arXiv subjects

Animesh Jha

Publications and source records attributed to Animesh Jha.

9 recordsLinked to original sources

arXiv2026

What to Forget in Unlearning? Forget Set Curation for Language Models

Machine unlearning aims to remove targeted data or behaviors from a trained model without retraining from scratch. Yet most evaluations assume that the examples to forget are already known. In realistic language-model deployments, a requester may ask a model to stop reproducing a song or book without knowing which spans, documents, quotations, or near-duplicates in a trillion-token corpus support that behavior. We study this missing upstream problem, forget set curation: mapping a suppression request to the data passed to an unlearning algorithm. We introduce CleanSlate, a benchmark for verbatim output suppression over songs and books, with model-specific extraction profiles, content-grounded QA, and capability-retention evaluations. CleanSlate exposes two failure modes. Natural lexical and exact-substring curators often yield forget sets that lead to weak suppression. An evaluation-aware curator suppresses requested continuations almost completely, but causes collateral regression on non-requested content and model-dependent capability loss. These results show that practical unlearning is not only an optimization problem once a forget set is given: the data chosen for forgetting determines both what can be unlearnt and what else is damaged.

cs.CL↗

arXiv2026

MAEB: Massive Audio Embedding Benchmark

We introduce the Massive Audio Embedding Benchmark (MAEB), a large-scale benchmark covering 30 tasks across speech, music, environmental sounds, and cross-modal audio-text reasoning in 100+ languages. We evaluate 50+ models and find that no single model dominates across all tasks: contrastive audio-text models excel at environmental sound classification (e.g., ESC50) but score near random on multilingual speech tasks (e.g., SIB-FLEURS), while speech-pretrained models show the opposite pattern. Clustering remains challenging for all models, with even the best-performing model achieving only modest results. We observe that models excelling on acoustic understanding often perform poorly on linguistic tasks, and vice versa. We also show that the performance of audio encoders on MAEB correlates highly with their performance when used in audio large language models. MAEB is derived from MAEB+, a collection of 98 tasks. MAEB is designed to maintain task diversity while reducing evaluation cost, and it integrates into the MTEB ecosystem for unified evaluation across text, image, and audio modalities. We release MAEB and all 98 tasks along with code and a leaderboard at https://github.com/embeddings-benchmark/mteb.

cs.SD↗

arXiv2025

RL-Guided Data Selection for Language Model Finetuning

Data selection for finetuning Large Language Models (LLMs) can be framed as a budget-constrained optimization problem: maximizing a model's downstream performance under a strict training data budget. Solving this problem is generally intractable, and existing approximate approaches are pretraining-oriented and transfer poorly to the fine-tuning setting. We reformulate this problem as a tractable Markov Decision Process (MDP) and train agents using various Reinforcement Learning (RL) methods to learn optimal data selection policies, guided by an efficient, proxy-model-based reward signal. Across four datasets, training on a $5\%$ subset selected by our approach matches or outperforms fine-tuning on the full dataset by up to $10.8$ accuracy points, while cutting wall-clock training time by up to $2 \times$, highlighting the promise of RL-guided data selection.

cs.LG↗

arXiv2025

Certified Unlearning for Neural Networks

We address the problem of machine unlearning, where the goal is to remove the influence of specific training data from a model upon request, motivated by privacy concerns and regulatory requirements such as the "right to be forgotten." Unfortunately, existing methods rely on restrictive assumptions or lack formal guarantees. To this end, we propose a novel method for certified machine unlearning, leveraging the connection between unlearning and privacy amplification by stochastic post-processing. Our method uses noisy fine-tuning on the retain data, i.e., data that does not need to be removed, to ensure provable unlearning guarantees. This approach requires no assumptions about the underlying loss function, making it broadly applicable across diverse settings. We analyze the theoretical trade-offs in efficiency and accuracy and demonstrate empirically that our method not only achieves formal unlearning guarantees but also performs effectively in practice, outperforming existing baselines. Our code is available at https://github.com/stair-lab/certified-unlearning-neural-networks-icml-2025

cs.LG↗

arXiv2022

[Re] Differentiable Spatial Planning using Transformers

This report covers our reproduction effort of the paper 'Differentiable Spatial Planning using Transformers' by Chaplot et al. . In this paper, the problem of spatial path planning in a differentiable way is considered. They show that their proposed method of using Spatial Planning Transformers outperforms prior data-driven models and leverages differentiable structures to learn mapping without a ground truth map simultaneously. We verify these claims by reproducing their experiments and testing their method on new data. We also investigate the stability of planning accuracy with maps with increased obstacle complexity. Efforts to investigate and verify the learnings of the Mapper module were met with failure stemming from a paucity of computational resources and unreachable authors.

cs.RO↗

arXiv2022

Reproducibility Report: Contrastive Learning of Socially-aware Motion Representations

The following paper is a reproducibility report for "Social NCE: Contrastive Learning of Socially-aware Motion Representations" {\cite{liu2020snce}} published in ICCV 2021 as part of the ML Reproducibility Challenge 2021. The original code was made available by the author \footnote{\href{https://github.com/vita-epfl/social-nce}{https://github.com/vita-epfl/social-nce}}. We attempted to verify the results claimed by the authors and reimplemented their code in PyTorch Lightning.

cs.CV↗

arXiv2022

[Reproducibility Report] Path Planning using Neural A* Search

The following paper is a reproducibility report for "Path Planning using Neural A* Search" published in ICML2 2021 as part of the ML Reproducibility Challenge 2021. The original paper proposes the Neural A* planner, and claims it achieves an optimal balance between the reduction of node expansions and path accuracy. We verify this claim by reimplementing the model in a different framework and reproduce the data published in the original paper. We have also provided a code-flow diagram to aid comprehension of the code structure. As extensions to the original paper, we explore the effects of (1) generalizing the model by training it on a shuffled dataset, (2) introducing dropout, (3) implementing empirically chosen hyperparameters as trainable parameters in the model, (4) altering the network model to Generative Adversarial Networks (GANs) to introduce stochasticity, (5) modifying the encoder from Unet to Unet++, (6) incorporating cost maps obtained from the Neural A* module in other variations of A* search.

cs.AI↗

arXiv2021

Local NMPC on Global Optimised Path for Autonomous Racing

The paper presents a strategy for the control of anautonomous racing car on a pre-mapped track. Using a dynamic model of the vehicle, the optimal racing line is computed, taking track boundaries into account. With the optimal racing line as areference, a local nonlinear model predictive controller (NMPC) is proposed, which takes into account multiple local objectives like making more progress along the race line, avoiding collision with opponent vehicles, and use of drafting to achieve more progress.

cs.RO↗

arXiv2011

Three-dimensional mid-infrared photonic circuits in chalcogenide glass

We report the fabrication of single mode buried channel waveguides for the whole mid-infrared transparency range of chalcogenide sulphide glasses by means of direct laser writing. We have explored the potential of this technology by fabricating a prototype three-dimensional three-beam combiner for future application in stellar interferometry, which delivers a monochromatic interference visibility of 99.89% at 10.6 \mum, and an ultrahigh bandwidth (3-11 \mum) interference visibility of 21.3%. These results demonstrate that it is possible to harness the whole transparency range offered by chalcogenide glasses on a single on-chip instrument by means of direct laser writing, a finding that may be of key significance in future technologies such as astrophotonics and biochemical sensing.

physics.optics↗