Sahil Shah

Cursor Research, Aaron Chan, Ahmed Shalaby et al. · berkeley, microsoft-research

Composer 2 is a specialized model designed for agentic software engineering. The model demonstrates strong long-term planning and coding intelligence while maintaining the ability to efficiently solve problems for interactive use. The model is trained in two phases: first, continued pretraining to improve the model's knowledge and latent coding ability, followed by large-scale reinforcement learning to improve end-to-end coding performance through stronger reasoning, accurate multi-step execution, and coherence on long-horizon realistic coding problems. We develop infrastructure to support training in the same Cursor harness that is used by the deployed model, with equivalent tools and structure, and use environments that match real problems closely. To measure the ability of the model on increasingly difficult tasks, we introduce a benchmark derived from real software engineering problems in large codebases including our own. Composer 2 is a frontier-level coding model and demonstrates a process for training strong domain-specialized models. On our CursorBench evaluations the model achieves a major improvement in accuracy compared to previous Composer models (61.3). On public benchmarks the model scores 61.7 on Terminal-Bench and 73.7 on SWE-bench Multilingual in our harness, comparable to state-of-the-art systems.

Minkyu Choi, S P Sharan, Harsh Goel et al.

Current text-to-video (T2V) generation models are increasingly popular due to their ability to produce coherent videos from textual prompts. However, these models often struggle to generate semantically and temporally consistent videos when dealing with longer, more complex prompts involving multiple objects or sequential events. Additionally, the high computational costs associated with training or fine-tuning make direct improvements impractical. To overcome these limitations, we introduce NeuS-E, a novel zero-training video refinement pipeline that leverages neuro-symbolic feedback to automatically enhance video generation, achieving superior alignment with the prompts. Our approach first derives the neuro-symbolic feedback by analyzing a formal video representation and pinpoints semantically inconsistent events, objects, and their corresponding frames. This feedback then guides targeted edits to the original video. Extensive empirical evaluations on both open-source and proprietary T2V models demonstrate that NeuS-E significantly enhances temporal and logical alignment across diverse prompts by almost 40%

Shawn Liang, Sahil Shah, Chengwei Zhou et al.

Neuro-symbolic approaches to long-form video question answering (LVQA) have demonstrated significant accuracy improvements by grounding temporal reasoning in formal verification. However, existing methods incur prohibitive latency overheads, up to 90x slower than base VLM prompting, rendering them impractical for latency-sensitive edge deployments. We present LE-NeuS, a latency-efficient neuro-symbolic framework that preserves the accuracy benefits of temporal logic-guided video understanding while drastically reducing inference latency. Our key insight is that the dominant computational bottleneck arises from sequential and dense proposition detection across video frames during automaton construction. We address this through two principled optimizations: (1) CLIP guided two-stage adaptive sampling that exploits visual redundancy to skip semantically similar frames while preserving temporal boundaries, and (2) batched proposition detection that parallelizes VLM inference across temporal windows. Theoretically, we derive latency bounds as a function of video length, proposition complexity, and sampling density, establishing conditions under which latency efficiency is achievable. Empirically, on LongVideoBench and Video-MME benchmarks deployed on NVIDIA H100 GPUs, LE-NeuS reduces the latency gap from 90x to approximately 10x while maintaining >10% accuracy gains on temporally complex queries.

Sahil Shah, S P Sharan, Harsh Goel et al.

While vision-language models (VLMs) excel at tasks involving single images or short videos, they still struggle with Long Video Question Answering (LVQA) due to its demand for complex multi-step temporal reasoning. Vanilla approaches, which simply sample frames uniformly and feed them to a VLM along with the question, incur significant token overhead. This forces aggressive downsampling of long videos, causing models to miss fine-grained visual structure, subtle event transitions, and key temporal cues. Recent works attempt to overcome these limitations through heuristic approaches; however, they lack explicit mechanisms for encoding temporal relationships and fail to provide any formal guarantees that the sampled context actually encodes the compositional or causal logic required by the question. To address these foundational gaps, we introduce NeuS-QA, a training-free, plug-and-play neuro-symbolic pipeline for LVQA. NeuS-QA first translates a natural language question into a logic specification that models the temporal relationship between frame-level events. Next, we construct a video automaton to model the video's frame-by-frame event progression, and finally employ model checking to compare the automaton against the specification to identify all video segments that satisfy the question's logical requirements. Only these logic-verified segments are submitted to the VLM, thus improving interpretability, reducing hallucinations, and enabling compositional reasoning without modifying or fine-tuning the model. Experiments on the LongVideoBench and CinePile LVQA benchmarks show that NeuS-QA significantly improves performance by over 10%, particularly on questions involving event ordering, causality, and multi-step reasoning. We open-source our code at https://utaustin-swarmlab.github.io/NeuS-QA/.

Kaushik Lakshmiramanan, Vineeta Nair, Ching-Yi Lin et al.

This paper presents the design and implementation of an asynchronous delta modulator as a spike encoder for event-driven neural recording in a 65nm CMOS process. The proposed neuromorphic front-end converts analog signals into discrete, asynchronous ON and OFF spikes, effectively compressing continuous biopotentials into spike trains compatible with spiking neural networks (SNNs). Its asynchronous operation enables seamless integration with neuromorphic architectures for real-time decoding in closed-loop brain-machine interfaces (BMIs). Measurement results from silicon demonstrate an energy consumption of 60.73 nJ/spike, an F1-score of 80% compared to a behavioral model of the asynchronous delta modulator, and a compact pixel area of 73.45 um $\times$ 73.64 um.

Minkyu Choi, Harsh Goel, Mohammad Omama et al.

The unprecedented surge in video data production in recent years necessitates efficient tools to extract meaningful frames from videos for downstream tasks. Long-term temporal reasoning is a key desideratum for frame retrieval systems. While state-of-the-art foundation models, like VideoLLaMA and ViCLIP, are proficient in short-term semantic understanding, they surprisingly fail at long-term reasoning across frames. A key reason for this failure is that they intertwine per-frame perception and temporal reasoning into a single deep network. Hence, decoupling but co-designing semantic understanding and temporal reasoning is essential for efficient scene identification. We propose a system that leverages vision-language models for semantic understanding of individual frames but effectively reasons about the long-term evolution of events using state machines and temporal logic (TL) formulae that inherently capture memory. Our TL-based reasoning improves the F1 score of complex event identification by 9-15% compared to benchmarks that use GPT4 for reasoning on state-of-the-art self-driving datasets such as Waymo and NuScenes.

Ching-Yi Lin, Sahil Shah

ReRAM-based in-memory computing (IMC) architectures are promising candidates for energy-efficient matrix-vector multiplication. While scaling the size of ReRAM arrays allows for the amortization of power-hungry peripheral circuits like DACs and ADCs, it simultaneously introduces more parasitic along the signal path. Because of these challenges, current design methodologies often lack practical guidelines to balance these effects at early design stage, forcing designers to rely on time-consuming, iterative transistor-level simulations. In this work, we propose a comprehensive framework for design space exploration that enables the selection of optimal array size, ADC resolution, and system frequency without requiring exhaustive simulations. The framework utilizes a specialized testbench to extract parameters from a limited set of representative transistor-level simulations. These parameters are then used to accurately predict the performance of arbitrary architectures. We demonstrate the effectiveness of this framework through two realistic design cases aimed at maximizing energy efficiency (TOPs/s/W). The results show that the framework successfully identifies optimal architectural configurations under strict power and error constraints, providing an efficient path for high-performance IMC design.

Yihui Wang, Sheng-Yu Peng, Sahil Shah

This paper presents a fully synthesizable, treebased Address-Event Representation (AER) encoder designed for scalable neuromorphic computing systems. To achieve high throughput while maintaining strict compatibility with commercial EDA workflows, the asynchronous design employs a bundled-data protocol within a semi-decoupled micropipeline. The architecture replaces traditional transparent latches with standard edge-triggered flip-flops, enabling digital synthesis and place-and-route (PnR) using Cadence toolkits. A cross-coupled NAND-based random-priority arbiter is embedded within the encoder of each tree node to resolve event collisions efficiently. An 8-event AER prototype is fabricated in 65 nm CMOS technology utilizing a purely digital standard-cell flow. Post-fabrication silicon measurements validate the design, demonstrating a peak throughput of 33 MEvent/s and an average event latency of 50 ns, equating to a propagation delay of 17 ns/(event-bit). The design consumes only 435 fJ per encoded event.

Sahil Shah, Harsh Goel, Sai Shankar Narasimhan et al.

Modern video understanding systems excel at tasks such as scene classification, object detection, and short video retrieval. However, as video analysis becomes increasingly central to real-world applications, there is a growing need for proactive video agents for the systems that not only interpret video streams but also reason about events and take informed actions. A key obstacle in this direction is temporal reasoning: while deep learning models have made remarkable progress in recognizing patterns within individual frames or short clips, they struggle to understand the sequencing and dependencies of events over time, which is critical for action-driven decision-making. Addressing this limitation demands moving beyond conventional deep learning approaches. We posit that tackling this challenge requires a neuro-symbolic perspective, where video queries are decomposed into atomic events, structured into coherent sequences, and validated against temporal constraints. Such an approach can enhance interpretability, enable structured reasoning, and provide stronger guarantees on system behavior, all key properties for advancing trustworthy video agents. To this end, we present a grand challenge to the research community: developing the next generation of intelligent video agents that integrate three core capabilities: (1) autonomous video search and analysis, (2) seamless real-world interaction, and (3) advanced content generation. By addressing these pillars, we can transition from passive perception to intelligent video agents that reason, predict, and act, pushing the boundaries of video understanding.

Minkyu Choi, Yunhao Yang, Neel P. Bhatt et al.

Many robots (e.g., iRobot's Roomba) operate based on visual observations from live video streams, and such observations may inadvertently include privacy-sensitive objects, such as personal identifiers. Existing approaches for preserving privacy rely on deep learning models, differential privacy, or cryptography. They lack guarantees for the complete concealment of all sensitive objects. Guaranteeing concealment requires post-processing techniques and thus is inadequate for real-time video streams. We develop a method for privacy-constrained video streaming, PCVS, that conceals sensitive objects within real-time video streams. PCVS takes a logical specification constraining the existence of privacy-sensitive objects, e.g., never show faces when a person exists. It uses a detection model to evaluate the existence of these objects in each incoming frame. Then, it blurs out a subset of objects such that the existence of the remaining objects satisfies the specification. We then propose a conformal prediction approach to (i) establish a theoretical lower bound on the probability of the existence of these objects in a sequence of frames satisfying the specification and (ii) update the bound with the arrival of each subsequent frame. Quantitative evaluations show that PCVS achieves over 95 percent specification satisfaction rate in multiple datasets, significantly outperforming other methods. The satisfaction rate is consistently above the theoretical bounds across all datasets, indicating that the established bounds hold. Additionally, we deploy PCVS on robots in real-time operation and show that the robots operate normally without being compromised when PCVS conceals objects.

Ching-Yi Lin, Sahil Shah

Pre-trained vision transformers have achieved remarkable performance across various visual tasks but suffer from expensive computational and memory costs. While model quantization reduces memory usage by lowering precision, these models still incur significant computational overhead due to the dequantization before matrix operations. In this work, we analyze the computation graph and propose an integerization process based on operation reordering. Specifically, the process delays dequantization until after matrix operations. This enables integerized matrix multiplication and linear module by directly processing the quantized input. To validate our approach, we synthesize the self-attention module of ViT on a systolic array-based hardware. Experimental results show that our low-bit inference reduces per-PE power consumption for linear layer and matrix multiplication, bridging the gap between quantized models and efficient inference.

Everest Bloomer, Irem Didin, Ching-Yi Lin et al.

Computational workloads are growing exponentially, driving power consumption to unsustainable levels. Efficiently distributing large-scale networks is an NP-Complete problem equivalent to Boolean satisfiability (SAT), making it one of the core challenges in modern computation. To address this, physics and device inspired methods such as Ising systems have been explored for solving SAT more efficiently. In this work, we implement an Ising model equivalence of the 3-SAT problem using a ReRAM crossbar fabricated in the Skywater 130 nm CMOS process. Our ReRAM-based algorithm achieves $91.0\%$ accuracy in matrix representation across iterative reprogramming cycles. Additionally, we establish a foundational energy profile by measuring the energy costs of small sub-matrix structures within the problem space, demonstrating under linear growth trajectory for combining sub-matrices into larger problems. These results demonstrate a promising platform for developing scalable architectures to accelerate NP-Complete problem solving.

Mustafa Munir, Harsh Goel, Xiwen Wei et al.

Video editing and synthesis often introduce object inconsistencies, such as frame flicker and identity drift that degrade perceptual quality. To address these issues, we introduce ObjectAlign, a novel framework that seamlessly blends perceptual metrics with symbolic reasoning to detect, verify, and correct object-level and temporal inconsistencies in edited video sequences. The novel contributions of ObjectAlign are as follows: First, we propose learnable thresholds for metrics characterizing object consistency (i.e. CLIP-based semantic similarity, LPIPS perceptual distance, histogram correlation, and SAM-derived object-mask IoU). Second, we introduce a neuro-symbolic verifier that combines two components: (a) a formal, SMT-based check that operates on masked object embeddings to provably guarantee that object identity does not drift, and (b) a temporal fidelity check that uses a probabilistic model checker to verify the video's formal representation against a temporal logic specification. A frame transition is subsequently deemed "consistent" based on a single logical assertion that requires satisfying both the learned metric thresholds and this unified neuro-symbolic constraint, ensuring both low-level stability and high-level temporal correctness. Finally, for each contiguous block of flagged frames, we propose a neural network based interpolation for adaptive frame repair, dynamically choosing the interpolation depth based on the number of frames to be corrected. This enables reconstruction of the corrupted frames from the last valid and next valid keyframes. Our results show up to 1.4 point improvement in CLIP Score and up to 6.1 point improvement in warp error compared to SOTA baselines on the DAVIS and Pexels video datasets.

Sahil Shah, Vihari Piratla, Soumen Chakrabarti et al.

State-of-the-art NLP inference uses enormous neural architectures and models trained for GPU-months, well beyond the reach of most consumers of NLP. This has led to one-size-fits-all public API-based NLP service models by major AI companies, serving large numbers of clients. Neither (hardware deficient) clients nor (heavily subscribed) servers can afford traditional fine tuning. Many clients own little or no labeled data. We initiate a study of adaptation of centralized NLP services to clients, and present one practical and lightweight approach. Each client uses an unsupervised, corpus-based sketch to register to the service. The server uses an auxiliary network to map the sketch to an abstract vector representation, which then informs the main labeling network. When a new client registers with its sketch, it gets immediate accuracy benefits. We demonstrate the success of the proposed architecture using sentiment labeling, NER, and predictive language modeling

Kumar Ashutosh, Sarthak Consul, Bhishma Dedhia et al.

Policy Iteration (PI) is a classical family of algorithms to compute an optimal policy for any given Markov Decision Problem (MDP). The basic idea in PI is to begin with some initial policy and to repeatedly update the policy to one from an improving set, until an optimal policy is reached. Different variants of PI result from the (switching) rule used for improvement. An important theoretical question is how many iterations a specified PI variant will take to terminate as a function of the number of states $n$ and the number of actions $k$ in the input MDP. While there has been considerable progress towards upper-bounding this number, there are fewer results on lower bounds. In particular, existing lower bounds primarily focus on the special case of $k = 2$ actions. We devise lower bounds for $k \geq 3$. Our main result is that a particular variant of PI can take $Ω(k^{n/2})$ iterations to terminate. We also generalise existing constructions on $2$-action MDPs to scale lower bounds by a factor of $k$ for some common deterministic variants of PI, and by $\log(k)$ for corresponding randomised variants.

Sahil Shah, Naman Jain, Abhishek Sharma et al.

This paper provides a comprehensive and exhaustive study of adversarial attacks on human pose estimation models and the evaluation of their robustness. Besides highlighting the important differences between well-studied classification and human pose-estimation systems w.r.t. adversarial attacks, we also provide deep insights into the design choices of pose-estimation systems to shape future work. We benchmark the robustness of several 2D single person pose-estimation architectures trained on multiple datasets, MPII and COCO. In doing so, we also explore the problem of attacking non-classification networks including regression based networks, which has been virtually unexplored in the past. \par We find that compared to classification and semantic segmentation, human pose estimation architectures are relatively robust to adversarial attacks with the single-step attacks being surprisingly ineffective. Our study shows that the heatmap-based pose-estimation models are notably robust than their direct regression-based systems and that the systems which explicitly model anthropomorphic semantics of human body fare better than their other counterparts. Besides, targeted attacks are more difficult to obtain than un-targeted ones and some body-joints are easier to fool than the others. We present visualizations of universal perturbations to facilitate unprecedented insights into their workings on pose-estimation. Additionally, we show them to generalize well across different networks. Finally we perform a user study about perceptibility of these examples.