Akshayaa Magesh

Akshayaa Magesh, Venugopal V. Veeravalli, Anirban Roy et al.

We study the problem of Out-of-Distribution (OOD) detection, that is, detecting whether a learning algorithm's output can be trusted at inference time. While a number of tests for OOD detection have been proposed in prior work, a formal framework for studying this problem is lacking. We propose a definition for the notion of OOD that includes both the input distribution and the learning algorithm, which provides insights for the construction of powerful tests for OOD detection. We propose a multiple hypothesis testing inspired procedure to systematically combine any number of different statistics from the learning algorithm using conformal p-values. We further provide strong guarantees on the probability of incorrectly classifying an in-distribution sample as OOD. In our experiments, we find that threshold-based tests proposed in prior work perform well in specific settings, but not uniformly well across different types of OOD instances. In contrast, our proposed method that combines multiple statistics performs uniformly well across different datasets and neural networks.

Adarsh M. Subramaniam, Akshayaa Magesh, Venugopal V. Veeravalli

Compressed Stochastic Gradient Descent (SGD) algorithms have been recently proposed to address the communication bottleneck in distributed and decentralized optimization problems, such as those that arise in federated machine learning. Existing compressed SGD algorithms assume the use of non-adaptive step-sizes(constant or diminishing) to provide theoretical convergence guarantees. Typically, the step-sizes are fine-tuned in practice to the dataset and the learning algorithm to provide good empirical performance. Such fine-tuning might be impractical in many learning scenarios, and it is therefore of interest to study compressed SGD using adaptive step-sizes. Motivated by prior work on adaptive step-size methods for SGD to train neural networks efficiently in the uncompressed setting, we develop an adaptive step-size method for compressed SGD. In particular, we introduce a scaling technique for the descent step in compressed SGD, which we use to establish order-optimal convergence rates for convex-smooth and strong convex-smooth objectives under an interpolation condition and for non-convex objectives under a strong growth condition. We also show through simulation examples that without this scaling, the algorithm can fail to converge. We present experimental results on deep neural networks for real-world datasets, and compare the performance of our proposed algorithm with previously proposed compressed SGD methods in literature, and demonstrate improved performance on ResNet-18, ResNet-34 and DenseNet architectures for CIFAR-100 and CIFAR-10 datasets at various levels of compression.

Ankur Samanta, Akshayaa Magesh, Ayush Jain et al.

Contemporary reinforcement learning with verifiable reward methods post-train language models on multi-step reasoning by assigning a single outcome reward uniformly across all tokens in a trajectory. Such uniform assignment ignores which steps contributed to success or failure. Improving credit assignment can address this limitation by enabling targeted refinement of faulty reasoning steps, rather than updating entire trajectories uniformly. Resets are one such simple mechanism, enabling more precise credit assignment by returning to an intermediate state and resampling counterfactual continuations, so that outcome differences can be attributed to decisions made at that point. We propose two such methods: Random-Reset Policy Optimization (RRPO), where reset states are drawn randomly from reasoning steps, and Self-Reset Policy Optimization (SRPO), where the model self-localizes the erroneous step in an incorrect trajectory and resets there. We analyze these methods within the Conservative Policy Iteration (CPI) framework. Extending CPI with a credit-assignment oracle that targets improvable states yields provable improvements over random resets. Across models and reasoning benchmarks, SRPO consistently outperforms standard GRPO and RRPO by sampling multiple suffix continuations at a self-localized reset and learning from their rewards, using only the model itself with no external supervision.

Ankur Samanta, Akshayaa Magesh, Ayush Jain et al.

Self-correction in language models remains elusive. In this work, we explore whether language models can explicitly localize errors in incorrect reasoning, as a path toward building AI systems that can effectively correct themselves. We introduce a prompting method that structures reasoning as discrete, semantically coherent thought steps, and show that models are able to reliably localize errors within this structure, while failing to do so in conventional, unstructured chain-of-thought reasoning. Motivated by how the human brain monitors errors at discrete decision points and resamples alternatives, we introduce Iterative Correction Sampling of Thoughts (Thought-ICS), a self-correction framework. Thought-ICS iteratively prompts the model to generate reasoning one discrete and complete thought at a time--where each thought represents a deliberate decision by the model--creating natural boundaries for precise error localization. Upon verification, the model localizes the first erroneous step, and the system backtracks to generate alternative reasoning from the last correct point. When asked to correct reasoning verified as incorrect by an oracle, Thought-ICS achieves 20-40% self-correction lift. In a completely autonomous setting without external verification, it outperforms contemporary self-correction baselines.

Ankur Samanta, Akshayaa Magesh, Youliang Yu et al.

Language Models (LMs) are inconsistent reasoners, often generating contradictory responses to identical prompts. While inference-time methods can mitigate these inconsistencies, they fail to address the core problem: LMs struggle to reliably select reasoning pathways leading to consistent outcomes under exploratory sampling. To address this, we formalize self-consistency as an intrinsic property of well-aligned reasoning models and introduce Multi-Agent Consensus Alignment (MACA), a reinforcement learning framework that post-trains models to favor reasoning trajectories aligned with their internal consensus using majority/minority outcomes from multi-agent debate. These trajectories emerge from deliberative exchanges where agents ground reasoning in peer arguments, not just aggregation of independent attempts, creating richer consensus signals than single-round majority voting. MACA enables agents to teach themselves to be more decisive and concise, and better leverage peer insights in multi-agent settings without external supervision, driving substantial improvements across self-consistency (+27.6% on GSM8K), single-agent reasoning (+23.7% on MATH), sampling-based inference (+22.4% Pass@20 on MATH), and multi-agent ensemble decision-making (+42.7% on MathQA). These findings, coupled with strong generalization to unseen benchmarks (+16.3% on GPQA, +11.6% on CommonsenseQA), demonstrate robust self-alignment that more reliably unlocks latent reasoning potential of language models.

Jiawei Li, Akshayaa Magesh, Venugopal V. Veeravalli

While Large Language Models (LLMs) have emerged as powerful foundational models to solve a variety of tasks, they have also been shown to be prone to hallucinations, i.e., generating responses that sound confident but are actually incorrect or even nonsensical. In this work, we formulate the problem of detecting hallucinations as a hypothesis testing problem and draw parallels to the problem of out-of-distribution detection in machine learning models. We propose a multiple-testing-inspired method to solve the hallucination detection problem, and provide extensive experimental results to validate the robustness of our approach against state-of-the-art methods.

Akshayaa Magesh, Venugopal V. Veeravalli

We consider a multi-player multi-armed bandit setting in the presence of adversaries that attempt to negatively affect the rewards received by the players in the system. The reward distributions for any given arm are heterogeneous across the players. In the event of a collision (more than one player choosing the same arm), all the colliding users receive zero rewards. The adversaries use collisions to affect the rewards received by the players, i.e., if an adversary attacks an arm, any player choosing that arm will receive zero reward. At any time step, the adversaries may attack more than one arm. It is assumed that the players in the system do not deviate from a pre-determined policy used by all the players, and that the probability that none of the arms face adversarial attacks is strictly positive at every time step. In order to combat the adversarial attacks, the players are allowed to communicate using a single bit for $O(\log T)$ time units, where $T$ is the time horizon, and each player can only observe their own actions and rewards at all time steps. We propose a {policy that is used by all the players, which} achieves near order optimal regret of order $O(\log^{1+δ}T + W)$, where $W$ is total number of time units for which there was an adversarial attack on at least one arm.

Sourya Basu, Akshayaa Magesh, Harshit Yadav et al.

Recent works show that group equivariance as an inductive bias improves neural network performance for both classification and generation. However, designing group-equivariant neural networks is challenging when the group of interest is large and is unknown. Moreover, inducing equivariance can significantly reduce the number of independent parameters in a network with fixed feature size, affecting its overall performance. We address these problems by proving a new group-theoretic result in the context of equivariant neural networks that shows that a network is equivariant to a large group if and only if it is equivariant to smaller groups from which it is constructed. Using this result, we design a novel fast group equivariant construction algorithm, and a deep Q-learning-based search algorithm in a reduced search space, yielding what we call autoequivariant networks (AENs). AENs find the right balance between equivariance and network size when tested on new benchmark datasets, G-MNIST and G-Fashion-MNIST, obtained via group transformations on MNIST and Fashion-MNIST respectively that we release. Extending these results to group convolutional neural networks, where we optimize between equivariances, augmentations, and network sizes, we find group equivariance to be the most dominating factor in all high-performing GCNNs on several datasets like CIFAR10, SVHN, RotMNIST, ASL, EMNIST, and KMNIST.

Meghana Bande, Akshayaa Magesh, Venugopal V. Veeravalli

A stochastic multi-user multi-armed bandit framework is used to develop algorithms for uncoordinated spectrum access. In contrast to prior work, it is assumed that rewards can be non-zero even under collisions, thus allowing for the number of users to be greater than the number of channels. The proposed algorithm consists of an estimation phase and an allocation phase. It is shown that if every user adopts the algorithm, the system wide regret is order-optimal of order $O(\log T)$ over a time-horizon of duration $T$. The regret guarantees hold for both the cases where the number of users is greater than or less than the number of channels. The algorithm is extended to the dynamic case where the number of users in the system evolves over time, and is shown to lead to sub-linear regret.

Akshayaa Magesh, Venugopal V. Veeravalli

Multi-user multi-armed bandits have emerged as a good model for uncoordinated spectrum access problems. In this paper we consider the scenario where users cannot communicate with each other. In addition, the environment may appear differently to different users, ${i.e.}$, the mean rewards as observed by different users for the same channel may be different. With this setup, we present a policy that achieves a regret of $O (\log{T})$. This paper has been accepted at Asilomar Conference on Signals, Systems, and Computers 2019.

Akshayaa Magesh, Venugopal V. Veeravalli

We consider a fully decentralized multi-player stochastic multi-armed bandit setting where the players cannot communicate with each other and can observe only their own actions and rewards. The environment may appear differently to different players, $\textit{i.e.}$, the reward distributions for a given arm are heterogeneous across players. In the case of a collision (when more than one player plays the same arm), we allow for the colliding players to receive non-zero rewards. The time-horizon $T$ for which the arms are played is \emph{not} known to the players. Within this setup, where the number of players is allowed to be greater than the number of arms, we present a policy that achieves near order-optimal expected regret of order $O(\log^{1 + δ} T)$ for some $0 < δ< 1$ over a time-horizon of duration $T$. This paper is accepted at IEEE Transactions on Information Theory.