Akash Haridas

Vignesh Sundaresha, Akash Haridas, Vikram Appia et al.

Image generation has emerged as a mainstream application of large generative models. Just as test-time compute and reasoning have improved language model capabilities, similar benefits have been observed for image generation models. In particular, searching over noise samples for diffusion and flow models has been shown to scale well with test-time compute. While recent works explore allocating non-uniform inference-compute budgets across denoising steps, existing approaches rely on greedy heuristics and often allocate the compute budget ineffectively. In this work, we study this problem and propose a simple fix. We propose Verifier-Threshold, which automatically reallocates test-time compute and delivers substantial efficiency improvements. For the same performance on the GenEval benchmark, we achieve a 2-4x reduction in computational time over the state-of-the-art method.

Akash Haridas, Karim Hamadeh, Samarendra Chandan Bindu Dash

DAgger is an imitation algorithm that aggregates its original datasets by querying the expert on all samples encountered during training. In order to reduce the number of samples queried, we propose a modification to DAgger, known as DADAgger, which only queries the expert for state-action pairs that are out of distribution (OOD). OOD states are identified by measuring the variance of the action predictions of an ensemble of models on each state, which we simulate using dropout. Testing on the Car Racing and Half Cheetah environments achieves comparable performance to DAgger but with reduced expert queries, and better performance than a random sampling baseline. We also show that our algorithm may be used to build efficient, well-balanced training datasets by running with no initial data and only querying the expert to resolve uncertainty.

Parsa Ashrafi Fashi, Utkarsh Saxena, Mehdi Rezagholizadeh et al.

Hybrid sequence models that combine efficient Transformer components with linear sequence modeling blocks are a promising alternative to pure Transformers, but most are still pretrained from scratch and therefore fail to reuse existing Transformer checkpoints. We study upcycling as a practical path to convert pretrained Transformer LLMs into hybrid architectures while preserving short-context quality and improving long-context capability. We call our solution \emph{HyLo} (HYbrid LOng-context): a long-context upcycling recipe that combines architectural adaptation with efficient Transformer blocks, Multi-Head Latent Attention (MLA), and linear blocks (Mamba2 or Gated DeltaNet), together with staged long-context training and teacher-guided distillation for stable optimization. HyLo extends usable context length by up to $32\times$ through efficient post-training and reduces KV-cache memory by more than $90\%$, enabling up to 2M-token prefill and decoding in our \texttt{vLLM} inference stack, while comparable Llama baselines run out of memory beyond 64K context. Across 1B- and 3B-scale settings (Llama- and Qwen-based variants), HyLo delivers consistently strong short- and long-context performance and significantly outperforms state-of-the-art upcycled hybrid baselines on long-context evaluations such as RULER. Notably, at similar scale, HyLo-Qwen-1.7B trained on only 10B tokens significantly outperforms JetNemotron (trained on 400B tokens) on GSM8K, Lm-Harness common sense reasoning and RULER-64K.

Akash Haridas, Utkarsh Saxena, Parsa Ashrafi Fashi et al.

Diffusion Transformers process images as fixed-length sequences of tokens produced by a static $\textit{patchify}$ operation. While effective, this design spends uniform compute on low- and high-information regions alike, ignoring that images contain regions of varying detail and that the denoising process progresses from coarse structure at early timesteps to fine detail at late timesteps. We introduce the Dynamic Chunking Diffusion Transformer (DC-DiT), which augments the DiT backbone with a learned encoder-router-decoder scaffold that adaptively compresses the 2D input into a shorter token sequence in a data-dependent manner using a chunking mechanism learned end-to-end with diffusion training. The mechanism learns to compress uniform background regions into fewer tokens and detail-rich regions into more tokens, with meaningful visual segmentations emerging without explicit supervision. Furthermore, it also learns to adapt its compression across diffusion timesteps, using fewer tokens at noisy stages and more tokens as fine details emerge. On class-conditional ImageNet $256{\times}256$, DC-DiT consistently improves FID and Inception Score over both parameter-matched and FLOP-matched DiT baselines across $4{\times}$ and $16{\times}$ compression, showing this is a promising technique with potential further applications to pixel-space, video and 3D generation. Beyond accuracy, DC-DiT is practical: it can be upcycled from pretrained DiT checkpoints with minimal post-training compute (up to $8{\times}$ fewer training steps) and composes with other dynamic computation methods to further reduce generation FLOPs.

Akash Haridas, Imran N. Junejo

Applications of generating photo-realistic human avatars are many, however, high-fidelity avatar generation traditionally required expensive professional camera rigs and artistic labor, but recent research has enabled constructing them automatically from smartphones with RGB and IR sensors. However, these new methods still rely on the presence of high-resolution cameras on modern smartphones and often require offloading the processing to powerful servers with GPUs. Modern applications such as video conferencing call for the ability to generate these avatars from consumer-grade laptop webcams using limited compute available on-device. In this work, we develop a novel method based on 3D morphable models, landmark detection, photo-realistic texture GANs, and differentiable rendering to tackle the problem of low webcam image quality and edge computation. We build an automatic system to generate high-fidelity animatable avatars under these limitations, leveraging the neural compute capabilities of mobile chips.

Akash Haridas, Nagabhushana Rao Vadlamani, Yuki Minamoto

Information loss in numerical physics simulations can arise from various sources when solving discretized partial differential equations. In particular, errors related to numerical precision ("sub-precision errors") can accumulate in the quantities of interest when the simulations are performed using low-precision 16-bit floating-point arithmetic compared to an equivalent 64-bit simulation. On the other hand, low-precision computation is less resource intensive than high-precision computation. Several machine learning techniques proposed recently have been successful in correcting errors due to coarse spatial discretization. In this work, we extend these techniques to improve CFD simulations performed with low numerical precision. We quantify the precision-related errors accumulated in a Kolmogorov forced turbulence test case. Subsequently, we employ a Convolutional Neural Network together with a fully differentiable numerical solver performing 16-bit arithmetic to learn a tightly-coupled ML-CFD hybrid solver. Compared to the 16-bit solver, we demonstrate the efficacy of the hybrid solver towards improving various metrics pertaining to the statistical and pointwise accuracy of the simulation.