findings highlight that model-specific robustness patterns preclude one-size-fits-all defenses and offer empirical guidance for practitioners selecting VLM backbones for agentic systems operating in adversarial environments.

cs.CV↗

arXiv2026

On Strengths and Limitations of Single-Vector Embeddings

Recent work (Weller et al., 2025) introduced a naturalistic dataset called LIMIT and showed empirically that a wide range of popular single-vector embedding models suffer substantial drops in retrieval quality, raising concerns about the reliability of single-vector embeddings for retrieval. Although (Weller et al., 2025) proposed limited dimensionality as the main factor contributing to this, we show that dimensionality alone cannot explain the observed failures. We observe from results in (Alon et al., 2016) that $2k+1$-dimensional vector embeddings suffice for top-$k$ retrieval. This result points to other drivers of poor performance. Controlling for tokenization artifacts and linguistic similarity between attributes yields only modest gains. In contrast, we find that domain shift and misalignment between embedding similarities and the task's underlying notion of relevance are major contributors; finetuning mitigates these effects and can improve recall substantially. Even with finetuning, however, single-vector models remain markedly weaker than multi-vector representations, pointing to fundamental limitations. Moreover, finetuning single-vector models on LIMIT-like datasets leads to catastrophic forgetting (performance on MSMARCO drops by more than 40%), whereas forgetting for multi-vector models is minimal. To better understand the gap between performance of single-vector and multi-vector models, we study the drowning in documents paradox (Reimers \& Gurevych, 2021; Jacob et al., 2025): as the corpus grows, relevant documents are increasingly "drowned out" because embedding similarities behave, in part, like noisy statistical proxies for relevance. Through experiments and mathematical calculations on toy mathematical models, we illustrate why single-vector models are more susceptible to drowning effects compared to multi-vector models.

cs.IR↗

arXiv2025

Cisco Integrated AI Security and Safety Framework Report

Artificial intelligence (AI) systems are being readily and rapidly adopted, increasingly permeating critical domains: from consumer platforms and enterprise software to networked systems with embedded agents. While this has unlocked potential for human productivity gains, the attack surface has expanded accordingly: threats now span content safety failures (e.g., harmful or deceptive outputs), model and data integrity compromise (e.g., poisoning, supply-chain tampering), runtime manipulations (e.g., prompt injection, tool and agent misuse), and ecosystem risks (e.g., orchestration abuse, multi-agent collusion). Existing frameworks such as MITRE ATLAS, National Institute of Standards and Technology (NIST) AI 100-2 Adversarial Machine Learning (AML) taxonomy, and OWASP Top 10s for Large Language Models (LLMs) and Agentic AI Applications provide valuable viewpoints, but each covers only slices of this multi-dimensional space. This paper presents Cisco's Integrated AI Security and Safety Framework ("AI Security Framework"), a unified, lifecycle-aware taxonomy and operationalization framework that can be used to classify, integrate, and operationalize the full range of AI risks. It integrates AI security and AI safety across modalities, agents, pipelines, and the broader ecosystem. The AI Security Framework is designed to be practical for threat identification, red-teaming, risk prioritization, and it is comprehensive in scope and can be extensible to emerging deployments in multimodal contexts, humanoids, wearables, and sensory infrastructures. We analyze gaps in prevailing frameworks, discuss design principles for our framework, and demonstrate how the taxonomy provides structure for understanding how modern AI systems fail, how adversaries exploit these failures, and how organizations can build defenses across the AI lifecycle that evolve alongside capability advancements.

cs.CR↗

arXiv2025

Incorporating Token Importance in Multi-Vector Retrieval

ColBERT introduced a late interaction mechanism that independently encodes queries and documents using BERT, and computes similarity via fine-grained interactions over token-level vector representations. This design enables expressive matching while allowing efficient computation of scores, as the multi-vector document representations could be pre-computed offline. ColBERT models distance using a Chamfer-style function: for each query token, it selects the closest document token and sums these distances across all query tokens. In our work, we explore enhancements to the Chamfer distance function by computing a weighted sum over query token contributions, where weights reflect the token importance. Empirically, we show that this simple extension, requiring only token-weight training while keeping the multi-vector representations fixed, further enhances the expressiveness of late interaction multi-vector mechanism. In particular, on the BEIR benchmark, our method achieves an average improvement of 1.28\% in Recall@10 in the zero-shot setting using IDF-based weights, and 3.66\% through few-shot fine-tuning.

cs.IR↗

arXiv2023

Learning Arithmetic Formulas in the Presence of Noise: A General Framework and Applications to Unsupervised Learning

We present a general framework for designing efficient algorithms for unsupervised learning problems, such as mixtures of Gaussians and subspace clustering. Our framework is based on a meta algorithm that learns arithmetic circuits in the presence of noise, using lower bounds. This builds upon the recent work of Garg, Kayal and Saha (FOCS 20), who designed such a framework for learning arithmetic circuits without any noise. A key ingredient of our meta algorithm is an efficient algorithm for a novel problem called Robust Vector Space Decomposition. We show that our meta algorithm works well when certain matrices have sufficiently large smallest non-zero singular values. We conjecture that this condition holds for smoothed instances of our problems, and thus our framework would yield efficient algorithms for these problems in the smoothed setting.

cs.DS↗

arXiv2022

Low-depth arithmetic circuit lower bounds via shifted partials

We prove super-polynomial lower bounds for low-depth arithmetic circuits using the shifted partials measure [Gupta-Kamath-Kayal-Saptharishi, CCC 2013], [Kayal, ECCC 2012] and the affine projections of partials measure [Garg-Kayal-Saha, FOCS 2020], [Kayal-Nair-Saha, STACS 2016]. The recent breakthrough work of Limaye, Srinivasan and Tavenas [FOCS 2021] proved these lower bounds by proving lower bounds for low-depth set-multilinear circuits. An interesting aspect of our proof is that it does not require conversion of a circuit to a set-multilinear circuit, nor does it involve a random restriction. We are able to upper bound the measures for homogeneous formulas directly, without going via set-multilinearity. Our lower bounds hold for the iterated matrix multiplication as well as the Nisan-Wigderson design polynomials. We also define a subclass of homogeneous formulas which we call unique parse tree (UPT) formulas, and prove superpolynomial lower bounds for these. This generalizes the superpolynomial lower bounds for regular formulas in [Kayal-Saha-Saptharishi, STOC 2014], [Fournier-Limaye-Malod-Srinivasan, STOC 2014].

cs.CC↗

arXiv2022

OOD-DiskANN: Efficient and Scalable Graph ANNS for Out-of-Distribution Queries

State-of-the-art algorithms for Approximate Nearest Neighbor Search (ANNS) such as DiskANN, FAISS-IVF, and HNSW build data dependent indices that offer substantially better accuracy and search efficiency over data-agnostic indices by overfitting to the index data distribution. When the query data is drawn from a different distribution - e.g., when index represents image embeddings and query represents textual embeddings - such algorithms lose much of this performance advantage. On a variety of datasets, for a fixed recall target, latency is worse by an order of magnitude or more for Out-Of-Distribution (OOD) queries as compared to In-Distribution (ID) queries. The question we address in this work is whether ANNS algorithms can be made efficient for OOD queries if the index construction is given access to a small sample set of these queries. We answer positively by presenting OOD-DiskANN, which uses a sparing sample (1% of index set size) of OOD queries, and provides up to 40% improvement in mean query latency over SoTA algorithms of a similar memory footprint. OOD-DiskANN is scalable and has the efficiency of graph-based ANNS indices. Some of our contributions can improve query efficiency for ID queries as well.

cs.LG↗

arXiv2021

Near-Optimal Lower Bounds For Convex Optimization For All Orders of Smoothness

We study the complexity of optimizing highly smooth convex functions. For a positive integer $p$, we want to find an $\epsilon$-approximate minimum of a convex function $f$, given oracle access to the function and its first $p$ derivatives, assuming that the $p$th derivative of $f$ is Lipschitz. Recently, three independent research groups (Jiang et al., PLMR 2019; Gasnikov et al., PLMR 2019; Bubeck et al., PLMR 2019) developed a new algorithm that solves this problem with $\tilde{O}(1/\epsilon^{\frac{2}{3p+1}})$ oracle calls for constant $p$. This is known to be optimal (up to log factors) for deterministic algorithms, but known lower bounds for randomized algorithms do not match this bound. We prove a new lower bound that matches this bound (up to log factors), and holds not only for randomized algorithms, but also for quantum algorithms.

math.OC↗

arXiv2020

No quantum speedup over gradient descent for non-smooth convex optimization

We study the first-order convex optimization problem, where we have black-box access to a (not necessarily smooth) function $f:\mathbb{R}^n \to \mathbb{R}$ and its (sub)gradient. Our goal is to find an $\epsilon$-approximate minimum of $f$ starting from a point that is distance at most $R$ from the true minimum. If $f$ is $G$-Lipschitz, then the classic gradient descent algorithm solves this problem with $O((GR/\epsilon)^{2})$ queries. Importantly, the number of queries is independent of the dimension $n$ and gradient descent is optimal in this regard: No deterministic or randomized algorithm can achieve better complexity that is still independent of the dimension $n$. In this paper we reprove the randomized lower bound of $\Omega((GR/\epsilon)^{2})$ using a simpler argument than previous lower bounds. We then show that although the function family used in the lower bound is hard for randomized algorithms, it can be solved using $O(GR/\epsilon)$ quantum queries. We then show an improved lower bound against quantum algorithms using a different set of instances and establish our main result that in general even quantum algorithms need $\Omega((GR/\epsilon)^2)$ queries to solve the problem. Hence there is no quantum speedup over gradient descent for black-box first-order convex optimization without further assumptions on the function family.

cs.DS↗

arXiv2020

Towards Stronger Counterexamples to the Log-Approximate-Rank Conjecture

We give improved separations for the query complexity analogue of the log-approximate-rank conjecture i.e. we show that there are a plethora of total Boolean functions on $n$ input bits, each of which has approximate Fourier sparsity at most $O(n^3)$ and randomized parity decision tree complexity $\Theta(n)$. This improves upon the recent work of Chattopadhyay, Mande and Sherif (JACM '20) both qualitatively (in terms of designing a large number of examples) and quantitatively (improving the gap from quartic to cubic). We leave open the problem of proving a randomized communication complexity lower bound for XOR compositions of our examples. A linear lower bound would lead to new and improved refutations of the log-approximate-rank conjecture. Moreover, if any of these compositions had even a sub-linear cost randomized communication protocol, it would demonstrate that randomized parity decision tree complexity does not lift to randomized communication complexity in general (with the XOR gadget).

cs.CC↗

arXiv2020

Scalar Poincar\'e Implies Matrix Poincar\'e

We prove that every reversible Markov semigroup which satisfies a Poincar\'e inequality satisfies a matrix-valued Poincar\'e inequality for Hermitian $d\times d$ matrix valued functions, with the same Poincar\'e constant. This generalizes recent results [Aoun et al. 2019, Kathuria 2019] establishing such inequalities for specific semigroups and consequently yields new matrix concentration inequalities. The short proof follows from the spectral theory of Markov semigroup generators.

math.PR↗

arXiv2020

Learning sums of powers of low-degree polynomials in the non-degenerate case

We develop algorithms for writing a polynomial as sums of powers of low degree polynomials. Consider an $n$-variate degree-$d$ polynomial $f$ which can be written as $$f = c_1Q_1^{m} + \ldots + c_s Q_s^{m},$$ where each $c_i\in \mathbb{F}^{\times}$, $Q_i$ is a homogeneous polynomial of degree $t$, and $t m = d$. In this paper, we give a $\text{poly}((ns)^t)$-time learning algorithm for finding the $Q_i$'s given (black-box access to) $f$, if the $Q_i's$ satisfy certain non-degeneracy conditions and $n$ is larger than $d^2$. The set of degenerate $Q_i$'s (i.e., inputs for which the algorithm does not work) form a non-trivial variety and hence if the $Q_i$'s are chosen according to any reasonable (full-dimensional) distribution, then they are non-degenerate with high probability (if $s$ is not too large). Our algorithm is based on a scheme for obtaining a learning algorithm for an arithmetic circuit model from a lower bound for the same model, provided certain non-degeneracy conditions hold. The scheme reduces the learning problem to the problem of decomposing two vector spaces under the action of a set of linear operators, where the spaces and the operators are derived from the input circuit and the complexity measure used in a typical lower bound proof. The non-degeneracy conditions are certain restrictions on how the spaces decompose.

cs.CC↗

arXiv2019

Towards a theory of non-commutative optimization: geodesic first and second order methods for moment maps and polytopes

This paper initiates a systematic development of a theory of non-commutative optimization. It aims to unify and generalize a growing body of work from the past few years which developed and analyzed algorithms for natural geodesically convex optimization problems on Riemannian manifolds that arise from the symmetries of non-commutative groups. These algorithms minimize the moment map (a non-commutative notion of the usual gradient) and test membership in null cones and moment polytopes (a vast class of polytopes, typically of exponential vertex and facet complexity, which arise from this a priori non-convex, non-linear setting). This setting captures a diverse set of problems in different areas of computer science, mathematics, and physics. Several of them were solved efficiently for the first time using non-commutative methods; the corresponding algorithms also lead to solutions of purely structural problems and to many new connections between disparate fields. In the spirit of standard convex optimization, we develop two general methods in the geodesic setting, a first order and a second order method, which respectively receive first and second order information on the "derivatives" of the function to be optimized. These in particular subsume all past results. The main technical work goes into identifying the key parameters of the underlying group actions which control convergence to the optimum in each of these methods. These non-commutative analogues of "smoothness" are far more complex and require significant algebraic and analytic machinery. Despite this complexity, the way in which these parameters control convergence in both methods is quite simple and elegant. We show how to bound these parameters and hence obtain efficient algorithms for null cone membership in several concrete situations. Our work points to intriguing open problems and suggests further research directions.

math.OC↗

arXiv2019

Search problems in algebraic complexity, GCT, and hardness of generator for invariant rings

We consider the problem of computing succinct encodings of lists of generators for invariant rings for group actions. Mulmuley conjectured that there are always polynomial sized such encodings for invariant rings of $\SL_n(\C)$-representations. We provide simple examples that disprove this conjecture (under standard complexity assumptions). We develop a general framework, denoted \emph{algebraic circuit search problems}, that captures many important problems in algebraic complexity and computational invariant theory. This framework encompasses various proof systems in proof complexity and some of the central problems in invariant theory as exposed by the Geometric Complexity Theory (GCT) program, including the aforementioned problem of computing succinct encodings for generators for invariant rings.

cs.CC↗

arXiv2019

More barriers for rank methods, via a "numeric to symbolic" transfer

We prove new barrier results in arithmetic complexity theory, showing severe limitations of natural lifting (aka escalation) techniques. For example, we prove that even optimal rank lower bounds on $k$-tensors cannot yield non-trivial lower bounds on the rank of $d$-tensors, for any constant $d>k$. This significantly extends recent barrier results on the limits of (matrix) rank methods by Efremenko, Garg, Oliveira and Wigderson, which handles the (very important) case $k=2$. Our generalization requires the development of new technical tools and results in algebraic geometry, which are interesting in their own right and possibly applicable elsewhere. The basic issue they probe is the relation between numeric and symbolic rank of tensors, essential in the proofs of previous and current barriers. Our main technical result implies that for every symbolic $k$-tensor (namely one whose entries are polynomials in some set of variables), if the tensor rank is small for every evaluation of the variables, then it is small symbolically. This statement is obvious for $k=2$. To prove an analogous statement for $k>2$ we develop a "numeric to symbolic" transfer of algebraic relations to algebraic functions, somewhat in the spirit of the implicit function theorem. It applies in the general setting of inclusion of images of polynomial maps, in the form appearing in Raz's elusive functions approach to proving VP $\neq$ VNP. We give a toy application showing how our transfer theorem may be useful in pursuing this approach to prove arithmetic complexity lower bounds.

cs.CC↗

arXiv2018

Recent progress on scaling algorithms and applications

Scaling problems have a rich and diverse history, and thereby have found numerous applications in several fields of science and engineering. For instance, the matrix scaling problem has had applications ranging from theoretical computer science to telephone forecasting, economics, statistics, optimization, among many other fields. Recently, a generalization of matrix scaling known as operator scaling has found applications in non-commutative algebra, invariant theory, combinatorics and algebraic complexity; and a further generalization (tensor scaling) has found more applications in quantum information theory, geometric complexity theory and invariant theory. In this survey, we will describe in detail the scaling problems mentioned above, showing how alternating minimization algorithms naturally arise in this setting, and we shall present a general framework to rigorously analyze such algorithms. These simple problems and algorithms are not just applicable to diverse mathematical and CS areas, but also serve to bring out deep connections between them. As this framework makes extensive use of concepts from invariant theory, we also provide a very gentle introduction to basic concepts of invariant theory and how they are used to analyze alternating minimization algorithms for the scaling problems. This survey is intended for a general computer science audience, and the only background required is basic knowledge of calculus and linear algebra, thereby making it accessible to graduate students and even to advanced undergraduates.

cs.DS↗