Massoud Pedram
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Massoud Pedram has a total of 273 co-authors. Here are the top co-authors by number of shared publications:
- Mehdi Kamal
University of Southern California, USA
56 shared publications - Ali Afzali-Kusha
University of Tehran
39 shared publications - Sasan Razmkhah
Ming Hsieh Department of Electrical and Computer Engineering, University of Southern California, Los Angeles, CA, USA
21 shared publications - Arash Fayyazi
University of Southern California Los Angeles California USA
17 shared publications - Mustafa Altay Karamuftuoglu
Ming Hsieh Department of Electrical and Computer Engineering, University of Southern California, Los Angeles, CA, USA
15 shared publications - Mahdi Nazemi
Department of Electrical & Computer Engineering University of Southern California Los Angeles CA USA
12 shared publications
Latest Publications
- Open Access
- 3 November 2026
- Published by MDPI AG in Preprints.org
Reasoning for Generation, Generation for Reasoning: A Survey of Unified Multimodal Models
Unified multimodal models (UMMs) integrate visual understanding and image generation, creating the opportunity for each capability to strengthen the other. However, architectural unification does not guarantee such interaction: models frequently understand visual requirements that they fail to realize in generated images, and generate intermediate images that their reasoning does not use. This survey examines both directions of transfer and why each remains challenging. We review the architectural foundations and representative model families, and the tasks in which understanding and generation must interact. For understanding → generation, we organize methods into training-free approaches that improve inference with frozen components and training-based approaches that learn reasoning, alignment, verification, or refinement, and compare them by stage of intervention, required access, and computational cost. For generation → reasoning, we review visual-imagination methods that reason with generated images, latent visual tokens, geometric representations, drawing tools, or external world models. We further examine evaluation protocols that distinguish individual task performance, cross-modal consistency, and genuine synergy. Across both directions, a shared challenge emerges: a correct intermediate, whether a reasoning trace or an imagined image, is not necessarily used. We conclude with open problems in reasoning-to-generation alignment, verifiable imagination, reliable self-correction, and compute-matched evaluation.
- 3 November 2026
- Published by MDPI AG in Preprints.org
Open AccessReasoning for Generation, Generation for Reasoning: A Survey of Unified Multimodal Models
Unified multimodal models (UMMs) integrate visual understanding and image generation, creating the opportunity for each capability to strengthen the other. However, architectural unification does not guarantee such interaction: models frequently understand visual requirements that they fail to realize in generated images, and generate intermediate images that their reasoning does not use. This survey examines both directions of transfer and why each remains challenging. We review the architectural foundations and representative model families, and the tasks in which understanding and generation must interact. For understanding → generation, we organize methods into training-free approaches that improve inference with frozen components and training-based approaches that learn reasoning, alignment, verification, or refinement, and compare them by stage of intervention, required access, and computational cost. For generation → reasoning, we review visual-imagination methods that reason with generated images, latent visual tokens, geometric representations, drawing tools, or external world models. We further examine evaluation protocols that distinguish individual task performance, cross-modal consistency, and genuine synergy. Across both directions, a shared challenge emerges: a correct intermediate, whether a reasoning trace or an imagined image, is not necessarily used. We conclude with open problems in reasoning-to-generation alignment, verifiable imagination, reliable self-correction, and compute-matched evaluation.
- Open Access
- 15 October 2026
- Published by MDPI AG in Preprints.org
Looped Language Models: A Survey and Meta-Analysis of Recurrent-Depth Transformers
Looped language models apply one shared block of Transformer layers repeatedly to the same tokens. Their effective depth, the number of layer applications per token, can grow without adding parameters and can be chosen at inference time. Recurrent depth is thus a way to scale test-time compute in latent space instead of in generated tokens. The idea appears under many names, including looped Transformers, LoopLMs, recurrent-depth models, Universal Transformers, Mixture-of-Recursions, and recursive reasoning models such as HRM and TRM. In two years it has moved from synthetic tasks to open models such as Huginn and Ouro and to pretraining on up to 28 trillion tokens. According to press reports, it is also used in a frontier-lab model. Its literature has split across architecture, training, systems, theory and safety. We present the first survey dedicated to looped language models, built on a systematic corpus of 557 core papers screened from 7,323 candidates. We introduce a unified prelude–core–coda formalism that makes explicit how parameters, compute, memory and depth are counted. We organize the field from design, training and scaling to inference systems, theory, mechanisms, capabilities, other modalities and safety. A meta-analysis classifies 822 verified empirical claims by comparison regime. It shows that most apparent contradictions about whether “a recurrence is worth a layer” come from comparing parameter-matched with compute-matched results. The evidence fits one reading, in which a recurrence is worth less than a fresh block of parameters. Looping therefore helps when parameters are the limiting resource and hurts when it replaces them at fixed compute. We derive a research agenda from measured gaps, including the absence of work on robustness, alignment and privacy. We release the corpus, the extraction records with verbatim evidence for every number, and the claims ledger.
- 15 October 2026
- Published by MDPI AG in Preprints.org
Open AccessLooped Language Models: A Survey and Meta-Analysis of Recurrent-Depth Transformers
Looped language models apply one shared block of Transformer layers repeatedly to the same tokens. Their effective depth, the number of layer applications per token, can grow without adding parameters and can be chosen at inference time. Recurrent depth is thus a way to scale test-time compute in latent space instead of in generated tokens. The idea appears under many names, including looped Transformers, LoopLMs, recurrent-depth models, Universal Transformers, Mixture-of-Recursions, and recursive reasoning models such as HRM and TRM. In two years it has moved from synthetic tasks to open models such as Huginn and Ouro and to pretraining on up to 28 trillion tokens. According to press reports, it is also used in a frontier-lab model. Its literature has split across architecture, training, systems, theory and safety. We present the first survey dedicated to looped language models, built on a systematic corpus of 557 core papers screened from 7,323 candidates. We introduce a unified prelude–core–coda formalism that makes explicit how parameters, compute, memory and depth are counted. We organize the field from design, training and scaling to inference systems, theory, mechanisms, capabilities, other modalities and safety. A meta-analysis classifies 822 verified empirical claims by comparison regime. It shows that most apparent contradictions about whether “a recurrence is worth a layer” come from comparing parameter-matched with compute-matched results. The evidence fits one reading, in which a recurrence is worth less than a fresh block of parameters. Looping therefore helps when parameters are the limiting resource and hurts when it replaces them at fixed compute. We derive a research agenda from measured gaps, including the absence of work on robustness, alignment and privacy. We release the corpus, the extraction records with verbatim evidence for every number, and the claims ledger.
- Open Access
- 15 June 2026
- Published by Association for Computing Machinery (ACM) in ACM Transactions on Design Automation of Electronic Systems
- Vol. 31 (6) , 1-19
GEMA+: A Hardware Accelerator for Exact Sequence Mapping Using Learned Indexing with Reduced Sensitivity to Read Length
This article presents the GEMA+ sequence mapping technique, characterized by its minimal sensitivity to the length of input read sequences. Unlike previously proposed methods, including GEMA and other similar learned index mapping techniques, GEMA+ effectively addresses the challenge of non-uniformity in the ascending trend of the mapping speed with the increasing read length. This challenge is tackled by our innovative process that eliminates the need for padding, which has traditionally been a source of performance variability. GEMA+ retains the efficient data structures of the original GEMA, ensuring no additional memory overhead. Additionally, we propose modifications to the hardware architecture of GEMA, resulting in an efficient implementation of GEMA+ on FPGA devices. The performance of the proposed technique is evaluated through post place-and-route simulations, which reveal that GEMA+ achieves a consistent and uniform increase in mapping speed as read length increases. By reducing performance sensitivity to input sequence length, GEMA+ demonstrates an average performance improvement of 42.5% for short reads and 11.8% for long reads, compared with the original GEMA.
- 15 June 2026
- Published by Association for Computing Machinery (ACM) in ACM Transactions on Design Automation of Electronic Systems
- Vol. 31 (6) , 1-19
Open AccessGEMA+: A Hardware Accelerator for Exact Sequence Mapping Using Learned Indexing with Reduced Sensitivity to Read Length
This article presents the GEMA+ sequence mapping technique, characterized by its minimal sensitivity to the length of input read sequences. Unlike previously proposed methods, including GEMA and other similar learned index mapping techniques, GEMA+ effectively addresses the challenge of non-uniformity in the ascending trend of the mapping speed with the increasing read length. This challenge is tackled by our innovative process that eliminates the need for padding, which has traditionally been a source of performance variability. GEMA+ retains the efficient data structures of the original GEMA, ensuring no additional memory overhead. Additionally, we propose modifications to the hardware architecture of GEMA, resulting in an efficient implementation of GEMA+ on FPGA devices. The performance of the proposed technique is evaluated through post place-and-route simulations, which reveal that GEMA+ achieves a consistent and uniform increase in mapping speed as read length increases. By reducing performance sensitivity to input sequence length, GEMA+ demonstrates an average performance improvement of 42.5% for short reads and 11.8% for long reads, compared with the original GEMA.
- 31 March 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Applied Superconductivity
- Vol. 36 (5) , 1-7
Superconducting Echo State Network for High-Speed SQUID Readout
Echo State Networks (ESNs), a type of reservoir computing that utilizes a recurrent neural network with a sparsely connected hidden layer, enable efficient temporal data processing with minimal training requirements. Conventional CMOS-based ESNs are limited in terms of speed and energy efficiency, which motivates the development of alternative implementations. We present a superconducting ESN that features ultra-low power, high-speed operation, and the nonlinear dynamics of Josephson Junctions to reconstruct waveforms from a digital Superconducting Quantum Interference Device (SQUID) magnetometer. The superconducting reservoir exhibits rich, re current dynamics, providing short-term memory and nonlinear transformation capabilities suitable for temporal pattern recognition. Experiments show that the superconducting ESN achieves high accuracy in reconstructing SQUID magnetic waveforms at GHz frequencies, producing real-time voltage output. The intrinsic dynamics of superconductivity enable low-power, high speed processing, outperforming conventional readout methods, and allowing direct in situ integration with sensors. These results establish superconducting ESNs as a scalable, noise-tolerant, and energy-efficient platform for high-speed SQUID readout and neuromorphic processing in cryogenic quantum systems.
- 31 March 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Applied Superconductivity
- Vol. 36 (5) , 1-7
Superconducting Echo State Network for High-Speed SQUID Readout
Echo State Networks (ESNs), a type of reservoir computing that utilizes a recurrent neural network with a sparsely connected hidden layer, enable efficient temporal data processing with minimal training requirements. Conventional CMOS-based ESNs are limited in terms of speed and energy efficiency, which motivates the development of alternative implementations. We present a superconducting ESN that features ultra-low power, high-speed operation, and the nonlinear dynamics of Josephson Junctions to reconstruct waveforms from a digital Superconducting Quantum Interference Device (SQUID) magnetometer. The superconducting reservoir exhibits rich, re current dynamics, providing short-term memory and nonlinear transformation capabilities suitable for temporal pattern recognition. Experiments show that the superconducting ESN achieves high accuracy in reconstructing SQUID magnetic waveforms at GHz frequencies, producing real-time voltage output. The intrinsic dynamics of superconductivity enable low-power, high speed processing, outperforming conventional readout methods, and allowing direct in situ integration with sensors. These results establish superconducting ESNs as a scalable, noise-tolerant, and energy-efficient platform for high-speed SQUID readout and neuromorphic processing in cryogenic quantum systems.
- 10 February 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Applied Superconductivity
- Vol. 36 (5) , 1-6
AC Power Delivery for Fast Phase Logic
Biasing remains one of the key scalability bottle necks in superconducting digital circuits, as traditional DC bias networks impose excessive power dissipation, large direct current requirements resulting in significant power overhead. Fast Phase Logic (FPL), an all-Josephson Junction (JJ) logic family, achieves ultra-high density by eliminating resistors and large inductors; however, existing FPL designs still rely on DC bias. This paper introduces a novel architecture: AC-biased FPL (ACFPL), which integrates single-phase AC/SFQ converters with FPL cells to deliver bias currents as quantized pulses, thereby eliminating the need for a global DC distribution network. Circuit-level simulations at 10 GHz demonstrate that AC biasing achieves up to 25× lower power consumption compared to DC bias for RSFQ circuits, and 14.3× reduction in power consumption for FPL circuits in our design-under-test (DUT), which combines FPL JTL, FPL DFF, and FPL AND cells. These results estab lish ACFPL as a scalable and energy-efficient alternative for large-scale superconducting digital systems, effectively combining compact FPL cell design with practical AC/SFQ bias delivery, allowing superconductor VLSI.
- 10 February 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Applied Superconductivity
- Vol. 36 (5) , 1-6
AC Power Delivery for Fast Phase Logic
Biasing remains one of the key scalability bottle necks in superconducting digital circuits, as traditional DC bias networks impose excessive power dissipation, large direct current requirements resulting in significant power overhead. Fast Phase Logic (FPL), an all-Josephson Junction (JJ) logic family, achieves ultra-high density by eliminating resistors and large inductors; however, existing FPL designs still rely on DC bias. This paper introduces a novel architecture: AC-biased FPL (ACFPL), which integrates single-phase AC/SFQ converters with FPL cells to deliver bias currents as quantized pulses, thereby eliminating the need for a global DC distribution network. Circuit-level simulations at 10 GHz demonstrate that AC biasing achieves up to 25× lower power consumption compared to DC bias for RSFQ circuits, and 14.3× reduction in power consumption for FPL circuits in our design-under-test (DUT), which combines FPL JTL, FPL DFF, and FPL AND cells. These results estab lish ACFPL as a scalable and energy-efficient alternative for large-scale superconducting digital systems, effectively combining compact FPL cell design with practical AC/SFQ bias delivery, allowing superconductor VLSI.
- 4 February 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Applied Superconductivity
- Vol. 36 (5) , 1-6
Programmable Bistable Vortex Logic for Scalable Superconductor Electronics
We present a reconfigurable logic circuit for Superconductor Electronics (SCE) based on a modified bistable vortex memory (BVM). Unlike conventional cell-based designs, this approach enables the direct modification of logic functions at the cell level by application of a control signal, thereby improving resource utilization and reducing redundant circuitry. The non-volatility of BVM enables on-chip logic reconfiguration without hardware changes or continuous power. Its high integration density and simple gate structure support higher logic density than most SCE logic families. This in-field programmability simplifies clocking, reduces area overhead, and supports scalable, energy-efficient designs. Circuit-level simulations verify the functionality and programmability of the proposed logic, while analysis of reconfiguration latency and energy efficiency demonstrates its feasibility for advanced superconductor computing architectures.
- 4 February 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Applied Superconductivity
- Vol. 36 (5) , 1-6
Programmable Bistable Vortex Logic for Scalable Superconductor Electronics
We present a reconfigurable logic circuit for Superconductor Electronics (SCE) based on a modified bistable vortex memory (BVM). Unlike conventional cell-based designs, this approach enables the direct modification of logic functions at the cell level by application of a control signal, thereby improving resource utilization and reducing redundant circuitry. The non-volatility of BVM enables on-chip logic reconfiguration without hardware changes or continuous power. Its high integration density and simple gate structure support higher logic density than most SCE logic families. This in-field programmability simplifies clocking, reduces area overhead, and supports scalable, energy-efficient designs. Circuit-level simulations verify the functionality and programmability of the proposed logic, while analysis of reconfiguration latency and energy efficiency demonstrates its feasibility for advanced superconductor computing architectures.
- 26 January 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Applied Superconductivity
- Vol. 36 (5) , 1-7
System-Level Comparison of Superconductor-Semiconductor Interface Circuits
Cryogenic computing systems often require communication between single flux quantum (SFQ) circuits at 4 K and CMOS circuits at a higher temperature stage. Interface circuits such as a 4JL (four-junction logic) gate and a SQUID (superconducting quantum interference device) stack have previously been used for SFQ-CMOS data links, but they have only been used for systems contained at 4 K. When communicating with CMOS circuits at a higher temperature stage, these interface circuits must be implemented in accordance with a holistic analysis of the entire cryogenic computing system. We have conducted an analysis in which we compare the impact of a 4JL gate and a SQUID stack on a system designed for the HYPRES ICE-T. For our analysis, we use the CORE e4 RSFQ microprocessor as the SFQ circuitry at 4 K. We provide guidelines for deciding between a 4JL gate and a SQUID stack as superconductor-semiconductor interface circuits in hybrid computing systems with multiple temperature stages.
- 26 January 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Applied Superconductivity
- Vol. 36 (5) , 1-7
System-Level Comparison of Superconductor-Semiconductor Interface Circuits
Cryogenic computing systems often require communication between single flux quantum (SFQ) circuits at 4 K and CMOS circuits at a higher temperature stage. Interface circuits such as a 4JL (four-junction logic) gate and a SQUID (superconducting quantum interference device) stack have previously been used for SFQ-CMOS data links, but they have only been used for systems contained at 4 K. When communicating with CMOS circuits at a higher temperature stage, these interface circuits must be implemented in accordance with a holistic analysis of the entire cryogenic computing system. We have conducted an analysis in which we compare the impact of a 4JL gate and a SQUID stack on a system designed for the HYPRES ICE-T. For our analysis, we use the CORE e4 RSFQ microprocessor as the SFQ circuitry at 4 K. We provide guidelines for deciding between a 4JL gate and a SQUID stack as superconductor-semiconductor interface circuits in hybrid computing systems with multiple temperature stages.
- 30 December 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Very Large Scale Integration (VLSI) Systems
- Vol. 34 (3) , 836-849
Flex-NTT: Design of a Flexible and Compact Number Theoretic Transform Architecture for Homomorphic Encryption Applications
This article presents Flex-NTT, a flexible (configurable) and area-efficient number theoretic transform (NTT) architecture featuring novel unified memory access patterns. The proposed design offers compile-time configurability (CTC) to support various parallel butterfly units (BUs) and run-time configurability (RTC) to accommodate diverse NTT operation sizes without recompilation. A single hardware instance is reconfigurable for NTT, inverse NTT (INTT), and point-wise multiplication, improving hardware utilization. The proposed coefficient and twiddle-factor memory access schemes achieve the theoretical minimum memory sizes, eliminate intermediate buffers, and enable natural-order outputs without additional reordering. FPGA evaluations demonstrate that Flex-NTT achieves up to $4.25 \times $ BRAM savings and $2.82 \times $ performance improvements over prior NTT/INTT designs. When applied to polynomial multiplication, it delivers up to $2.35 \times $ performance and $2.93 \times $ BRAM utilization improvements under the same metric.
- 30 December 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Very Large Scale Integration (VLSI) Systems
- Vol. 34 (3) , 836-849
Flex-NTT: Design of a Flexible and Compact Number Theoretic Transform Architecture for Homomorphic Encryption Applications
This article presents Flex-NTT, a flexible (configurable) and area-efficient number theoretic transform (NTT) architecture featuring novel unified memory access patterns. The proposed design offers compile-time configurability (CTC) to support various parallel butterfly units (BUs) and run-time configurability (RTC) to accommodate diverse NTT operation sizes without recompilation. A single hardware instance is reconfigurable for NTT, inverse NTT (INTT), and point-wise multiplication, improving hardware utilization. The proposed coefficient and twiddle-factor memory access schemes achieve the theoretical minimum memory sizes, eliminate intermediate buffers, and enable natural-order outputs without additional reordering. FPGA evaluations demonstrate that Flex-NTT achieves up to $4.25 \times $ BRAM savings and $2.82 \times $ performance improvements over prior NTT/INTT designs. When applied to polynomial multiplication, it delivers up to $2.35 \times $ performance and $2.93 \times $ BRAM utilization improvements under the same metric.
- 20 October 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Applied Superconductivity
- Vol. 35 (9) , 1-8
Scalable Asynchronous Single Flux Quantum Up-Down Counter Using Josephson Trapping Lines and $\alpha$-Cells
We present a scalable, asynchronous up-down counter architecture implemented using single-flux quantum logic to enable efficient state management in superconductor digital systems. The proposed design eliminates the reliance on clocked storage elements by introducing the Josephson trapping line. This bidirectional pulse-trapping structure enables persistent, nonvolatile state storage without clocking. The counter integrates $\alpha$-cells with a splitter element to make bidirectional data propagation possible and support multifanout connectivity. The design supports increment, decrement, and read operations and includes a control unit that guarantees correct output behavior across all valid state transitions. Circuit-level simulations based on SPICE models demonstrate robust bidirectional functionality across a 3-bit state range [−4 to +4] at an operating frequency of 4 GHz. The proposed counter offers a modular and scalable solution suitable for integration into larger superconducting systems targeting quantum computing, neuromorphic processing, and cryogenic sensing applications.
- 20 October 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Applied Superconductivity
- Vol. 35 (9) , 1-8
Scalable Asynchronous Single Flux Quantum Up-Down Counter Using Josephson Trapping Lines and $\alpha$-Cells
We present a scalable, asynchronous up-down counter architecture implemented using single-flux quantum logic to enable efficient state management in superconductor digital systems. The proposed design eliminates the reliance on clocked storage elements by introducing the Josephson trapping line. This bidirectional pulse-trapping structure enables persistent, nonvolatile state storage without clocking. The counter integrates $\alpha$-cells with a splitter element to make bidirectional data propagation possible and support multifanout connectivity. The design supports increment, decrement, and read operations and includes a control unit that guarantees correct output behavior across all valid state transitions. Circuit-level simulations based on SPICE models demonstrate robust bidirectional functionality across a 3-bit state range [−4 to +4] at an operating frequency of 4 GHz. The proposed counter offers a modular and scalable solution suitable for integration into larger superconducting systems targeting quantum computing, neuromorphic processing, and cryogenic sensing applications.
- Open Access
- 1 September 2025
- Published by International Joint Conferences on Artificial Intelligence
- Vol. 1, 367-375
Efficient Counterexample-Guided Fairness Verification and Repair of Neural Networks Using Satisfiability Modulo Convex Programming
Ensuring fairness is essential for ethical decision-making in various domains. Informally, a neural network is considered fair if and only if it treats similar individuals similarly in a given task. We introduce FaVeR (Fairness Verification and Repair), a framework for efficiently verifying and repairing pre-trained neural networks with respect to individual fairness properties. FaVeR ensures fairness via iterative search of high-sensitivity neurons and backward adjustment of their weights, guided by counterexamples generated from fairness verification using satisfiability modulo convex programming. By addressing fairness at the neuron level, FaVeR minimizes the impact of neural network repair on the overall performance. Experimental evaluations on common fairness datasets show that FaVeR achieves a 100% fairness repair rate across all models, with accuracy reduction of less than 2.27%. Moreover, its significantly lower average runtime makes it suitable for practical applications.
- 1 September 2025
- Published by International Joint Conferences on Artificial Intelligence
- Vol. 1, 367-375
Open AccessEfficient Counterexample-Guided Fairness Verification and Repair of Neural Networks Using Satisfiability Modulo Convex Programming
Ensuring fairness is essential for ethical decision-making in various domains. Informally, a neural network is considered fair if and only if it treats similar individuals similarly in a given task. We introduce FaVeR (Fairness Verification and Repair), a framework for efficiently verifying and repairing pre-trained neural networks with respect to individual fairness properties. FaVeR ensures fairness via iterative search of high-sensitivity neurons and backward adjustment of their weights, guided by counterexamples generated from fairness verification using satisfiability modulo convex programming. By addressing fairness at the neuron level, FaVeR minimizes the impact of neural network repair on the overall performance. Experimental evaluations on common fairness datasets show that FaVeR achieves a 100% fairness repair rate across all models, with accuracy reduction of less than 2.27%. Moreover, its significantly lower average runtime makes it suitable for practical applications.