Purbesh Mitra

Purbesh Mitra, Sennur Ulukus

Long context reasoning in large language models (LLMs) has demonstrated enhancement of their cognitive capabilities via chain-of-thought (CoT) inference. Training such models is usually done via reinforcement learning with verifiable rewards (RLVR) in reasoning based problems, like math and programming. However, RLVR is limited by several bottlenecks, such as, lack of dense reward, and inadequate sample efficiency. As a result, it requires significant compute resources in post-training phase. To overcome these limitations, in this work, we propose \textbf{Semantic Soft Bootstrapping (SSB)}, a self-distillation technique, in which the same base language model plays the role of both teacher and student, but receives different semantic contexts about the correctness of its outcome at training time. The model is first prompted with a math problem and several rollouts are generated. From them, the correct and most common incorrect response are filtered, and then provided to the model in context to produce a more robust, step-by-step explanation with a verified final answer. This pipeline automatically curates a paired teacher-student training set from raw problem-answer data, without any human intervention. This generation process also produces a sequence of logits, which is what the student model tries to match in the training phase just from the bare question alone. In our experiment, Qwen2.5-3B-Instruct on GSM8K dataset via parameter-efficient fine-tuning. We then tested its accuracy on MATH500, and AIME2024 benchmarks. Our experiments show a jump of 10.6%, and 10% improvements in accuracy, respectively, over group relative policy optimization (GRPO), which is a commonly used RLVR algorithm. Our code is available at https://github.com/purbeshmitra/semantic-soft-bootstrapping, and the model, curated dataset is available at https://huggingface.co/purbeshmitra/semantic-soft-bootstrapping.

Purbesh Mitra, Sennur Ulukus

We consider a gossip network, consisting of $n$ nodes, which tracks the information at a source. The source updates its information with a Poisson arrival process and also sends updates to the nodes in the network. The nodes themselves can exchange information among themselves to become as timely as possible. However, the network structure is sparse and irregular, i.e., not every node is connected to every other node in the network, rather, the order of connectivity is low, and varies across different nodes. This asymmetry of the network implies that the nodes in the network do not perform equally in terms of timelines. Due to the gossiping nature of the network, some nodes are able to track the source very timely, whereas, some nodes fall behind versions quite often. In this work, we investigate how the rate-constrained source should distribute its update rate across the network to maintain fairness regarding timeliness, i.e., the overall worst case performance of the network can be minimized. Due to the continuous search space for optimum rate allocation, we formulate this problem as a continuum-armed bandit problem and employ Gaussian process based Bayesian optimization to meet a trade-off between exploration and exploitation sequentially.

Purbesh Mitra, Priyanka Kaswan, Sennur Ulukus

Mixture-of-Agents (MoA) has recently been proposed as a method to enhance performance of large language models (LLMs), enabling multiple individual LLMs to work together for collaborative inference. This collaborative approach results in improved responses to user prompts compared to relying on a single LLM. In this paper, we consider such an MoA architecture in a distributed setting, where LLMs operate on individual edge devices, each uniquely associated with a user and equipped with its own distributed computing power. These devices exchange information using decentralized gossip algorithms, allowing different device nodes to talk without the supervision of a centralized server. In the considered setup, different users have their own LLM models to address user prompts. Additionally, the devices gossip either their own user-specific prompts or augmented prompts to generate more refined answers to certain queries. User prompts are temporarily stored in the device queues when their corresponding LLMs are busy. Given the memory limitations of edge devices, it is crucial to ensure that the average queue sizes in the system remain bounded. In this paper, we address this by theoretically calculating the queuing stability conditions for the device queues under reasonable assumptions, which we validate experimentally as well. Further, we demonstrate through experiments, leveraging open-source LLMs for the implementation of distributed MoA, that certain MoA configurations produce higher-quality responses compared to others, as evaluated on AlpacaEval 2.0 benchmark. The implementation is available at: https://github.com/purbeshmitra/distributed_moa.

Purbesh Mitra, Sennur Ulukus

Recent advancements in the reasoning capabilities of large language models (LLMs) show that employing group relative policy optimization (GRPO) algorithm for reinforcement learning (RL) training allows the models to use more thinking/reasoning tokens for generating better responses. However, LLMs can generate only a finite amount of tokens while maintaining attention to the previously generated tokens. This limit, also known as the context size of an LLM, is a bottleneck in LLM reasoning with arbitrarily large number of tokens. To think beyond the limit of context size, an LLM must employ a modular thinking strategy to reason over multiple rounds. In this work, we propose $\textbf{MOTIF: Modular Thinking via Reinforcement Finetuning}$ -- an RL training method for generating thinking tokens in multiple rounds, effectively allowing the model to think with additional context size. We trained the open-source model Qwen2.5-3B-Instruct on GSM8K dataset via parameter efficient fine-tuning and tested its accuracy on MATH500 and AIME2024 benchmarks. Our experiments show 3.8\% and 3.3\% improvements over vanilla GRPO based training in the respective benchmarks. Furthermore, this improvement was achieved with only 15\% of samples, thus demonstrating sample efficiency of MOTIF. Our code and models are available at https://github.com/purbeshmitra/MOTIF and https://huggingface.co/purbeshmitra/MOTIF, respectively.

Purbesh Mitra, Sennur Ulukus

We consider an asynchronous decentralized learning system, which consists of a network of connected devices trying to learn a machine learning model without any centralized parameter server. The users in the network have their own local training data, which is used for learning across all the nodes in the network. The learning method consists of two processes, evolving simultaneously without any necessary synchronization. The first process is the model update, where the users update their local model via a fixed number of stochastic gradient descent steps. The second process is model mixing, where the users communicate with each other via randomized gossiping to exchange their models and average them to reach consensus. In this work, we investigate the staleness criteria for such a system, which is a sufficient condition for convergence of individual user models. We show that for network scaling, i.e., when the number of user devices $n$ is very large, if the gossip capacity of individual users scales as $Ω(\log n)$, we can guarantee the convergence of user models in finite time. Furthermore, we show that the bounded staleness can only be guaranteed by any distributed opportunistic scheme by $Ω(n)$ scaling.

Purbesh Mitra, Sennur Ulukus

We consider an asynchronous hierarchical federated learning (AHFL) setting with a client-edge-cloud framework. The clients exchange the trained parameters with their corresponding edge servers, which update the locally aggregated model. This model is then transmitted to all the clients in the local cluster. The edge servers communicate to the central cloud server for global model aggregation. The goal of each client is to converge to the global model, while maintaining timeliness of the clients, i.e., having optimum training iteration time. We investigate the convergence criteria for such a system with dense clusters. Our analysis shows that for a system of $n$ clients with fixed average timeliness, the convergence in finite time is probabilistically guaranteed, if the nodes are divided into $O(1)$ number of clusters, that is, if the system is built as a sparse set of edge servers with dense client bases each.