Ganesh Venkatesh

Yuanxin Wang, Pawel Filipczuk, Anisha Garg et al.

Modern LLM reasoning relies on extensive test-time computation, driven by internal model training and external agentic orchestration. However, this synergy is often inefficient, as model verbosity and poor instruction following lead to wasted compute. We analyze this capability-cost trade-off and introduce an optimized reasoning workflow (\cepo) that empowers smaller open-source models to outperform models multiple times their size. We will open-source this workflow to enable further research. Our work demonstrates a clear path toward co-designing orchestration frameworks with the underlying model capabilities to unlock powerful reasoning in small-to-medium sized models.

Hyojin Park, Jayeon Yoo, Seohyeong Jeong et al.

Current state-of-the-art approaches for Semi-supervised Video Object Segmentation (Semi-VOS) propagates information from previous frames to generate segmentation mask for the current frame. This results in high-quality segmentation across challenging scenarios such as changes in appearance and occlusion. But it also leads to unnecessary computations for stationary or slow-moving objects where the change across frames is minimal. In this work, we exploit this observation by using temporal information to quickly identify frames with minimal change and skip the heavyweight mask generation step. To realize this efficiency, we propose a novel dynamic network that estimates change across frames and decides which path -- computing a full network or reusing previous frame's feature -- to choose depending on the expected similarity. Experimental results show that our approach significantly improves inference speed without much accuracy degradation on challenging Semi-VOS datasets -- DAVIS 16, DAVIS 17, and YouTube-VOS. Furthermore, our approach can be applied to multiple Semi-VOS methods demonstrating its generality. The code is available in https://github.com/HYOJINPARK/Reuse_VOS.

Hyojin Park, Ganesh Venkatesh, Nojun Kwak

Semi-supervised video object segmentation (semi-VOS) is widely used in many applications. This task is tracking class-agnostic objects from a given target mask. For doing this, various approaches have been developed based on online-learning, memory networks, and optical flow. These methods show high accuracy but are hard to be utilized in real-world applications due to slow inference time and tremendous complexity. To resolve this problem, template matching methods are devised for fast processing speed but sacrificing lots of performance in previous models. We introduce a novel semi-VOS model based on a template matching method and a temporal consistency loss to reduce the performance gap from heavy models while expediting inference time a lot. Our template matching method consists of short-term and long-term matching. The short-term matching enhances target object localization, while long-term matching improves fine details and handles object shape-changing through the newly proposed adaptive template attention module. However, the long-term matching causes error-propagation due to the inflow of the past estimated results when updating the template. To mitigate this problem, we also propose a temporal consistency loss for better temporal coherence between neighboring frames by adopting the concept of a transition matrix. Our model obtains 79.5% J&F score at the speed of 73.8 FPS on the DAVIS16 benchmark. The code is available in https://github.com/HYOJINPARK/TTVOS.

Anisha Garg, Ganesh Venkatesh

Group-relative Policy Optimization's (GRPO) simplicity makes it highly desirable for adapting LLMs to become experts at specific tasks. But this simplicity also makes it ill-specified as we seek to enhance RL training with richer, non-binary feedback. When using ordinal rewards to give partial credit, GRPO's simplicity starts to hurt, as its group-average baseline often assigns a positive advantage to failed trajectories and reinforces incorrect behavior. We introduce Correctness Relative Policy Optimization (CoRPO), a new formulation that solves this flaw. CoRPO uses an adaptive baseline that enforces a minimum quality threshold, ensuring failed solutions are never positively reinforced. Once the policy consistently meets this threshold, the baseline automatically transitions to a relative preference mode, pushing the model to find optimal solutions rather than just "acceptable" ones. We empirically validate CoRPO on a code verification task, where it demonstrates more stable convergence and better out-of-domain generalization. This work represents a critical step in our broader research program to enable LLMs to learn genuinely new capabilities through reinforcement learning. We achieve this by enabling LLMs to learn from rich, multi-dimensional feedback - progressing from binary to ordinal rewards in this work, and onward to denser, per-step supervision.

Nishit Neema, Srinjoy Mukherjee, Sapan Shah et al.

Large Language Models (LLMs) excel at general tasks but underperform in specialized domains like economics and psychology, which require deep, principled understanding. To address this, we introduce ACER (Automated Curriculum-Enhanced Regimen) that transforms generalist models into domain experts without sacrificing their broad capabilities. ACER first synthesizes a comprehensive, textbook-style curriculum by generating a table of contents for a subject and then creating question-answer (QA) pairs guided by Bloom's taxonomy. This ensures systematic topic coverage and progressively increasing difficulty. The resulting synthetic corpus is used for continual pretraining with an interleaved curriculum schedule, aligning learning across both content and cognitive dimensions. Experiments with Llama 3.2 (1B and 3B) show significant gains in specialized MMLU subsets. In challenging domains like microeconomics, where baselines struggle, ACER boosts accuracy by 5 percentage points. Across all target domains, we observe a consistent macro-average improvement of 3 percentage points. Notably, ACER not only prevents catastrophic forgetting but also facilitates positive cross-domain knowledge transfer, improving performance on non-target domains by 0.7 points. Beyond MMLU, ACER enhances performance on knowledge-intensive benchmarks like ARC and GPQA by over 2 absolute points, while maintaining stable performance on general reasoning tasks. Our results demonstrate that ACER offers a scalable and effective recipe for closing critical domain gaps in LLMs.

Vithursan Thangarasa, Ganesh Venkatesh, Mike Lasby et al.

Large language models have driven significant progress in natural language processing, but their deployment requires substantial compute and memory resources. As models scale, compression techniques become essential for balancing model quality with computational efficiency. Structured pruning, which removes less critical components of the model, is a promising strategy for reducing complexity. However, one-shot pruning often results in significant quality degradation, particularly in tasks requiring multi-step reasoning. To recover lost quality, supervised fine-tuning (SFT) is commonly applied, but it can lead to catastrophic forgetting by shifting the model's learned data distribution. Therefore, addressing the degradation from both pruning and SFT is essential to preserve the original model's quality. In this work, we utilize self-data distilled fine-tuning to address these challenges. Our approach leverages the original, unpruned model to generate a distilled dataset that preserves semantic richness and mitigates catastrophic forgetting by maintaining alignment with the base model's knowledge. Empirically, we demonstrate that self-data distillation consistently outperforms standard SFT, improving average accuracy by up to 8% on the HuggingFace OpenLLM Leaderboard v1. Specifically, when pruning six decoder blocks on Llama3.1-8B Instruct (i.e., 32 to 26 layers, reducing the model size from 8.03B to 6.72B parameters), our method retains 91.2% of the original model's accuracy compared to 81.7% with SFT, while reducing real-world FLOPs by 16.3%. Furthermore, combining self-data distilled models through model merging yields enhanced quality retention. Additionally, leveraging these pruned models in speculative decoding increases token acceptance rates, thereby improving inference efficiency in applied settings.

Yuanxin Wang, Ganesh Venkatesh

Large Language Models (LLMs) have demonstrated remarkable proficiency in understanding text and generating high-quality responses. However, a critical distinction from human cognition is their typical lack of a distinct internal `reading' or deliberation phase before `speaking' (i.e., generating text). Humans often engage in silent reading to comprehend context and formulate thoughts prior to articulation. This paper investigates methods to imbue LLMs with a similar capacity for internal processing. We introduce and evaluate techniques that encourage LLMs to `read silently.' Our findings indicate that even a straightforward approach, such as providing the model with an initial contextual prompt or `reading space' before it begins predicting subsequent tokens for the final output, can yield significant performance improvements. We further enhance this concept by developing a `reading buddy' architecture, where an auxiliary component silently processes the input and provides refined contextual insights to the primary generation model. These approaches aim to foster deeper understanding from LLMs so that they can produce better reasoned responses, moving them one step closer to more human-like text processing. Our results indicate that these simple techniques can provide surprisingly strong impact on accuracy with multiple point accuracy boost.

Aarti Ghatkesar, Ganesh Venkatesh

Achieving deep alignment between vision and language remains a central challenge for Multimodal Large Language Models (MLLMs). These models often fail to fully leverage visual input, defaulting to strong language priors. Our approach first provides insights into how MLLMs internally build visual understanding of image regions and then introduces techniques to amplify this capability. Specifically, we explore techniques designed both to deepen the model's understanding of visual content and to ensure that these visual insights actively guide language generation. We demonstrate the superior multimodal understanding of our resultant model through a detailed upstream analysis quantifying its ability to predict visually-dependent tokens as well as 10 pt boost on visually challenging tasks.

Anisha Garg, Engin Tekin, Yash More et al.

Advanced test-time computing strategies are essential for scaling reasoning models, but their effectiveness is capped by the models' poor self-evaluation. We propose a pairwise Explanatory Verifier, trained via reinforcement learning (GRPO), that produces calibrated confidence scores and associated natural language reasoning for generated solutions. Our verifier improves the accuracy and efficiency of test-time strategies like best-of-n and self-reflection. Crucially, it excels at identifying challenging failure modes, such as when both candidate solutions are identically incorrect, succeeding where standard methods like majority voting fail.

Zhaoshuo Li, Wei Ye, Dilin Wang et al.

Temporally consistent depth estimation is crucial for online applications such as augmented reality. While stereo depth estimation has received substantial attention as a promising way to generate 3D information, there is relatively little work focused on maintaining temporal stability. Indeed, based on our analysis, current techniques still suffer from poor temporal consistency. Stabilizing depth temporally in dynamic scenes is challenging due to concurrent object and camera motion. In an online setting, this process is further aggravated because only past frames are available. We present a framework named Consistent Online Dynamic Depth (CODD) to produce temporally consistent depth estimates in dynamic scenes in an online setting. CODD augments per-frame stereo networks with novel motion and fusion networks. The motion network accounts for dynamics by predicting a per-pixel SE3 transformation and aligning the observations. The fusion network improves temporal depth consistency by aggregating the current and past estimates. We conduct extensive experiments and demonstrate quantitatively and qualitatively that CODD outperforms competing methods in terms of temporal consistency and performs on par in terms of per-frame accuracy.

Haichuan Yang, Yuan Shangguan, Dilin Wang et al.

From wearables to powerful smart devices, modern automatic speech recognition (ASR) models run on a variety of edge devices with different computational budgets. To navigate the Pareto front of model accuracy vs model size, researchers are trapped in a dilemma of optimizing model accuracy by training and fine-tuning models for each individual edge device while keeping the training GPU-hours tractable. In this paper, we propose Omni-sparsity DNN, where a single neural network can be pruned to generate optimized model for a large range of model sizes. We develop training strategies for Omni-sparsity DNN that allows it to find models along the Pareto front of word-error-rate (WER) vs model size while keeping the training GPU-hours to no more than that of training one singular model. We demonstrate the Omni-sparsity DNN with streaming E2E ASR models. Our results show great saving on training time and resources with similar or better accuracy on LibriSpeech compared to individually pruned sparse models: 2%-6.6% better WER on Test-other.

Dilin Wang, Yuan Shangguan, Haichuan Yang et al.

Automatic speech recognition (ASR) has become increasingly ubiquitous on modern edge devices. Past work developed streaming End-to-End (E2E) all-neural speech recognizers that can run compactly on edge devices. However, E2E ASR models are prone to overfitting and have difficulties in generalizing to unseen testing data. Various techniques have been proposed to regularize the training of ASR models, including layer normalization, dropout, spectrum data augmentation and speed distortions in the inputs. In this work, we present a simple yet effective noisy training strategy to further improve the E2E ASR model training. By introducing random noise to the parameter space during training, our method can produce smoother models at convergence that generalize better. We apply noisy training to improve both dense and sparse state-of-the-art Emformer models and observe consistent WER reduction. Specifically, when training Emformers with 90% sparsity, we achieve 12% and 14% WER improvements on the LibriSpeech Test-other and Test-clean data set, respectively.

David Eriksson, Pierce I-Jen Chuang, Samuel Daulton et al.

When tuning the architecture and hyperparameters of large machine learning models for on-device deployment, it is desirable to understand the optimal trade-offs between on-device latency and model accuracy. In this work, we leverage recent methodological advances in Bayesian optimization over high-dimensional search spaces and multi-objective Bayesian optimization to efficiently explore these trade-offs for a production-scale on-device natural language understanding model at Facebook.

Varun Nagaraja, Yangyang Shi, Ganesh Venkatesh et al.

On-device speech recognition requires training models of different sizes for deploying on devices with various computational budgets. When building such different models, we can benefit from training them jointly to take advantage of the knowledge shared between them. Joint training is also efficient since it reduces the redundancy in the training procedure's data handling operations. We propose a method for collaboratively training acoustic encoders of different sizes for speech recognition. We use a sequence transducer setup where different acoustic encoders share a common predictor and joiner modules. The acoustic encoders are also trained using co-distillation through an auxiliary task for frame level chenone prediction, along with the transducer loss. We perform experiments using the LibriSpeech corpus and demonstrate that the collaboratively trained acoustic encoders can provide up to a 11% relative improvement in the word error rate on both the test partitions.

Asit Mishra, Jorge Albericio Latorre, Jeff Pool et al.

As neural network model sizes have dramatically increased, so has the interest in various techniques to reduce their parameter counts and accelerate their execution. An active area of research in this field is sparsity - encouraging zero values in parameters that can then be discarded from storage or computations. While most research focuses on high levels of sparsity, there are challenges in universally maintaining model accuracy as well as achieving significant speedups over modern matrix-math hardware. To make sparsity adoption practical, the NVIDIA Ampere GPU architecture introduces sparsity support in its matrix-math units, Tensor Cores. We present the design and behavior of Sparse Tensor Cores, which exploit a 2:4 (50%) sparsity pattern that leads to twice the math throughput of dense matrix units. We also describe a simple workflow for training networks that both satisfy 2:4 sparsity pattern requirements and maintain accuracy, verifying it on a wide range of common tasks and model architectures. This workflow makes it easy to prepare accurate models for efficient deployment on Sparse Tensor Cores.

Ganesh Venkatesh, Alagappan Valliappan, Jay Mahadeokar et al.

Recurrent transducer models have emerged as a promising solution for speech recognition on the current and next generation smart devices. The transducer models provide competitive accuracy within a reasonable memory footprint alleviating the memory capacity constraints in these devices. However, these models access parameters from off-chip memory for every input time step which adversely effects device battery life and limits their usability on low-power devices. We address transducer model's memory access concerns by optimizing their model architecture and designing novel recurrent cell designs. We demonstrate that i) model's energy cost is dominated by accessing model weights from off-chip memory, ii) transducer model architecture is pivotal in determining the number of accesses to off-chip memory and just model size is not a good proxy, iii) our transducer model optimizations and novel recurrent cell reduces off-chip memory accesses by 4.5x and model size by 2x with minimal accuracy impact.

Yutong Bai, Haoqi Fan, Ishan Misra et al.

Leveraging temporal information has been regarded as essential for developing video understanding models. However, how to properly incorporate temporal information into the recent successful instance discrimination based contrastive self-supervised learning (CSL) framework remains unclear. As an intuitive solution, we find that directly applying temporal augmentations does not help, or even impair video CSL in general. This counter-intuitive observation motivates us to re-design existing video CSL frameworks, for better integration of temporal knowledge. To this end, we present Temporal-aware Contrastive self-supervised learningTaCo, as a general paradigm to enhance video CSL. Specifically, TaCo selects a set of temporal transformations not only as strong data augmentation but also to constitute extra self-supervision for video understanding. By jointly contrasting instances with enriched temporal transformations and learning these transformations as self-supervised signals, TaCo can significantly enhance unsupervised video representation learning. For instance, TaCo demonstrates consistent improvement in downstream classification tasks over a list of backbones and CSL approaches. Our best model achieves 85.1% (UCF-101) and 51.6% (HMDB-51) top-1 accuracy, which is a 3% and 2.4% relative improvement over the previous state-of-the-art.

Paulius Micikevicius, Sharan Narang, Jonah Alben et al.

Deep neural networks have enabled progress in a wide variety of applications. Growing the size of the neural network typically results in improved accuracy. As model sizes grow, the memory and compute requirements for training these models also increases. We introduce a technique to train deep neural networks using half precision floating point numbers. In our technique, weights, activations and gradients are stored in IEEE half-precision format. Half-precision floating numbers have limited numerical range compared to single-precision numbers. We propose two techniques to handle this loss of information. Firstly, we recommend maintaining a single-precision copy of the weights that accumulates the gradients after each optimizer step. This single-precision copy is rounded to half-precision format during training. Secondly, we propose scaling the loss appropriately to handle the loss of information with half-precision gradients. We demonstrate that this approach works for a wide variety of models including convolution neural networks, recurrent neural networks and generative adversarial networks. This technique works for large scale models with more than 100 million parameters trained on large datasets. Using this approach, we can reduce the memory consumption of deep learning models by nearly 2x. In future processors, we can also expect a significant computation speedup using half-precision hardware units.

Ganesh Venkatesh, Eriko Nurvitadhi, Debbie Marr

We explore techniques to significantly improve the compute efficiency and performance of Deep Convolution Networks without impacting their accuracy. To improve the compute efficiency, we focus on achieving high accuracy with extremely low-precision (2-bit) weight networks, and to accelerate the execution time, we aggressively skip operations on zero-values. We achieve the highest reported accuracy of 76.6% Top-1/93% Top-5 on the Imagenet object classification challenge with low-precision network\footnote{github release of the source code coming soon} while reducing the compute requirement by ~3x compared to a full-precision network that achieves similar accuracy. Furthermore, to fully exploit the benefits of our low-precision networks, we build a deep learning accelerator core, dLAC, that can achieve up to 1 TFLOP/mm^2 equivalent for single-precision floating-point operations (~2 TFLOP/mm^2 for half-precision).