arXiv subjects

Akhil Bhimaraju

Publications and source records attributed to Akhil Bhimaraju.

9 recordsLinked to original sources

arXiv2026

Skip-It? Theoretical Conditions for Layer Skipping in Vision-Language Models

Vision-language models achieve incredible performance across a wide range of tasks, but their large size makes inference costly. Recent work has shown that multimodal processing contains significant redundancies, making it possible to skip certain layers with minimal performance loss. Yet current pruning techniques remain ad-hoc, relying on heuristics or hyperparameter sweeps rather than principled criteria for determining when layer skipping is beneficial. In this paper, we propose a unified framework that characterizes the redundancy conditions under which pruning can enhance efficiency without sacrificing performance. Central to our approach are experimentally verifiable and interpretable notions of redundancy that can be evaluated without requiring downstream task performance as a metric. Applying this framework, we corroborate prior findings that both early and late vision tokens are redundant across models, and we validate our conditions by showing they align with actual performance degradation. Beyond these empirical results, our framework provides a theoretically grounded understanding of redundancy in VLMs and unifies many of the ideas behind modern layer-skipping techniques.

cs.AI↗

arXiv2026

Watermarking Discrete Diffusion Language Models

Watermarking has emerged as a promising technique to track AI-generated content and differentiate it from authentic human creations. While prior work extensively studies watermarking for autoregressive large language models (LLMs) and image diffusion models, it remains comparatively underexplored for discrete diffusion language models (DDLMs), which are becoming popular due to their high inference throughput. In this paper, we introduce one of the first watermarking methods for DDLMs. Our approach applies a distribution-preserving Gumbel-max sampling trick at every diffusion step and seeds the randomness by sequence position to enable reliable detection. We empirically demonstrate reliable detectability on LLaDA, a state-of-the-art DDLM. We also analytically prove that the watermark is distortion-free, with a false detection probability that decays exponentially in the sequence length. A key practical advantage is that our method realizes desired watermarking properties with no expensive hyperparameter tuning, making it straightforward to deploy and scale across models and benchmarks.

cs.CR↗

arXiv2026

DeepInsert: Early Layer Bypass for Efficient and Performant Multimodal Understanding

Hyperscaling of data and parameter count in LLMs is yielding diminishing improvement when weighed against training costs, underlining a growing need for more efficient finetuning and inference without sacrificing performance. This is especially so for multimodal language models (MLMs), where the overhead of processing multimodal tokens can limit their practical viability. Parallely, recent work has uncovered implicit cross-modal alignment in the deeper layers of large MLMs, deepening our understanding of how MLMs process and encode information. Motivated by this, and our observation that MLMs naturally defer most cross-modal token interactions to deeper layers of the model, we propose a simple modification. Instead of concatenation with the language prompt at the start, we insert multimodal tokens directly into the middle, allowing them to entirely bypass the early layers. Our results with diverse modalities, (i) LLaVA \& BLIP for vision, (ii) LTU for audio, and (iii) MoLCA for molecular data, and model sizes, starting from 350M to 13B parameters, indicate that our method reduces both training and inference costs, while at least preserving, if not surpassing the performance of existing baselines.

cs.CV↗

arXiv2025

Dynamic Batching of Online Arrivals to Leverage Economies of Scale

Many settings, such as matching riders to drivers in ride-hailing platforms or in-stream video advertising, require handling arrivals over time. In such applications, it is often beneficial to group the arriving orders or requests into batches and process the larger batches rather than individual arrivals. However, waiting too long to create larger batches incurs a waiting cost for past arrivals. On the other hand, processing the arrivals too soon leads to higher processing costs by missing the economies of scale of grouping larger numbers of arrivals into larger batches. Moreover, the timing of the next arrival is often unknown, meaning fixed-size batches or fixed waiting times tend to be poor choices. In this work, we consider the problem of finding the optimal batching schedule to minimize the sum of waiting time and processing cost under both offline and online settings. In the offline problem in which all arrival times are known a priori, we show that the optimal batching schedule can be found in polynomial time by reducing it to a shortest path problem on a weighted acyclic graph. For the online problem with unknown arrival times, we develop algorithms that are provably competitive for a broad range of processing-cost functions. We also provide a lower bound on the competitive ratio that no online algorithm can beat. Finally, we run numerical experiments on simulated and real data to demonstrate the effectiveness of our algorithms against the offline benchmark.

cs.DS↗

arXiv2024

Fractional Budget Allocation for Influence Maximization under General Marketing Strategies

We consider the fractional influence maximization problem, i.e., identifying users on a social network to be incentivized with potentially partial discounts to maximize the influence on the network. The larger the discount given to a user, the higher the likelihood of its activation (adopting a new product or innovation), who then attempts to activate its neighboring users, causing a cascade effect of influence through the network. Our goal is to devise efficient algorithms that assign initial discounts to the network's users to maximize the total number of activated users at the end of the cascade, subject to a constraint on the total sum of discounts given. In general, the activation likelihood could be any non-decreasing function of the discount, whereas, our focus lies on the case when the activation likelihood is an affine function of the discount, potentially varying across different users. As this problem is shown to be NP-hard, we propose and analyze an efficient (1-1/e)-approximation algorithm. Furthermore, we run experiments on real-world social networks to show the performance and scalability of our method.

cs.SI↗

arXiv2023

Dynamic Resource Allocation to Minimize Concave Costs of Shortfalls

We study a resource allocation problem over time, where a finite (random) resource needs to be distributed among a set of users at each time instant. Shortfalls in the resource allocated result in user dissatisfaction, which we model as an increasing function of the long-term average shortfall for each user. In many scenarios such as wireless multimedia streaming, renewable energy grid, or supply chain logistics, a natural choice for this cost function turns out to be concave, rather than usual convex cost functions. We consider minimizing the (normalized) cumulative cost across users. Depending on whether users' mean consumption rates are known or unknown, this problem can be reduced to two different structured non-convex problems. The "known" case is a concave minimization problem subject to a linear constraint. By exploiting a well-chosen linearization of the cost functions, we solve this provably within $\mathcal{O}\left(\frac{1}{m}\right)$ of the optimum, in $\mathcal{O}\left(m \log{m}\right)$ time, where $m$ is the number of users in the system. In the "unknown" case, we are faced with minimizing the sum of functions that are concave on part of the domain and convex on the rest, subject to a linear constraint. We present a provably exact algorithm when the cost functions and prior distributions on mean consumption are the same across all users.

math.OC↗

arXiv2022

Multi-channel Resource Allocation for Smooth Streaming: Non-convexity and Bandits

User dissatisfaction due to buffering pauses during streaming is a significant cost to the system, which we model as a non-decreasing function of the frequency of buffering pause. Minimization of total user dissatisfaction in a multi-channel cellular network leads to a non-convex problem. Utilizing a combinatorial structure in this problem, we first propose a polynomial time joint admission control and channel allocation algorithm which is provably (almost) optimal. This scheme assumes that the base station (BS) knows the frame statistics of the streams. In a more practical setting, where these statistics are not available a priori at the BS, a learning based scheme with provable guarantees is developed. This learning based scheme has relation to regret minimization in multi-armed bandits with non-i.i.d. and delayed reward (cost). All these algorithms require none to minimal feedback from the user equipment to the base station regarding the states of the media player buffer at the application layer, and hence, are of practical interest.

cs.NI↗

arXiv2021

Expected Extinction Times of Epidemics with State-Dependent Infectiousness

We model an epidemic where the per-person infectiousness in a network of geographic localities changes with the total number of active cases. This would happen as people adopt more stringent non-pharmaceutical precautions when the population has a larger number of active cases. We show that there exists a sharp threshold such that when the curing rate for the infection is above this threshold, the mean time for the epidemic to die out is logarithmic in the initial infection size, whereas when the curing rate is below this threshold, the mean time for epidemic extinction is infinite. We also show that when the per-person infectiousness goes to zero asymptotically as a function of the number of active cases, the mean extinction times all have the same asymptote independent of network structure. Simulations bear out these results, while also demonstrating that if the per-person infectiousness is large when the epidemic size is small (i.e., the precautions are lax when the epidemic is small and only get stringent after the epidemic has become large), it might take a very long time for the epidemic to die out. We also provide some analytical insight into these observations.

cs.SI↗

arXiv2020

Non-clairvoyant Scheduling of Coflows

The coflow scheduling problem is considered: given an input/output switch with each port having a fixed capacity, find a scheduling algorithm that minimizes the weighted sum of the coflow completion times respecting the port capacities, where each flow of a coflow has a demand per input/output port, and coflow completion time is the finishing time of the last flow of the coflow. The objective of this paper is to present theoretical guarantees on approximating the sum of coflow completion time in the non-clairvoyant setting, where on a coflow arrival, only the number of flows, and their input-output port is revealed, while the critical demand volumes for each flow on the respective input-output port is unknown. The main result of this paper is to show that the proposed BlindFlow algorithm is $8p$-approximate, where $p$ is the largest number of input-output port pairs that a coflow uses. This result holds even in the online case, where coflows arrive over time and the scheduler has to use only causal information. Simulations reveal that the experimental performance of BlindFlow is far better than the theoretical guarantee.

cs.DS↗