arXiv subjects

Anant Khandelwal

Publications and source records attributed to Anant Khandelwal.

At least 19 recordsLinked to original sources

arXiv2026

How Edge of Stability Hinders SCAFFOLD in Federated Optimization

In federated learning, it is well known that heterogeneous data can (in theory) slow down optimization, and much effort has been directed at designing optimization algorithms that are unaffected by data heterogeneity, such as the SCAFFOLD algorithm. Yet, despite strong theoretical guarantees, SCAFFOLD does not usually outperform the much simpler FedAvg in practice. In this work, we propose that this gap is due to the presence of Edge of Stability (EoS) and progressive sharpening in federated optimization, supported by extensive empirical probing. First, we find that EoS-like dynamics occur with both FedAvg and SCAFFOLD under a variety of architectures and hyperparameters. We observe that the equilibrium value of the sharpness is inversely proportional to the learning rate (as in GD), and interestingly, the degree of data heterogeneity (but not the number of local steps) also affects the equilibrium value. Most importantly, we observe that SCAFFOLD's ability to estimate the gradient of the global objective is severely degraded at the EoS, as measured by the correlation between sharpness and SCAFFOLD's error in estimating the global gradient along the optimization trajectory. This suggests a mechanism for SCAFFOLD's lackluster performance in deep learning: with high sharpness at the EoS, SCAFFOLD cannot reliably estimate the global gradient.

cs.LG↗

arXiv2026

Beyond Imitation: Self-Improving Robot Policies via Off-Policy Q-Planning

Behaviour Cloning (BC) has driven remarkable progress in robot manipulation, yet it is fundamentally limited by its inability to self-improve: a policy that fails cannot learn from that failure without additional human demonstrations. Reinforcement Learning fine-tuning offers a path to self-improvement but has proven difficult to scale to the multi-billion-parameter models underpinning modern robot policies. We propose Q-Planning, which equips a large visuomotor BC policy with a small off-policy Q-function. Because a Q-function estimates value rather than imitates actions, it can be trained on the same successful demonstrations as the BC policy and later absorb both successful and failed deployment rollouts, an asymmetry BC does not have. We exploit this asymmetry to enable value-guided action selection at inference (a single-step Q-weighted average over BC draws) and online self-improvement that fine-tunes only the Q-function, leaving the BC weights untouched. On LIBERO and bimanual RoboTwin, ten iterations of self-improvement lift every benchmark score we tested (LIBERO-10 93% to 99%, RoboTwin 83.8% to 91.4%) and shorten successful episodes on the near-ceiling suites (LIBERO-Object, LIBERO-Goal). On two contact-rich bimanual real-robot tasks, the same loop (BC frozen, no human intervention) improves purely from its own deployment rollouts: stack-cups 40% to 90% and insert-wallet 25% to 80% in five iterations, whereas SFT on successful rollouts alone stalls at 55% and 30%. Under an identical online budget Q-Planning is the only method, among Best-of-N, filtered SFT, IBRL, DSRL, and DAWR, that improves stably from failures without training an auxiliary actor.

cs.RO↗

arXiv2026

Read the Trace, Steer the Path: Trajectory-Aware Reinforcement Learning for Diffusion Language Models

Diffusion large language models (dLLMs) generate responses by iteratively unmasking and revising many positions in parallel. This process leaves a rich denoising trace depicting which tokens become confident, which remain unstable, and when commitments form. Existing dLLM reinforcement learning methods use this signal only weakly. Flat rollouts are cheap, but assign a single outcome reward to the whole trajectory. Tree rollouts provide finer, verifiable training signals by branching partial trajectories and propagating leaf rewards upward, but are compute intensive. We ask whether the denoising trace itself can provide tree-like supervision without tree-level compute. We introduce CAPR (Cached-Amortized Path Refinement), a dLLM-RL algorithm that summarizes the denoising trace into a compact path state, uses cached trajectory states to generate cheap sibling continuations, and trains a block-level value head for local block-wise supervision. Under a block-wise unmasking schedule, CAPR records path-state and block-progress features, then redistributes the final outcome reward across blocks according to the tokens revealed in each block. This trains the value head to convert one sparse reward into block-level PPO weights. CAPR therefore recovers much of the granularity of tree search while avoiding full tree expansion, reducing rollout-generation cost to roughly 0.75x that of flat rollouts and 0.6x that of tree rollouts (under standard settings). Across 4x4 Sudoku, Countdown, GSM8K, and Math500, on dense and mixture-of-experts LLaDA backbones, CAPR sets a new state of the art for RL-tuned dLLMs at 256- and 512-token budgets. On Sudoku, it matches the strongest tree-structured baseline at less than one third of the per-step compute.

cs.CL↗

arXiv2026

Brain alignment of reasoning and action representations from vision-language and action models during naturalistic gameplay

Understanding how humans and artificial intelligence systems predict and plan by interacting with their environment is a fundamental challenge at the intersection of neuroscience and machine learning. Most brain-encoding studies focus on aligning artificial models with brain activity during language comprehension or passive visual processing, while interactive brain-alignment studies have to date been largely limited to reinforcement-learning (RL) agents and theory-based models. To address this gap, we study brain alignment of representative models from two foundation-model families, namely vision-language models (VLMs) and large-action models (LAMs), using fMRI recordings from participants playing naturalistic Atari-style video games. Specifically, we examine how action-focused and reasoning-focused prompts shape model's internal representations and align with fMRI brain activity. First, we find that both VLMs and LAMs exhibit significantly exhibit voxel-wise encoding performance than RL baselines, with the advantage holding even under matched feature dimensionality. Second, prompt-driven gains scale with the cortical processing hierarchy: the largest improvements appear in frontal-parietal and motor-planning regions, while early visual cortex gains roughly half as much. Third, variance partitioning reveals a qualitatively different representational organization: VLM is prompt-symmetric (12.5% unique action vs. 13.6% unique reasoning), whereas LAM is prompt-asymmetric (27% unique action vs. -5% unique reasoning), with the asymmetry strongest in frontal-motor cortex. Together, these results demonstrate that action-specialized fine-tuning reorganizes multimodal representations toward action-relevant neural computations even when whole-brain prediction accuracy is statistically equivalent between VLM and LAM.

q-bio.NC↗

arXiv2026

Linguistic properties and model scale in brain encoding: from small to compressed language models

Recent work has shown that scaling large language models (LLMs) improves their alignment with human brain activity, yet it remains unclear what drives these gains and which representational properties are responsible. Although larger models often yield better task performance and brain alignment, they are increasingly difficult to analyze mechanistically. This raises a fundamental question: what is the minimal model capacity required to capture brain-relevant representations? To address this question, we systematically investigate how constraining model scale and numerical precision affects brain alignment. We compare full-precision LLMs, small language models (SLMs), and compressed variants (quantized and pruned) by predicting fMRI responses during naturalistic language comprehension. Across model families up to 14B parameters, we find that 3B SLMs achieve brain predictivity indistinguishable from larger LLMs, whereas 1B models degrade substantially, particularly in semantic language regions. Brain alignment is remarkably robust to compression: most quantization and pruning methods preserve neural predictivity, with GPTQ as a consistent exception. Linguistic probing reveals a dissociation between task performance and brain predictivity: compression degrades discourse, syntax, and morphology, yet brain predictivity remains largely unchanged. Overall, brain alignment saturates at modest model scales and is resilient to compression, challenging common assumptions about neural scaling and motivating compact models for brain-aligned language modeling.

q-bio.NC↗

arXiv2026

How does longer temporal context enhance multimodal narrative video processing in the brain?

Understanding how humans and artificial intelligence systems process complex narrative videos is a fundamental challenge at the intersection of neuroscience and machine learning. This study investigates how the temporal context length of video clips (3--24 s clips) and the narrative-task prompting shape brain-model alignment during naturalistic movie watching. Using fMRI recordings from participants viewing full-length movies, we examine how brain regions sensitive to narrative context dynamically represent information over varying timescales and how these neural patterns align with model-derived features. We find that increasing clip duration substantially improves brain alignment for multimodal large language models (MLLMs), whereas unimodal video models show little to no gain. Further, shorter temporal windows align with perceptual and early language regions, while longer windows preferentially align higher-order integrative regions, mirrored by a layer-to-cortex hierarchy in MLLMs. Finally, experiments with four narrative-task prompts show that they elicit task-specific, region-dependent brain alignment patterns and context-dependent shifts in clip-level tuning in higher-order regions. Our work positions long-form narrative movies as a principled testbed for studying long-timescale temporal integration in long-context MLLMs and its relationship to cortical responses during narrative comprehension.

q-bio.NC↗

arXiv2025

PICore: Physics-Informed Unsupervised Coreset Selection for Data Efficient Neural Operator Training

Neural operators offer a powerful paradigm for solving partial differential equations (PDEs) that cannot be solved analytically by learning mappings between function spaces. However, there are two main bottlenecks in training neural operators: they require a significant amount of training data to learn these mappings, and this data needs to be labeled, which can only be accessed via expensive simulations with numerical solvers. To alleviate both of these issues simultaneously, we propose PICore, an unsupervised coreset selection framework that identifies the most informative training samples without requiring access to ground-truth PDE solutions. PICore leverages a physics-informed loss to select unlabeled inputs by their potential contribution to operator learning. After selecting a compact subset of inputs, only those samples are simulated using numerical solvers to generate labels, reducing annotation costs. We then train the neural operator on the reduced labeled dataset, significantly decreasing training time as well. Across four diverse PDE benchmarks and multiple coreset selection strategies, PICore achieves up to 78% average increase in training efficiency relative to supervised coreset selection methods with minimal changes in accuracy. We provide code at https://github.com/Asatheesh6561/PICore.

cs.LG↗

arXiv2025

HistoryBankQA: Multilingual Temporal Question Answering on Historical Events

Temporal reasoning about historical events is a critical skill for NLP tasks like event extraction, historical entity linking, temporal question answering, timeline summarization, temporal event clustering and temporal natural language inference. Yet efforts on benchmarking temporal reasoning capabilities of large language models (LLMs) are rather limited. Existing temporal reasoning datasets are limited in scale, lack multilingual coverage and focus more on contemporary events. To address these limitations, we present HistoryBank, a multilingual database of 10M+ historical events extracted from Wikipedia timeline pages and article infoboxes. Our database provides unprecedented coverage in both historical depth and linguistic breadth with 10 languages. Additionally, we construct a comprehensive question answering benchmark for temporal reasoning across all languages. This benchmark covers a diverse set of 6 temporal QA reasoning tasks, and we evaluate a suite of popular language models (LLaMA-3-8B, Mistral-7B, Gemma-2-9b, Qwen3-8B, GPT4o) to assess their performance on these tasks. As expected GPT4o performs best across all answer types and languages; Gemma-2 outperforms the other small language models. Our work aims to provide a comprehensive resource for advancing multilingual and temporally-aware natural language understanding of historical events. To facilitate further research, we will make our code and datasets publicly available upon acceptance of this paper.

cs.CL↗

arXiv2025

CoCoA: Confidence and Context-Aware Adaptive Decoding for Resolving Knowledge Conflicts in Large Language Models

Faithful generation in large language models (LLMs) is challenged by knowledge conflicts between parametric memory and external context. Existing contrastive decoding methods tuned specifically to handle conflict often lack adaptability and can degrade performance in low conflict settings. We introduce CoCoA (Confidence- and Context-Aware Adaptive Decoding), a novel token-level algorithm for principled conflict resolution and enhanced faithfulness. CoCoA resolves conflict by utilizing confidence-aware measures (entropy gap and contextual peakedness) and the generalized divergence between the parametric and contextual distributions. Crucially, CoCoA maintains strong performance even in low conflict settings. Extensive experiments across multiple LLMs on diverse Question Answering (QA), Summarization, and Long-Form Question Answering (LFQA) benchmarks demonstrate CoCoA's state-of-the-art performance over strong baselines like AdaCAD. It yields significant gains in QA accuracy, up to 9.2 points on average compared to the strong baseline AdaCAD, and improves factuality in summarization and LFQA by up to 2.5 points on average across key benchmarks. Additionally, it demonstrates superior sensitivity to conflict variations. CoCoA enables more informed, context-aware, and ultimately more faithful token generation.

cs.CL↗

arXiv2025

FlexiClip: Locality-Preserving Free-Form Character Animation

Animating clipart images with seamless motion while maintaining visual fidelity and temporal coherence presents significant challenges. Existing methods, such as AniClipart, effectively model spatial deformations but often fail to ensure smooth temporal transitions, resulting in artifacts like abrupt motions and geometric distortions. Similarly, text-to-video (T2V) and image-to-video (I2V) models struggle to handle clipart due to the mismatch in statistical properties between natural video and clipart styles. This paper introduces FlexiClip, a novel approach designed to overcome these limitations by addressing the intertwined challenges of temporal consistency and geometric integrity. FlexiClip extends traditional Bézier curve-based trajectory modeling with key innovations: temporal Jacobians to correct motion dynamics incrementally, continuous-time modeling via probability flow ODEs (pfODEs) to mitigate temporal noise, and a flow matching loss inspired by GFlowNet principles to optimize smooth motion transitions. These enhancements ensure coherent animations across complex scenarios involving rapid movements and non-rigid deformations. Extensive experiments validate the effectiveness of FlexiClip in generating animations that are not only smooth and natural but also structurally consistent across diverse clipart types, including humans and animals. By integrating spatial and temporal modeling with pre-trained video diffusion models, FlexiClip sets a new standard for high-quality clipart animation, offering robust performance across a wide range of visual content. Project Page: https://creative-gen.github.io/flexiclip.github.io/

cs.CV↗

arXiv2024

FloCoDe: Unbiased Dynamic Scene Graph Generation with Temporal Consistency and Correlation Debiasing

Dynamic scene graph generation (SGG) from videos requires not only a comprehensive understanding of objects across scenes but also a method to capture the temporal motions and interactions with different objects. Moreover, the long-tailed distribution of visual relationships is a crucial bottleneck for most dynamic SGG methods. This is because many of them focus on capturing spatio-temporal context using complex architectures, leading to the generation of biased scene graphs. To address these challenges, we propose FloCoDe: Flow-aware Temporal Consistency and Correlation Debiasing with uncertainty attenuation for unbiased dynamic scene graphs. FloCoDe employs feature warping using flow to detect temporally consistent objects across frames. To address the long-tail issue of visual relationships, we propose correlation debiasing and a label correlation-based loss to learn unbiased relation representations for long-tailed classes. Specifically, we propose to incorporate label correlations using contrastive loss to capture commonly co-occurring relations, which aids in learning robust representations for long-tailed classes. Further, we adopt the uncertainty attenuation-based classifier framework to handle noisy annotations in the SGG data. Extensive experimental evaluation shows a performance gain as high as 4.1%, demonstrating the superiority of generating more unbiased scene graphs.

cs.CV↗

arXiv2024

PromptSync: Bridging Domain Gaps in Vision-Language Models through Class-Aware Prototype Alignment and Discrimination

The potential for zero-shot generalization in vision-language (V-L) models such as CLIP has spurred their widespread adoption in addressing numerous downstream tasks. Previous methods have employed test-time prompt tuning to adapt the model to unseen domains, but they overlooked the issue of imbalanced class distributions. In this study, we explicitly address this problem by employing class-aware prototype alignment weighted by mean class probabilities obtained for the test sample and filtered augmented views. Additionally, we ensure that the class probabilities are as accurate as possible by performing prototype discrimination using contrastive learning. The combination of alignment and discriminative loss serves as a geometric regularizer, preventing the prompt representation from collapsing onto a single class and effectively bridging the distribution gap between the source and test domains. Our method, named PromptSync, synchronizes the prompts for each test sample on both the text and vision branches of the V-L model. In empirical evaluations on the domain generalization benchmark, our method outperforms previous best methods by 2.33% in overall performance, by 1% in base-to-novel generalization, and by 2.84% in cross-dataset transfer tasks.

cs.CV↗

arXiv2024

RePoseDM: Recurrent Pose Alignment and Gradient Guidance for Pose Guided Image Synthesis

Pose-guided person image synthesis task requires re-rendering a reference image, which should have a photorealistic appearance and flawless pose transfer. Since person images are highly structured, existing approaches require dense connections for complex deformations and occlusions because these are generally handled through multi-level warping and masking in latent space. The feature maps generated by convolutional neural networks do not have equivariance, and hence multi-level warping is required to perform pose alignment. Inspired by the ability of the diffusion model to generate photorealistic images from the given conditional guidance, we propose recurrent pose alignment to provide pose-aligned texture features as conditional guidance. Due to the leakage of the source pose in conditional guidance, we propose gradient guidance from pose interaction fields, which output the distance from the valid pose manifold given a predicted pose as input. This helps in learning plausible pose transfer trajectories that result in photorealism and undistorted texture details. Extensive results on two large-scale benchmarks and a user study demonstrate the ability of our proposed approach to generate photorealistic pose transfer under challenging scenarios. Additionally, we demonstrate the efficiency of gradient guidance in pose-guided image generation on the HumanArt dataset with fine-tuned stable diffusion.

cs.CV↗

arXiv2023

InFusion: Inject and Attention Fusion for Multi Concept Zero-Shot Text-based Video Editing

Large text-to-image diffusion models have achieved remarkable success in generating diverse, high-quality images. Additionally, these models have been successfully leveraged to edit input images by just changing the text prompt. But when these models are applied to videos, the main challenge is to ensure temporal consistency and coherence across frames. In this paper, we propose InFusion, a framework for zero-shot text-based video editing leveraging large pre-trained image diffusion models. Our framework specifically supports editing of multiple concepts with pixel-level control over diverse concepts mentioned in the editing prompt. Specifically, we inject the difference in features obtained with source and edit prompts from U-Net residual blocks of decoder layers. When these are combined with injected attention features, it becomes feasible to query the source contents and scale edited concepts along with the injection of unedited parts. The editing is further controlled in a fine-grained manner with mask extraction and attention fusion, which cut the edited part from the source and paste it into the denoising pipeline for the editing prompt. Our framework is a low-cost alternative to one-shot tuned models for editing since it does not require training. We demonstrated complex concept editing with a generalised image model (Stable Diffusion v1.5) using LoRA. Adaptation is compatible with all the existing image diffusion techniques. Extensive experimental results demonstrate the effectiveness of existing methods in rendering high-quality and temporally consistent videos.

cs.CV↗

arXiv2023

SegDA: Maximum Separable Segment Mask with Pseudo Labels for Domain Adaptive Semantic Segmentation

Unsupervised Domain Adaptation (UDA) aims to solve the problem of label scarcity of the target domain by transferring the knowledge from the label rich source domain. Usually, the source domain consists of synthetic images for which the annotation is easily obtained using the well known computer graphics techniques. However, obtaining annotation for real world images (target domain) require lot of manual annotation effort and is very time consuming because it requires per pixel annotation. To address this problem we propose SegDA module to enhance transfer performance of UDA methods by learning the maximum separable segment representation. This resolves the problem of identifying visually similar classes like pedestrian/rider, sidewalk/road etc. We leveraged Equiangular Tight Frame (ETF) classifier inspired from Neural Collapse for maximal separation between segment classes. This causes the source domain pixel representation to collapse to a single vector forming a simplex vertices which are aligned to the maximal separable ETF classifier. We use this phenomenon to propose the novel architecture for domain adaptation of segment representation for target domain. Additionally, we proposed to estimate the noise in labelling the target domain images and update the decoder for noise correction which encourages the discovery of pixels for classes not identified in pseudo labels. We have used four UDA benchmarks simulating synthetic-to-real, daytime-to-nighttime, clear-to-adverse weather scenarios. Our proposed approach outperforms +2.2 mIoU on GTA -> Cityscapes, +2.0 mIoU on Synthia -> Cityscapes, +5.9 mIoU on Cityscapes -> DarkZurich, +2.6 mIoU on Cityscapes -> ACDC.

cs.CV↗

arXiv2023

Large Scale Generative Multimodal Attribute Extraction for E-commerce Attributes

E-commerce websites (e.g. Amazon) have a plethora of structured and unstructured information (text and images) present on the product pages. Sellers often either don't label or mislabel values of the attributes (e.g. color, size etc.) for their products. Automatically identifying these attribute values from an eCommerce product page that contains both text and images is a challenging task, especially when the attribute value is not explicitly mentioned in the catalog. In this paper, we present a scalable solution for this problem where we pose attribute extraction problem as a question-answering task, which we solve using \textbf{MXT}, consisting of three key components: (i) \textbf{M}AG (Multimodal Adaptation Gate), (ii) \textbf{X}ception network, and (iii) \textbf{T}5 encoder-decoder. Our system consists of a generative model that \emph{generates} attribute-values for a given product by using both textual and visual characteristics (e.g. images) of the product. We show that our system is capable of handling zero-shot attribute prediction (when attribute value is not seen in training data) and value-absent prediction (when attribute value is not mentioned in the text) which are missing in traditional classification-based and NER-based models respectively. We have trained our models using distant supervision, removing dependency on human labeling, thus making them practical for real-world applications. With this framework, we are able to train a single model for 1000s of (product-type, attribute) pairs, thus reducing the overhead of training and maintaining separate models. Extensive experiments on two real world datasets show that our framework improves the absolute recall@90P by 10.16\% and 6.9\% from the existing state of the art models. In a popular e-commerce store, we have deployed our models for 1000s of (product-type, attribute) pairs.

cs.CV↗

arXiv2023

DomainInv: Domain Invariant Fine Tuning and Adversarial Label Correction For QA Domain Adaptation

Existing Question Answering (QA) systems limited by the capability of answering questions from unseen domain or any out-of-domain distributions making them less reliable for deployment to real scenarios. Most importantly all the existing QA domain adaptation methods are either based on generating synthetic data or pseudo labeling the target domain data. The domain adaptation methods based on synthetic data and pseudo labeling suffers either from the requirement of computational resources or an extra overhead of carefully selecting the confidence threshold to separate the noisy examples from being in the training dataset. In this paper, we propose the unsupervised domain adaptation for unlabeled target domain by transferring the target representation near to source domain while still using the supervision from source domain. Towards that we proposed the idea of domain invariant fine tuning along with adversarial label correction to identify the target instances which lie far apart from the source domain, so that the feature encoder can be learnt to minimize the distance between such target instances and source instances class wisely, removing the possibility of learning the features of target domain which are still near to source support but are ambiguous. Evaluation of our QA domain adaptation method namely, DomainInv on multiple target QA dataset reveal the performance improvement over the strongest baseline.

cs.CL↗

arXiv2023

MASIL: Towards Maximum Separable Class Representation for Few Shot Class Incremental Learning

Few Shot Class Incremental Learning (FSCIL) with few examples per class for each incremental session is the realistic setting of continual learning since obtaining large number of annotated samples is not feasible and cost effective. We present the framework MASIL as a step towards learning the maximal separable classifier. It addresses the common problem i.e forgetting of old classes and over-fitting to novel classes by learning the classifier weights to be maximally separable between classes forming a simplex Equiangular Tight Frame. We propose the idea of concept factorization explaining the collapsed features for base session classes in terms of concept basis and use these to induce classifier simplex for few shot classes. We further adds fine tuning to reduce any error occurred during factorization and train the classifier jointly on base and novel classes without retaining any base class samples in memory. Experimental results on miniImageNet, CIFAR-100 and CUB-200 demonstrate that MASIL outperforms all the benchmarks.

cs.CV↗