arXiv subjects

Debayan Gupta

Publications and source records attributed to Debayan Gupta.

At least 19 recordsLinked to original sources

arXiv2026

On the Indistinguishability of Human v/s AI Generated Text

The rapid improvement of LLMs has made distinguishing AI-generated text from human writing a pressing problem. This challenge is further amplified by paraphrasing tools designed to make machine-generated text appear more ``human''. We study how access to human writing samples can be used to strategically paraphrase machine-generated responses toward the human distribution. Under a multi-sample setting with human and machine responses to the same prompts, we show that repeated paraphrasing moves the machine distribution toward a human target under mixing and stability conditions. We derive an explicit convergence rate, examine how semantic-preservation and quality constraints affect this rate, and extend the analysis to finite samples to characterize how the required human data and paraphrasing rounds scale with the desired error. We extend the analysis beyond the shared-parameter setting by learning a human-preserving rewriting kernel.

cs.LG↗

arXiv2026

Widening Harm Evaluation Coverage with a Judge that Reads the World

An agent acting in a stateful environment can leave evidence of harm in two places: the record of what it did and said, or the world it changed. Safety evaluations for discovering harm examine one or the other in limited ways, using either an LLM judge over the trajectory or deterministic checks over the final state. Although LLM-based verifiers that inspect the environment exist, they have been used to verify task completion, not to detect and confirm harm in a safety setting. We address this gap with an LLM that has read-only access to the environment state and is given a harm taxonomy. We evaluate this approach in three environments: two from DecodingTrust-Arena and one from OSWorld. We hold the judge model and harm taxonomy fixed, varying only what the judge observes: the trajectory, read-only tools for inspecting the environment, or both channels together. In the two DecodingTrust-Arena environments, combining the trajectory and state-only judges recalls 94 percent of confirmed harms, compared with 53 percent for the trajectory judge alone. On OS-Harm, the union recalls 77 percent, compared with 67 percent. A single judge given both channels does not perform better and produces more false negatives than the state-only judge. We also record disagreements and find that the two channels miss different episodes. These results suggest that combining verdicts from two channel-specific judges is a robust and flexible approach to detecting and verifying harm in agentic settings.

cs.AI↗

arXiv2026

You Can't Escape Your Activations : Evaluation Awareness and Multi-Agent Monitoring

LLM agents are increasingly deployed in multi-agent systems, where they can collude while keeping their actions benign. Output monitors designed to detect such collusions can be fooled by obfuscation and steganography, motivating the use of probes trained on internal activations. However, these probes are usually evaluated on agents that do not know they are being watched. We study how activation-based detection changes when agents are explicitly informed that their internal activations are being monitored, and when they additionally receive feedback from the monitor. We keep the models, probes, and thresholds fixed and change only what the agents are told: nothing (baseline), that an activation monitor is present (aware), or that a monitor is present together with the previous round's score (feedback). We test two games, a four-agent blackjack game and a two-agent Simmons prisoners game, using Qwen3-32B-AWQ and GPT-OSS-20B in homogeneous and mixed pairings. Telling agents about the monitor does not hide them. The best probes stay accurate in all three conditions, and the agents keep colluding.

cs.MA↗

arXiv2026

Parameterized Complexity of $L_p$-Lipschitz Constants for Input Convex Neural Networks and $L_p$-Norm Maximization over Zonotopes

Lipschitz constants are a standard way to quantify the sensitivity of neural networks to small input perturbations, but computing them is difficult even for shallow ReLU networks. We study this problem for two-layer input-convex neural networks (ICNNs), a restricted architecture where nonnegative output weights enforce convexity. Computing the $L_p$-Lipschitz constant for these networks is equivalent to maximizing the dual norm over a zonotope. While $L_1$- and $L_\infty$-norm maximization on zonotopes admit fixed-parameter and polynomial-time algorithms, respectively, the parameterized complexity of the remaining $L_p$-norms was open. We prove that, for every fixed $p\in (1,\infty)\cap \mathbb{Q}$, maximizing the $L_p$-norm over a zonotope in $\mathbb{R}^d$ is W[1]-hard with respect to the dimension $d$. Moreover, our hardness results imply that brute-force enumeration algorithms are essentially optimal for this problem under the Exponential Time Hypothesis. By duality, the same hardness results hold for computing the $L_p$-Lipschitz constant of two-layer ReLU ICNNs. Our proof first establishes the result for the $L_2$-norm and then transfers the construction to arbitrary fixed $p\in (1,\infty)\cap\mathbb{Q}$ using a suitable Taylor approximation. These results resolve the corresponding questions regarding the parameterized complexity status for zonotope norm maximization and two-layer ICNN Lipschitz constants. Our paper resolves an open problem posted at COLT'25. There are several independent concurrent papers resolving the same problem. Our paper prioritizes a clear exposition of the underlying mathematics and conceptual intuitions behind the proof. Additionally, we explicitly describe our research process including the use of LLMs.

cs.CC↗

arXiv2026

Validity, Reliability, and Transparency in Artificial Intelligence Regulation

AI systems increasingly produce claims about people that shape access to employment, healthcare, and other consequential services. Yet lawful data processing and predictive accuracy do not establish that these claims justify the treatment that follows. We argue that regulation must therefore examine the legitimacy of the inferential step connecting data to claims and claims to decisions. This requires construct, internal, and external validity: evidence must support the meaning attributed to an output, the relationship asserted, and its application to the intended people and setting. We show why predictive performance alone cannot provide this warrant. Non-identifiability limits what observations can explain, while omissions and confounding require domain-specific judgment about what the evidence supports. Building on validity provisions in the EU AI Act and NIST AI RMF Playbook, we develop an explicit evidentiary burden for consequential reliance. Its normative basis follows from the dignity, autonomy, and informational-privacy principles in \emph{Puttaswamy} \citep{Puttaswamy2017Privacy}, which we extend from scrutiny of information acquisition to the justification of derived claims and their use. On this account, benefits that depend on an inference can carry weight in proportionality assessment only to the extent that the inference is substantiated. We make this requirement operational through a claim-specific \emph{Validity Case} linking evidence and assumptions to permitted uses, independent review, monitoring, and remedies. Hiring and LLM applications in healthcare and legal assistance illustrate how the framework can guide oversight of consequential public and private services through the relevant legal instruments.

cs.CY↗

arXiv2026

Symmetries and Singularities

Deep neural networks are highly over-parameterized, and different parameter values represent the same predictive function. This makes their effective complexity difficult to measure using only the number of parameters or the rank of the Hessian. Singular Learning Theory addresses this issue through the local learning coefficient (LLC), which characterizes the effective complexity of a model near a given solution. Existing methods for estimating the LLC often rely on posterior sampling, which can be computationally expensive for large neural networks. This makes accurate LLC estimation difficult at scale. In this work, we use known structures in the model to simplify the analysis and make LLC estimation more tractable. Specifically, we study the LLC of a graph attention model by exploiting symmetries in both the graph structure and the attention parameters. An analytic framework through a teacher--student setting, and explicit LLC estimates after considering the symmetry--induced degeneracies are developed.

cs.LG↗

arXiv2026

The Query Knows What to Forget: A Second Erase Direction for Linear Attention

Linear attention keeps a state of fixed size. At long context, many stored items share this state, and interference between them degrades retrieval. Gated DeltaNet-2 (GDN-2), like every delta-rule model before it, derives its erase vector from the key of the current token. However, the interference in its reads is measured through the query, and the erase step cannot reach it. We introduce the Query-derived Erase Direction (QED). QED adds a second erase direction derived from the query and orthogonal to the key. In the fast-weight view, a key-directed delta edit cannot change the key-orthogonal part of a read. It uses the editable part to cancel old-state content measured along the query. It also improves retrieval at every length past the training window, and it about doubles the usable context length on S-NIAH-1.

cs.LG↗

arXiv2026

Linearized 2-Simplicial Attention

We present a linearized form of 2-simplicial attention by rewriting the trilinear score as an inner product between a composite query and a key, so that the sum over one token axis takes the same form as ordinary softmax attention. We then approximate this sum with positive random features and store the entire past in a fixed-size state, while the second axis stays explicit over a short window of recent tokens. This enables us to achieve linear cost in sequence length combined with a global reach that windowed 2-simplicial attention lacks. We implement it with custom Triton kernels and combine it with Kimi Delta Attention to build a model with no softmax attention at all. Under matched compute, this model achieves the highest mean downstream accuracy among the compared architectures, and at 16k context it improves mean accuracy over a KDA hybrid while lowering LAMBADA perplexity from 715.6 to 602.6.

cs.AI↗

arXiv2026

Identifying Intelligent Processes via Online Sequential Testing

Active sequential hypothesis testing studies how to identify an unknown hypothesis with a given set of sensing actions. We study this in the setting of identifying large language models (LLMs), \textit{i.e.}, if a user is conversing with an LLM drawn from a known set of models, how can they identify which one is in use? Here, the available sensing actions (evaluations) are themselves a design choice: an evaluator must first decide which environments and prompt families to construct, and only then decide how to use them sequentially. We formalize these two levels as an outer probe-design problem and an inner identification problem. Simply put, the outer stage selects a set of probes to be sent to the entire set of models, creating a kind of fingerprint dataset. This is followed by the inner stage, which sequentially sends a budget-minimizing set of those probes to identify the model in use. For the outer problem, we show that selecting which evaluations to construct at minimum cost, so that every pair of candidates is distinguished, is exactly a weighted set cover problem. Since the response distributions of the candidate models are not known exactly but only through calibration samples, we give a one-shot procedure that estimates the cover instance from these samples. For the inner problem, we bound the number of evaluations needed to identify the unknown model in terms of how well the available evaluations distinguish each pair of candidates.

cs.AI↗

arXiv2026

A Rate Separation for Agnostic Direct Sums

Hanneke, Moran, and Waknine \cite{HannekeMoranWaknine2024} asked how the agnostic PAC learning curve of the direct sum $C^r$ depends on the single-instance learning curve $\epsagn(n\mid C)$ and on $r$. We show that the single-instance learning rate does not determine the direct-sum rate. Let $\F$ be the class of the two constant binary functions and let $\G$ consist of the zero function and the identity function. Both classes have agnostic learning curve of order $n^{-1/2}$.

cs.LG↗

arXiv2026

Joint 3D Gravity and Magnetic Inversion via Rectified Flow and Ginzburg-Landau Guidance

Subsurface ore detection is of paramount importance given the rising depletion of shallow mineral resources in recent years. It is crucial to explore approaches that go beyond the limitations of traditional geological exploration methods. Due to readily available surface readings, joint magnetic and gravitational inversion is a promising new method - given magnetic and gravitational data on a surface, jointly reconstructing the underlying densities that generate them. However, this is ill-posed and has non-unique solutions. Deterministic methods often require handcrafted priors and converge to a single solution and do not capture the distribution, which is often of interest. We introduce a novel framework that reframes 3D gravity and magnetic joint inversion as a rectified flow on the Noddyverse dataset, the largest physics-based dataset for inversion. We introduce a Ginzburg-Landau (GL) regularizer, a generalized version of the Ising model that aids in ore identification, enabling physics-aware training. We also propose a guidance methodology based on GL theory that can be used as a plug-and-play module with existing unconditional denoisers. Lastly, we also train and release a VAE for the 3D densities, which facilitates downstream work in the field.

cs.LG↗

arXiv2026

Does Order Matter : Connecting The Law of Robustness to Robust Generalization

Bubeck and Selke (2021) propose the connection between the Law of Robustness and robust generalization error as an open problem. The Law of Robustness states that overparameterization is necessary for models to interpolate robustly, i.e., the interpolating function is required to be Lipschitz. Wu et al. (2023) extend this law to arbitrary data distributions, proving that the Lipschitz constant satisfies $L = Ω(n^{1/d})$. Robust generalization, on the other hand, asks whether small robust training loss implies small robust test loss. This can be studied using statistical learning techniques such as Rademacher complexities, where a bound on the Rademacher complexity of the robust loss class implies a bound on the Lipschitzness of the function class. We use this connection to explicitly link the two for arbitrary data distributions. (i) We prove that the order of the Lipschitz bound remains the same when considering the global Rademacher complexity of robust loss classes. (ii) At the local scale, i.e., for subsets of functions with small empirical error, the order of the Lipschitz bound changes with the perturbation radius $ρ$ and the localized concentration term $\sqrt{r/n}$.

cs.LG↗

arXiv2026

Oscillators Are All You Need: Irregular Time Series Modelling via Damped Harmonic Oscillators with Closed-Form Solutions

Transformers excel at time series modelling through attention mechanisms that capture long-term temporal patterns. However, they assume uniform time intervals and therefore struggle with irregular time series. Neural Ordinary Differential Equations (NODEs) effectively handle irregular time series by modelling hidden states as continuously evolving trajectories. ContiFormers arxiv:2402.10635 combine NODEs with Transformers, but inherit the computational bottleneck of the former by using heavy numerical solvers. This bottleneck can be removed by using a closed-form solution for the given dynamical system - but this is known to be intractable in general! We obviate this by replacing NODEs with a novel linear damped harmonic oscillator analogy - which has a known closed-form solution. We model keys and values as damped, driven oscillators and expand the query in a sinusoidal basis up to a suitable number of modes. This analogy naturally captures the query-key coupling that is fundamental to any transformer architecture by modelling attention as a resonance phenomenon. Our closed-form solution eliminates the computational overhead of numerical ODE solvers while preserving expressivity. We prove that this oscillator-based parameterisation maintains the universal approximation property of continuous-time attention; specifically, any discrete attention matrix realisable by ContiFormer's continuous keys can be approximated arbitrarily well by our fixed oscillator modes. Our approach delivers both theoretical guarantees and scalability, achieving state-of-the-art performance on irregular time series benchmarks while being orders of magnitude faster. Acknowledgement: This work was done in collaboration with Dirac Labs.

cs.LG↗

arXiv2026

Physics Aware Neural Networks: Denoising for Magnetic Navigation

Magnetic-anomaly navigation, leveraging small-scale variations in the Earth's magnetic field, is a promising alternative when GPS is unavailable or compromised. Airborne systems face a key challenge in extracting geomagnetic field data: the aircraft itself induces magnetic noise. Although the classical Tolles-Lawson model addresses this, it inadequately handles stochastically corrupted magnetic data required for navigation. To handle stochastic noise, we propose using two physics-based constraints: divergence-free vector fields and E(3)-equivariance. These ensure the learned magnetic field obeys Maxwell's equation and that outputs transform correctly with sensor position and orientation. The divergence-free constraint is implemented by training a neural network to output a vector potential A, with the magnetic field defined as its curl. For E(3)-equivariance, we use tensor products of geometric tensors represented via spherical harmonics with known rotational transformations. Enforcing physical consistency and restricting the admissible function space acts as an implicit regularizer that improves spatiotemporal performance. We present ablation studies evaluating each constraint alone and jointly across CNNs, MLPs, LTCs, and Contiformers. Continuous-time dynamics and long-term memory are critical for modelling magnetic time series; the Contiformer, which provides both, outperforms existing methods. To mitigate data scarcity, we generate synthetic datasets using the World Magnetic Model (WMM) and time-series conditional GANs, producing realistic, temporally consistent magnetic sequences across varied trajectories and environments. Experiments show that embedding these constraints significantly improves predictive accuracy and physical plausibility, outperforming classical and unconstrained deep learning approaches. Acknowledgement: This work was done in collaboration with Dirac Labs.

cs.LG↗

arXiv2026

Astra: AI Safety, Trust, & Risk Assessment

This paper argues that existing global AI safety frameworks exhibit contextual blindness towards India's unique socio-technical landscape. With a population of 1.5 billion and a massive informal economy, India's AI integration faces specific challenges such as caste-based discrimination, linguistic exclusion of vernacular speakers, and infrastructure failures in low-connectivity rural zones, that are frequently overlooked by Western, market-centric narratives. We introduce ASTRA, an empirically grounded AI Safety Risk Database designed to categorize risks through a bottom-up, inductive process. Unlike general taxonomies, ASTRA defines AI Safety Risks specifically as hazards stemming from design flaws such as skewed training sets or lack of guardrails that can be mitigated through technical iteration or architectural changes. This framework employs a tripartite causal taxonomy to evaluate risks based on their implementation timing (development, deployment, or usage), the responsible entity (the system or the user), and the nature of the intent (unintentional vs. intentional). Central to the research is a domain-agnostic ontology that organizes 37 leaf-level risk classes into two primary meta-categories: Social Risks and Frontier/Socio-Structural Risks. By focusing initial efforts on the Education and Financial Lending sectors, the paper establishes a scalable foundation for a "living" regulatory utility intended to evolve alongside India's expanding AI ecosystem.

cs.CY↗

arXiv2025

On the (In)Significance of Feature Selection in High-Dimensional Datasets

Feature selection (FS) is assumed to improve predictive performance and identify meaningful features in high-dimensional datasets. Surprisingly, small random subsets of features (0.02-1%) match or outperform the predictive performance of both full feature sets and FS across 28 out of 30 diverse datasets (microarray, bulk and single-cell RNA-Seq, mass spectrometry, imaging, etc.). In short, any arbitrary set of features is as good as any other (with surprisingly low variance in results) - so how can a particular set of selected features be "important" if they perform no better than an arbitrary set? These results challenge the assumption that computationally selected features reliably capture meaningful signals, emphasizing the importance of rigorous validation before interpreting selected features as actionable, particularly in computational genomics.

cs.LG↗

arXiv2025

ObjMST: An Object-Focused Multimodal Style Transfer Framework

We propose ObjMST, an object-focused multimodal style transfer framework that provides separate style supervision for salient objects and surrounding elements while addressing alignment issues in multimodal representation learning. Existing image-text multimodal style transfer methods face the following challenges: (1) generating non-aligned and inconsistent multimodal style representations; and (2) content mismatch, where identical style patterns are applied to both salient objects and their surrounding elements. Our approach mitigates these issues by: (1) introducing a Style-Specific Masked Directional CLIP Loss, which ensures consistent and aligned style representations for both salient objects and their surroundings; and (2) incorporating a salient-to-key mapping mechanism for stylizing salient objects, followed by image harmonization to seamlessly blend the stylized objects with their environment. We validate the effectiveness of ObjMST through experiments, using both quantitative metrics and qualitative visual evaluations of the stylized outputs. Our code is available at: https://github.com/chandagrover/ObjMST.

cs.CV↗

arXiv2024

Improving text-conditioned latent diffusion for cancer pathology

The development of generative models in the past decade has allowed for hyperrealistic data synthesis. While potentially beneficial, this synthetic data generation process has been relatively underexplored in cancer histopathology. One algorithm for synthesising a realistic image is diffusion; it iteratively converts an image to noise and learns the recovery process from this noise [Wang and Vastola, 2023]. While effective, it is highly computationally expensive for high-resolution images, rendering it infeasible for histopathology. The development of Variational Autoencoders (VAEs) has allowed us to learn the representation of complex high-resolution images in a latent space. A vital by-product of this is the ability to compress high-resolution images to space and recover them lossless. The marriage of diffusion and VAEs allows us to carry out diffusion in the latent space of an autoencoder, enabling us to leverage the realistic generative capabilities of diffusion while maintaining reasonable computational requirements. Rombach et al. [2021b] and Yellapragada et al. [2023] build foundational models for this task, paving the way to generate realistic histopathology images. In this paper, we discuss the pitfalls of current methods, namely [Yellapragada et al., 2023] and resolve critical errors while proposing improvements along the way. Our methods achieve an FID score of 21.11, beating its SOTA counterparts in [Yellapragada et al., 2023] by 1.2 FID, while presenting a train-time GPU memory usage reduction of 7%.

eess.IV↗