arXiv subjects

Mayank Mishra

Publications and source records attributed to Mayank Mishra.

At least 19 recordsLinked to original sources

arXiv2026

Monolithic Integration of Piezo-Optomechanical Photonics and CMOS Electronics

Next-generation photonic architectures for AI, sensing, and quantum computing require thousands to millions of reprogrammable photonic devices on a chip[1]. The monolithic integration of Electronically-backed Photonic Integrated Circuits (EPICs) allows for very high density electrical interconnection and electronic drivers that can scale with photonics. Piezo-optomechanical photonic integrated circuits (POMPICs) offer low power consumption, high speed modulation, cryogenic compatibility and broadband optical transparency from ultraviolet to infrared wavelengths[2,3], but have not been demonstrated with monolithically integrated CMOS electronics. Here, we show a fully monolithic, all-CMOS fabricated platform for POMPICs co-fabricated with commercial control electronics. 200 millimeter photonic wafers are constructed directly on completed CMOS driver wafers by back-end-of-line processing, connecting integrated piezoelectric actuators under broadband silicon nitride waveguides to a high-density digital backplane comprising >2 million electrical connections per die with 6.4x6.4 micron electrode pitch. We introduce segmented POMPIC components as Photonic Digital-to-Analog converters (PDACs) that convert low-voltage digital electronic signals to multi-bit analog optical phase and amplitude modulation, and we demonstrate parallel control of optical phase shifters, Mach-Zehnder interferometers, optical routing trees, and tunable ring resonators using a standard HDMI interface to program CMOS electronics. We test multiple reticles and perform electronic and photonic characterization across the entire wafer to establish uniformity and yield, demonstrating wafer-scale integration of POMPICs on an electronic backplane and enabling dense, scalable electronic control of piezo-optomechanical circuits.

physics.optics↗

arXiv2026

Recover, Discover, Plan: Learning Skills and Concepts from Robot Failures

Intelligent robots should not only recover from failures, but also acquire the abstract knowledge needed to avoid them in the future. While reinforcement learning (RL) can learn reactive recovery behaviors, training a separate policy for every distinct failure mode is highly inefficient. We introduce Recovery-Driven Synthesis of Relational Concepts (ReSYNC), the first approach that progressively discovers and refines state abstractions (relational predicates) from failure-recovery experience to support abstract planning. Unlike purely reactive methods, ReSYNC jointly learns skills and concepts through an incremental dual-learning process. In the skill-learning phase, the robot uses RL to learn to recover from failures seen in training tasks. In the concept-learning phase, the robot discovers new relational predicates and refines its abstract planning model to explain and generalize the learned recovery behaviors. This interaction enables ReSYNC to convert local recoveries seen during training into global failure avoidance at test time. Across four simulated domains, we show that ReSYNC's ability to continually expand and refine its abstraction library allows it to solve long-horizon, previously unseen problems, outperforming strong baselines by over 50%. Additionally, we demonstrate sim-to-real transfer of ReSYNC, where it performs real-world non-prehensile manipulation skills and generalizes to unseen scenarios through abstract planning. Overall, ReSYNC represents a significant step toward robots that autonomously acquire abstractions for scalable, failure-aware planning in the physical world.

cs.RO↗

arXiv2026

Taylor-Calibrate: Principled Initialization for Hybrid Linear Attention Distillation

Hybrid linear attention models offer an appealing path to faster long-context inference: they reduce the quadratic cost and KV-cache burden of full softmax attention while retaining much of the quality of Transformer models. A practical way to obtain such models is to convert a pretrained Transformer instead of pretraining a new architecture from scratch, but this conversion is still brittle. Simply copying the teacher attention projections into a Gated DeltaNet (GDN) student does not specify the new recurrent decay, write, and output-gating dynamics. As a result, the converted model often starts in a poor dynamical regime and must spend many distillation tokens repairing initialization rather than learning the remaining teacher behavior. We propose Taylor-Calibrate, a lightweight initialization method for hybrid GDN students. The method uses Taylor-guided teacher attention statistics to set the value projection, memory timescale, write gates, and output gate, then applies a short per-layer alignment step to match each converted layer to the teacher output. Across four teacher settings and three retained-layer policies, Taylor-Calibrate gives substantially stronger zero-shot students, with up to an 88x improvement in a representative ablation, and reaches matched recovery targets with 4.9x--9.2x fewer training tokens than naive conversion.

cs.LG↗

arXiv2026

Description and Discussion on DCASE 2026 Challenge Task 4: Spatial Semantic Segmentation of Sound Scenes

This paper presents an overview of the Detection and Classification of Acoustic Scenes and Events (DCASE) 2026 Challenge Task 4, Spatial Semantic Segmentation of Sound Scenes (S5). The S5 task focuses on the joint detection and separation of sound events in complex spatial audio mixtures, contributing to the foundation of immersive communication. First introduced in DCASE 2025, the S5 task continues in DCASE 2026 Task 4 with key changes to better reflect real-world conditions, including allowing mixtures to contain multiple sources of the same class and to contain no target sources. In this paper, we describe task setting, along with the corresponding updates to the evaluation metrics and dataset. The experimental results of the submitted systems are also reported and analyzed. The official access point for data and code is https://github.com/nttcslab/dcase2026_task4_baseline.

eess.AS↗

arXiv2026

M$^2$RNN: Non-Linear RNNs with Matrix-Valued States for Scalable Language Modeling

Transformers are highly parallel but are limited to computations in the TC$^0$ complexity class, excluding tasks such as entity tracking and code execution that provably require greater expressive power. Motivated by this limitation, we revisit non-linear Recurrent Neural Networks (RNNs) for language modeling and introduce Matrix-to-Matrix RNN (M$^2$RNN): an architecture with matrix-valued hidden states and expressive non-linear state transitions. We demonstrate that the language modeling performance of non-linear RNNs is limited by their state size, and show how the state size expansion mechanism enables efficient use of tensor cores. Empirically, M$^2$RNN achieves perfect state tracking generalization at sequence lengths not seen during training. These benefits also translate to large-scale language modeling. In hybrid settings that interleave recurrent layers with attention, Hybrid M$^2$RNN outperforms equivalent Gated DeltaNet hybrids by $0.4$-$0.5$ perplexity points on a 7B MoE model, while using $3\times$ smaller state sizes for the recurrent layers. Notably, replacing even a single recurrent layer with M$^2$RNN in an existing hybrid architecture yields accuracy gains comparable to Hybrid M$^2$RNN with minimal impact on training throughput. Further, the Hybrid Gated DeltaNet models with a single M$^2$RNN layer also achieve superior long-context generalization, outperforming state-of-the-art hybrid linear attention architectures by up to $8$ points on LongBench. Together, these results establish non-linear RNN layers as a compelling building block for efficient and scalable language models.

cs.LG↗

arXiv2026

Ultra-low loss piezo-optomechanical low-confinement silicon nitride platform for visible wavelength quantum photonic circuits

Realizing photonic quantum computing at scale requires integrated circuits that combine ultra-low loss with fast, low-power, low-hysteresis, and low-crosstalk reconfiguration. These requirements are particularly challenging at visible wavelengths, where many quantum resource-state generators, including single-photon sources and quantum memories, naturally operate. Low-confinement silicon nitride waveguides offer the requisite loss performance, but conventional thermo-optic modulators dissipate significant static power, driving thermal crosstalk and precluding cryogenic operation. Visible-wavelength piezo-optomechanical circuits avoid these drawbacks, but existing demonstrations rely either on high-confinement waveguides with propagation losses of 35-100 dB/m, or on low-confinement platforms using foundry-incompatible PZT. Here we combine piezo-optomechanical actuation with a CMOS-foundry-fabricated, low-confinement silicon nitride platform, achieving 2.6 dB/m propagation loss at 780 nm, megahertz-scale modulation bandwidth, a half-wave voltage-length product of 2.8 Vm, and negligible hysteresis. We demonstrate reconfigurable Mach-Zehnder interferometers with 0.63 dB loss per spiral phase shifter, enabling deep, actively reconfigurable visible-wavelength quantum photonic circuits.

physics.optics↗

arXiv2025

Distilling to Hybrid Attention Models via KL-Guided Layer Selection

Distilling pretrained softmax attention Transformers into more efficient hybrid architectures that interleave softmax and linear attention layers is a promising approach for improving the inference efficiency of LLMs without requiring expensive pretraining from scratch. A critical factor in the conversion process is layer selection, i.e., deciding on which layers to convert to linear attention variants. This paper describes a simple and efficient recipe for layer selection that uses layer importance scores derived from a small amount of training on generic text data. Once the layers have been selected we use a recent pipeline for the distillation process itself \citep[RADLADS;][]{goldstein2025radlads}, which consists of attention weight transfer, hidden state alignment, KL-based distribution matching, followed by a small amount of finetuning. We find that this approach is more effective than existing approaches for layer selection, including heuristics that uniformly interleave linear attentions based on a fixed ratio, as well as more involved approaches that rely on specialized diagnostic datasets.

cs.CL↗

arXiv2025

SonicMoE: Accelerating MoE with IO and Tile-aware Optimizations

Mixture of Experts (MoE) models have emerged as the de facto architecture for scaling up language models without significantly increasing the computational cost. Recent MoE models demonstrate a clear trend towards high expert granularity (smaller expert intermediate dimension) and higher sparsity (constant number of activated experts with a higher number of total experts), which improve model quality per FLOP. However, fine-grained MoEs suffer from increased activation memory footprint and reduced hardware efficiency due to higher IO costs, while sparser MoEs suffer from wasted computations due to padding in Grouped GEMM kernels. In response, we propose a memory-efficient algorithm to compute the forward and backward passes of MoEs with minimal activation caching for the backward pass. We also design GPU kernels that overlap memory IO with computation, benefiting all MoE architectures. Finally, we propose a novel "token rounding" method that minimizes the wasted compute due to padding in Grouped GEMM kernels. As a result, our method SonicMoE reduces activation memory by 45% and achieves a 1.86x compute throughput improvement on Hopper GPUs compared to ScatterMoE's BF16 MoE kernel for a fine-grained 7B MoE. Concretely, SonicMoE on 64 H100s achieves a training throughput of 213 billion tokens per day, comparable to ScatterMoE's 225 billion tokens per day on 96 H100s for a 7B MoE model training with FSDP-2 using the lm-engine codebase. On Blackwell GPUs, SonicMoE also achieves a 25% and 15% relative speedup on the forward and backward pass respectively compared to a highly optimized DeepGEMM baseline on OLMoE-sized 7B MoE models. Under high MoE sparsity settings, our tile-aware token rounding algorithm yields an additional 1.16x speedup on kernel execution time compared to vanilla top-K routing while maintaining similar downstream performance on Hopper GPUs. We open-source all our kernels.

cs.LG↗

arXiv2025

Metric Analysis for Spatial Semantic Segmentation of Sound Scenes

Spatial semantic segmentation of sound scenes (S5) consists of jointly performing audio source separation and sound event classification from a multichannel audio mixture. Evaluating S5 systems with separation and classification metrics individually makes system comparison difficult, whereas existing joint metrics, such as the class-aware signal-to-distortion ratio (CA-SDR), can conflate separation and labeling errors. In particular, CA-SDR relies on predicted class labels for source matching, which may obscure label swaps or misclassifications when the underlying source estimates remain perceptually correct. In this work, we introduce the class and source-aware signal-to-distortion ratio (CASA-SDR), a new metric that performs permutation-invariant source matching before computing classification errors, thereby shifting from a classification-focused approach to a separation-focused approach. We first analyze CA-SDR in controlled scenarios with oracle separation and synthetic classification errors, as well as under controlled cross-contamination between sources, and compare its behavior to that of the classical SDR and CASA-SDR. We also study the impact of classification errors on the metrics by introducing error-based and source-based aggregation strategies. Finally, we compare CA-SDR and CASA-SDR on systems submitted to Task 4 of the DCASE 2025 challenge, highlighting the cases where CA-SDR over-penalizes label swaps or poorly separated sources, while CASA-SDR provides a more interpretable separation-centric assessment of S5 performance.

cs.SD↗

arXiv2025

Description and Discussion on DCASE 2025 Challenge Task 4: Spatial Semantic Segmentation of Sound Scenes

Spatial Semantic Segmentation of Sound Scenes (S5) aims to enhance technologies for sound event detection and separation from multi-channel input signals that mix multiple sound events with spatial information. This is a fundamental basis of immersive communication. The ultimate goal is to separate sound event signals with 6 Degrees of Freedom (6DoF) information into dry sound object signals and metadata about the object type (sound event class) and representing spatial information, including direction. However, because several existing challenge tasks already provide some of the subset functions, this task for this year focuses on detecting and separating sound events from multi-channel spatial input signals. This paper outlines the S5 task setting of the Detection and Classification of Acoustic Scenes and Events (DCASE) 2025 Challenge Task 4 and the DCASE2025 Task 4 Dataset, newly recorded and curated for this task. We also report experimental results for an S5 system trained and evaluated on this dataset. The full version of this paper will be published after the challenge results are made public.

cs.SD↗

arXiv2025

FlashFormer: Whole-Model Kernels for Efficient Low-Batch Inference

The size and compute characteristics of modern large language models have led to an increased interest in developing specialized kernels tailored for particular training and inference workloads. Existing kernels primarily optimize for compute utilization, targeting the large-batch training and inference settings. However, low-batch inference, where memory bandwidth and kernel launch overheads are significant factors, remains important for many applications of interest such as in edge deployment and latency-sensitive applications. This paper describes FlashFormer, which fuses the entire transformer forward pass into a single kernel for accelerating low-batch inference of large language models. Across various model sizes and quantizations settings, FlashFormer achieves nontrivial speedups compared to existing inference kernels.

cs.LG↗

arXiv2025

PaTH Attention: Position Encoding via Accumulating Householder Transformations

The attention mechanism is a core primitive in modern large language models (LLMs) and AI more broadly. Since attention by itself is permutation-invariant, position encoding is essential for modeling structured domains such as language. Rotary position encoding (RoPE) has emerged as the de facto standard approach for position encoding and is part of many modern LLMs. However, in RoPE the key/query transformation between two elements in a sequence is only a function of their relative position and otherwise independent of the actual input. This limits the expressivity of RoPE-based transformers. This paper describes PaTH, a flexible data-dependent position encoding scheme based on accumulated products of Householder(like) transformations, where each transformation is data-dependent, i.e., a function of the input. We derive an efficient parallel algorithm for training through exploiting a compact representation of products of Householder matrices, and implement a FlashAttention-style blockwise algorithm. Across both targeted synthetic benchmarks and moderate-scale real-world language modeling experiments, we find that PaTH improves upon RoPE and other recent baselines. Finally, we show that we can convert pretrained RoPE transformers into PaTH with continued pretraining.

cs.CL↗

arXiv2025

Granite Vision: a lightweight, open-source multimodal model for enterprise Intelligence

We introduce Granite Vision, a lightweight large language model with vision capabilities, specifically designed to excel in enterprise use cases, particularly in visual document understanding. Our model is trained on a comprehensive instruction-following dataset, including document-related tasks, such as content extraction from tables, charts, diagrams, sketches, and infographics, as well as general image tasks. The architecture of Granite Vision is centered around visual modality alignment with a decoder-only, 2 billion parameter Granite large language model. Additionally, we introduce a dedicated safety classification approach in test-time that leverages a sparse set of attention vectors to identify potential harmful inputs. Despite its lightweight architecture, Granite Vision achieves strong results in standard benchmarks related to visual document understanding, as well as on the LiveXiv benchmark, which is designed to avoid test set contamination by using a constantly updated corpus of recently published Arxiv papers. We are releasing the model under the Apache-2 license, allowing for both research and commercial use, while offering complete visibility into the training data and other relevant details. See https://huggingface.co/ibm-granite/ for model weights.

cs.CV↗

arXiv2025

Ladder-residual: parallelism-aware architecture for accelerating large model inference with communication overlapping

Large language model inference is both memory-intensive and time-consuming, often requiring distributed algorithms to efficiently scale. Various model parallelism strategies are used in multi-gpu training and inference to partition computation across multiple devices, reducing memory load and computation time. However, using model parallelism necessitates communication of information between GPUs, which has been a major bottleneck and limits the gains obtained by scaling up the number of devices. We introduce Ladder Residual, a simple architectural modification applicable to all residual-based models that enables straightforward overlapping that effectively hides the latency of communication. Our insight is that in addition to systems optimization, one can also redesign the model architecture to decouple communication from computation. While Ladder Residual can allow communication-computation decoupling in conventional parallelism patterns, we focus on Tensor Parallelism in this paper, which is particularly bottlenecked by its heavy communication. For a Transformer model with 70B parameters, applying Ladder Residual to all its layers can achieve 29% end-to-end wall clock speed up at inference time with TP sharding over 8 devices. We refer the resulting Transformer model as the Ladder Transformer. We train a 1B and 3B Ladder Transformer from scratch and observe comparable performance to a standard dense transformer baseline. We also show that it is possible to convert parts of the Llama-3.1 8B model to our Ladder Residual architecture with minimal accuracy degradation by only retraining for 3B tokens. We release our code for training and inference for easier replication of experiments.

cs.LG↗

arXiv2024

Selective Self-Rehearsal: A Fine-Tuning Approach to Improve Generalization in Large Language Models

Fine-tuning Large Language Models (LLMs) on specific datasets is a common practice to improve performance on target tasks. However, this performance gain often leads to overfitting, where the model becomes too specialized in either the task or the characteristics of the training data, resulting in a loss of generalization. This paper introduces Selective Self-Rehearsal (SSR), a fine-tuning approach that achieves performance comparable to the standard supervised fine-tuning (SFT) while improving generalization. SSR leverages the fact that there can be multiple valid responses to a query. By utilizing the model's correct responses, SSR reduces model specialization during the fine-tuning stage. SSR first identifies the correct model responses from the training set by deploying an appropriate LLM as a judge. Then, it fine-tunes the model using the correct model responses and the gold response for the remaining samples. The effectiveness of SSR is demonstrated through experiments on the task of identifying unanswerable queries across various datasets. The results show that standard SFT can lead to an average performance drop of up to $16.7\%$ on multiple benchmarks, such as MMLU and TruthfulQA. In contrast, SSR results in close to $2\%$ drop on average, indicating better generalization capabilities compared to standard SFT.

cs.CL↗

arXiv2024

Power Scheduler: A Batch Size and Token Number Agnostic Learning Rate Scheduler

Finding the optimal learning rate for language model pretraining is a challenging task. This is not only because there is a complicated correlation between learning rate, batch size, number of training tokens, model size, and other hyperparameters but also because it is prohibitively expensive to perform a hyperparameter search for large language models with Billions or Trillions of parameters. Recent studies propose using small proxy models and small corpus to perform hyperparameter searches and transposing the optimal parameters to large models and large corpus. While the zero-shot transferability is theoretically and empirically proven for model size related hyperparameters, like depth and width, the zero-shot transfer from small corpus to large corpus is underexplored. In this paper, we study the correlation between optimal learning rate, batch size, and number of training tokens for the recently proposed WSD scheduler. After thousands of small experiments, we found a power-law relationship between variables and demonstrated its transferability across model sizes. Based on the observation, we propose a new learning rate scheduler, Power scheduler, that is agnostic about the number of training tokens and batch size. The experiment shows that combining the Power scheduler with Maximum Update Parameterization (muP) can consistently achieve impressive performance with one set of hyperparameters regardless of the number of training tokens, batch size, model size, and even model architecture. Our 3B dense and MoE models trained with the Power scheduler achieve comparable performance as state-of-the-art small language models. We open-source these pretrained models at https://ibm.biz/BdKhLa.

cs.CL↗

arXiv2024

Scaling Granite Code Models to 128K Context

This paper introduces long-context Granite code models that support effective context windows of up to 128K tokens. Our solution for scaling context length of Granite 3B/8B code models from 2K/4K to 128K consists of a light-weight continual pretraining by gradually increasing its RoPE base frequency with repository-level file packing and length-upsampled long-context data. Additionally, we also release instruction-tuned models with long-context support which are derived by further finetuning the long context base models on a mix of permissively licensed short and long-context instruction-response pairs. While comparing to the original short-context Granite code models, our long-context models achieve significant improvements on long-context tasks without any noticeable performance degradation on regular code completion benchmarks (e.g., HumanEval). We release all our long-context Granite code models under an Apache 2.0 license for both research and commercial use.

cs.AI↗

arXiv2024

Enhancing Training Efficiency Using Packing with Flash Attention

Padding is often used in tuning LLM models by adding special tokens to shorter training examples to match the length of the longest sequence in each batch. While this ensures uniformity for batch processing, it introduces inefficiencies by including irrelevant padding tokens in the computation and wastes GPU resources. Hugging Face SFT trainer has always offered the option to use packing to combine multiple training examples, allowing for maximal utilization of GPU resources. However, up till now, it did not offer proper masking of each packed training example. This capability has been added to Hugging Face Transformers 4.44. We analyse this new feature and show the benefits across different variations of packing.

cs.LG↗