arXiv subjects

Esin Tureci

Publications and source records attributed to Esin Tureci.

17 recordsLinked to original sources

arXiv2026

In-Context Learning Can Help Vision-Language Models Overcome Training Priors

Vision-language models (VLMs) learn strong statistical regularities during training, which can make them fail to properly respond to visual evidence when input images violate those regularities. It is challenging to disambiguate these prior-induced errors from errors caused by the model missing the underlying visual capabilities. In this work, we use controlled visual in-context learning to encourage the model to focus on the visual evidence, and demonstrate that this alone is frequently sufficient to correct the errors -- suggesting that the models in fact do possess the relevant visual capabilities. Specifically, we show that across five vision-centric tasks, demonstrations yield large gains on prior-conflicting counterfactual examples ($+27.2\%$ for Qwen3-VL 235B and $+15.1\%$ for Qwen3-VL 32B) while leaving accuracy on prior-aligned examples nearly unchanged. These findings generalize across five VLM families with gains from $+9.7\%$ to $+30.3\%$. We analyze visual representations and find that demonstrations indeed redirect attention toward grounding cues while reducing reliance on cues that support the learned prior. Finally, we find that in-context learning is limited when the underlying visual capability is weak or absent. Together, these results show that current VLMs possess visual grounding capabilities that standard evaluations may overlook, underscoring the need for controlled evaluation design in diagnosing and improving model capability.

cs.CV↗

arXiv2026

Robustness of Internal Signals for Hallucination Detection in Vision-Language Models

When given an image and a text prompt, vision-language models (VLMs) can produce seemingly sophisticated text responses. However, they are also prone to hallucinate objects in their output, which can have detrimental consequences in safety-critical applications. To detect these hallucinations, certain methods interestingly leverage internal signals from the VLM as it generates the text response; potentially, these methods use the same signals that led to a hallucination to detect it. We conduct an in-depth analysis to understand the robustness of these detection methods. First, we show that a naive baseline of a generated word's token position in the VLM text response achieves competitive performance. Further, we reveal how some detection methods rely on token position and use this heuristic as a shortcut. Second, we explore detection methods' sensitivity to language and vision modalities. Our experiments suggest that strong detection performance is often driven by detecting language-based hallucinations, and when methods rely on token position, they tend to be less visually grounded and exhibit reduced robustness when hallucinations occur earlier in a text response. Lastly, we demonstrate that different detection methods can carry complementary signals; when ensembling methods together, adding weaker methods to an ensemble of higher performing ones still offers performance improvements. Altogether, while there may be useful internal VLM signals to detect hallucinations, our investigation reveals important limitations of the current methods that try to leverage them.

cs.CV↗

arXiv2026

Adapter Thickets: Splitting an RLVR Budget Beats Concentrating It

Majority voting over sampled completions is the workhorse of test-time scaling, and reinforcement learning with verifiable rewards (RLVR) is the workhorse for making each completion better. The standard pipeline composes the two: train one policy with RLVR, then sample it many times and vote. We show that this composition is lossy. A vote can only overturn mistakes that its voters do not share, and RLVR sharpens a policy so that its samples increasingly make the same mistakes. With every method drawing exactly $160$ completions per problem, training a single LoRA adapter on the full RLVR budget raises single-sample accuracy on every model we test ($1.5$B-$8$B). Yet on three of four models it leaves the majority vote below that of the untrained base model, by up to $4.8$ points. The damage builds during training: voter errors grow steadily more correlated, and the majority vote accuracy peaks early before falling by up to $7.0$ points. The cause is concentration, not RLVR itself. We split the same data and training budget across $K$ LoRA adapters, each trained on its own random disjoint shard, and call the result an adapter thicket. Thickets out-vote the fully trained adapter in all $16$ (model, $K$) settings, and for $K{\geq}4$ they stay within $0.8$ points of the base model or above it. A single adapter stopped early, at a thicket member's step count, is a strong control that matches thickets for small $K$. For $K{\geq}8$, thickets keep more of RLVR's single-sample gain and out-vote this control in six of eight settings. The cost of concentration also grows with the number of votes: from $16$ to $160$ votes, the thicket's lead over the fully trained adapter widens from $1.3$ to $3.3$ points. When the plan is to sample and vote, an RLVR budget is better spent broad than deep.

cs.LG↗

arXiv2026

Depth as Time in One-Step Generative Models

The recent wave of one-step generative models, which compress the multi-step trajectory of diffusion via either distillation or learned flow maps, has reached an inflection point where they can generate high-quality images. Here, we ask a natural question that follows from these advances: what happens to the denoising trajectory of multi-step diffusion when generation is compressed into a single forward pass? We offer an empirical observation we call \textit{depth as time}: the denoising computation that multi-step diffusion performs across sampling steps appears to unfold across the depth of a single forward pass, and can be recovered by decoding intermediate layers with the model's own output head. Most interestingly, we show that this depthwise computation depends on the transport task a flow map is trained to solve. The most surprising case is MeanFlow, where probing shorter transport intervals reveals both denoising and renoising within a single network evaluation. In contrast, generators trained without a time-indexed transport task, such as drifting models, do not exhibit the same depthwise denoising. Consequently, we show that models that exhibit the depthwise denoising phenomenon are more compressible across the layerwise computation: a MeanFlow \texttt{SiT-L/2} model can be compressed by $16.6\times$ in parameters into a single time-conditioned block. We offer an explanation for this denoise-then-renoise behavior and show that, when we treat the layerwise computation explicitly as a flow, a single time-conditioned block can be trained to denoise across layers, compressing a MeanFlow \texttt{SiT-L/2} model by $16.6\times$ in parameters. Together, these results suggest that the temporal computation of diffusion is not eliminated by one-step generation, but reorganized across network depth.

cs.AI↗

arXiv2026

Beyond Objects: Contextual Synthetic Data Generation for Fine-Grained Classification

Text-to-image (T2I) models are increasingly used for synthetic dataset generation, but generating effective synthetic training data for classification remains challenging. Fine-tuning a T2I model with a few real examples can help improve the quality of synthetic training data; however, it may also cause overfitting and reduce diversity in the generated samples. We propose a fine-tuning strategy BOB (BeyondOBjects) to mitigate these concerns for fine-grained classification. Given a small set of real examples, we first extract class-agnostic attributes such as scene background and object pose. We then explicitly condition on these attributes during fine-tuning of the T2I model and marginalize them out during generation. This design mitigates overfitting, preserves the T2I model's generative prior, reduces estimation errors, and further minimizes unintended inter-class associations. Extensive experiments across multiple T2I models, backbones, and datasets show that our method achieves state-of-the-art performance in low-shot fine-grained classification when augmented with synthetic data. Concretely, BOB outperforms DataDream by 7.4% on the Aircraft dataset (from 50.0% to 57.4% when fine-tuning a CLIP classifier with five real images augmented with 100 synthetic images). In three of the four benchmarks, fine-tuning downstream models with 5 real images augmented with BOB achieves better performance than fine-tuning with 10 real images. Collectively, BOB outperforms prior art in 18 of 24 experimental settings, with 2+% accuracy improvements in 14 of these settings.

cs.CV↗

arXiv2026

Prioritize the Process, Not Just the Outcome: Rewarding Latent Thought Trajectories Improves Reasoning in Looped Language Models

Looped Language Models (LoopLMs) perform multi-step latent reasoning prior to token generation and outperform conventional LLMs on reasoning benchmarks at smaller parameter budgets. However, attempts to further improve LoopLM reasoning with reinforcement learning have failed - standard objectives such as Group Relative Policy Optimization (GRPO) only assign credit to the final latent state, creating a fundamental mismatch with the model's internal computation. To resolve this, we introduce RLTT (Reward Latent Thought Trajectories), a reinforcement learning framework which distributes reward across the full latent reasoning trajectory. RLTT provides dense, trajectory-level credit assignment without relying on external verifiers and can directly replace GRPO with negligible overhead. Across extensive experiments with Ouro-1.4B/2.6B-Thinking under identical training and inference conditions, RLTT yields statistically significant improvements over GRPO on challenging mathematical reasoning benchmarks, improving mean accuracy over MATH-500, AIME24/26, and BeyondAIME by +5.8% on the 1.4B scale, and +10.9% on the 2.6B scale. Despite being trained exclusively on mathematics, RLTT also transfers effectively to non-mathematical reasoning benchmarks, demonstrating the effectiveness of trajectory-level credit assignment for reinforcement learning in LoopLMs. Code is available at https://github.com/jonwill8/RLTT.git.

cs.LG↗

arXiv2026

Personalized Generative Models for Contextual Debiasing

Different visual patterns appear with different frequencies in the world: e.g., beach balls appear on sand more often than they do on a road. These statistics are reflected in vision datasets, and as a result trained models more easily recognize objects in common scenarios. However, recognizing a beach ball on a road may arguably be even more important than recognizing it on sand. We study how to mitigate this discrepancy. Since collecting uncommon images in the real world may be difficult, we explore whether generating images with less frequent contexts can serve as effective training augmentation. A key challenge is guiding generations to remain close to the original dataset distribution while creating diverse images with uncommon contexts. We introduce Decoupling Contextual Patterns with Generations (DecoupleGen), a method that personalizes text-to-image diffusion models to facilitate coherent synthesis of images with rare contexts while preserving original visual details. The generated images contain semantically meaningful content and remain visually aligned with the original datasets. We further apply verification constraints to ensure relevance of the augmented data. We evaluate our approach on object classification and recognition tasks on complex scene datasets. Our experiments demonstrate consistent improvements over previous approaches, and our analyses identify factors underlying these improvements.

cs.CV↗

arXiv2026

Visual Compositional Tuning

Visual instruction tuning (VIT) datasets have grown rapidly in scale, yet the informativeness of individual training samples has largely been overlooked. Recent dataset selection methods have shown that a small fraction of such datasets enriched with informative samples can lead to efficient finetuning of Multimodal Large Language Models. In this work, we explore the impact of sample complexity on informative data curation and introduce COMPACT (COMPositional Atomic-to-complex Visual Compositional Tuning), a compositional VIT data recipe that scales training sample complexity by combining multiple atomic visual capabilities in a single training example. Concretely, we synthesize rich and informative text questions for each image, allowing us to significantly reduce the number of training examples required for effective VIT. COMPACT demonstrates superior data efficiency compared to existing data reduction methods. When applied to the LLaVA-665K VIT dataset, COMPACT reduces the data budget by 90% while still achieving 100.2% of the full VIT performance (compared to only 97.5% by the state-of-the-art method) across eight multimodal benchmarks. Furthermore, training on the COMPACT data outperforms training on the full-scale VIT data on particularly complex benchmarks such as MM-Vet (+8.6%) and MMStar (+2.9%). COMPACT offers a scalable and efficient synthetic data generation recipe to improve on vision-language tasks.

cs.CV↗

arXiv2026

Bias at the End of the Score

Reward models (RMs) are inherently non-neutral value functions designed and trained to encode specific objectives, such as human preferences or text-image alignment. RMs have become crucial components of text-to-image (T2I) generation systems where they are used at various stages for dataset filtering, as evaluation metrics, as a supervisory signal during optimization of parameters, and for post-generation safety and quality filtering of T2I outputs. While specific problems with the integration of RMs into the T2I pipeline have been studied (e.g. reward hacking or mode collapse), their robustness and fairness as scoring functions remains largely unknown. We conduct a large scale audit of RM robustness with respect to demographic biases during T2I model training and generation. We provide quantitative and qualitative evidence that while originally developed as quality measures, RMs encode demographic biases, which cause reward-guided optimization to disproportionately sexualize female image subjects reinforce gender/racial stereotypes, and collapse demographic diversity. These findings highlight shortcomings in current reward models, challenge their reliability as quality metrics, and underscore the need for improved data collection and training procedures to enable more robust scoring.

cs.CV↗

arXiv2024

Tascade: Hardware Support for Atomic-free, Asynchronous and Efficient Reduction Trees

Graph search and sparse data-structure traversal workloads contain challenging irregular memory patterns on global data structures that need to be modified atomically. Distributed processing of these workloads has relied on server threads operating on their own data copies that are merged upon global synchronization. As parallelism increases within each server, the communication challenges that arose in distributed systems a decade ago are now being encountered within large manycore servers. Prior work has achieved scalability for sparse applications up to thousands of PUs on-chip, but does not scale further due to increasing communication distances and load-imbalance across PUs. To address these challenges we propose Tascade, a hardware-software co-design that offers support for storage-efficient data-private reductions as well as asynchronous and opportunistic reduction trees. Tascade introduces an execution model along with supporting hardware design that allows coalescing of data updates regionally and merges the data from these regions through cascaded updates. Together, Tascade innovations minimize communication and increase work balance in task-based parallelization schemes and scales up to a million PUs. We evaluate six applications and four datasets to provide a detailed analysis of Tascade's performance, power, and traffic-reduction gains over prior work. Our parallelization of Breadth-First-Search with RMAT-26 across a million PUs -- the largest of the literature -- reaches over 7600 GTEPS.

cs.AR↗

arXiv2024

Muchisim: A Simulation Framework for Design Exploration of Multi-Chip Manycore Systems

The design space exploration of scaled-out manycores for communication-intensive applications (e.g., graph analytics and sparse linear algebra) is hampered due to either lack of scalability or accuracy of existing frameworks at simulating data-dependent execution patterns. This paper presents MuchiSim, a novel parallel simulator designed to address these challenges when exploring the design space of distributed multi-chiplet manycore architectures. We evaluate MuchiSim at simulating systems with up to a million interconnected processing units (PUs) while modeling data movement and communication cycle by cycle. In addition to performance, MuchiSim reports the energy, area, and cost of the simulated system. It also comes with a benchmark application suite and two data visualization tools. MuchiSim supports various parallelization strategies and communication primitives such as task-based parallelization and message passing, making it highly relevant for architectures with software-managed coherence and distributed memory. Via a case study, we show that MuchiSim helps users explore the balance between memory and computation units and the constraints related to chiplet integration and inter-chip communication. MuchiSim enables evaluating new techniques or design parameters for systems at scales that are more realistic for modern parallel systems, opening the gate for further research in this area.

cs.AR↗

arXiv2024

DCRA: A Distributed Chiplet-based Reconfigurable Architecture for Irregular Applications

In recent years, the growing demand to process large graphs and sparse datasets has led to increased research efforts to develop hardware- and software-based architectural solutions to accelerate them. While some of these approaches achieve scalable parallelization with up to thousands of cores, adaptation of these proposals by the industry remained slow. To help solve this dissonance, we identified a set of questions and considerations that current research has not considered deeply. Starting from a tile-based architecture, we put forward a Distributed Chiplet-based Reconfigurable Architecture (DCRA) for irregular applications that carefully consider fabrication constraints that made prior work either hard or costly to implement or too rigid to be applied. We identify and study pre-silicon, package-time and compile-time configurations that help optimize DCRA for different deployments and target metrics. To enable that, we propose a practical path for manufacturing chip packages by composing variable numbers of DCRA and memory dies, with a software-configurable Torus network to connect them. We evaluate six applications and four datasets, with several configurations and memory technologies, to provide a detailed analysis of the performance, power, and cost of DCRA as a compute node for scale-out sparse data processing. Finally, we present our findings and discuss how DCRA's framework for design exploration can help guide architects to build scalable and cost-efficient systems for irregular applications.

cs.AR↗

arXiv2023

Co-Designed Superconducting Architecture for Lattice Surgery of Surface Codes with Quantum Interface Routing Card

Facilitating the ability to achieve logical qubit error rates below physical qubit error rates, error correction is anticipated to play an important role in scaling quantum computers. While many algorithms require millions of physical qubits to be executed with error correction, current superconducting qubit systems contain only hundreds of physical qubits. One of the most promising codes on the superconducting qubit platform is the surface code, requiring a realistically attainable error threshold and the ability to perform universal fault-tolerant quantum computing with local operations via lattice surgery and magic state injection. Surface code architectures easily generalize to single-chip planar layouts, however space and control hardware constraints point to limits on the number of qubits that can fit on one chip. Additionally, the planar routing on single-chip architectures leads to serialization of commuting gates and strain on classical decoding caused by large ancilla patches. A distributed multi-chip architecture utilizing the surface code can potentially solve these problems if one can optimize inter-chip gates, manage collisions in networking between chips, and minimize routing hardware costs. We propose QuIRC, a superconducting Quantum Interface Routing Card for Lattice Surgery between surface code modules inside of a single dilution refrigerator. QuIRC improves scaling by allowing connection of many modules, increases ancilla connectivity of surface code lattices, and offers improved transpilation of Pauli-based surface code circuits. QuIRC employs in-situ Entangled Pair (EP) generation protocols for communication. We explore potential topological layouts of QuIRC based on superconducting hardware fabrication constraints, and demonstrate reductions in ancilla patch size by up to 77.8%, and in layer transpilation size by 51.9% when compared to the single-chip case.

quant-ph↗

arXiv2023

Massive Data-Centric Parallelism in the Chiplet Era

Recent works have introduced task-based parallelization schemes to accelerate graph search and sparse data-structure traversal, where some solutions scale up to thousands of processing units (PUs) on a single chip. However parallelizing these memory-intensive workloads across millions of cores requires a scalable communication scheme as well as designing a cost-efficient computing node that makes multi-node systems practical, which have not been addressed in previous research. To address these challenges, we propose a task-oriented scalable chiplet architecture for distributed execution (Tascade), a multi-node system design that we evaluate with up to 256 distributed chips -- over a million PUs. We introduce an execution model that scales to this level via proxy regions and selective cascading, which reduce overall communication and improve load balancing. In addition, package-time reconfiguration of our chiplet-based design enables creating chip products that optimized post-silicon for different target metrics, such as time-to-solution, energy, or cost. We evaluate six applications and four datasets, with several configurations and memory technologies to provide a detailed analysis of the performance, power, and cost of data-centric execution at a massive scale. Our parallelization of Breadth-First-Search with RMAT-26 across a million PUs -- the largest of the literature -- reaches 3021 GTEPS.

cs.DC↗

arXiv2023

Microarchitectures for Heterogeneous Superconducting Quantum Computers

Noisy Intermediate-Scale Quantum Computing (NISQ) has dominated headlines in recent years, with the longer-term vision of Fault-Tolerant Quantum Computation (FTQC) offering significant potential albeit at currently intractable resource costs and quantum error correction (QEC) overheads. For problems of interest, FTQC will require millions of physical qubits with long coherence times, high-fidelity gates, and compact sizes to surpass classical systems. Just as heterogeneous specialization has offered scaling benefits in classical computing, it is likewise gaining interest in FTQC. However, systematic use of heterogeneity in either hardware or software elements of FTQC systems remains a serious challenge due to the vast design space and variable physical constraints. This paper meets the challenge of making heterogeneous FTQC design practical by introducing HetArch, a toolbox for designing heterogeneous quantum systems, and using it to explore heterogeneous design scenarios. Using a hierarchical approach, we successively break quantum algorithms into smaller operations (akin to classical application kernels), thus greatly simplifying the design space and resulting tradeoffs. Specializing to superconducting systems, we then design optimized heterogeneous hardware composed of varied superconducting devices, abstracting physical constraints into design rules that enable devices to be assembled into standard cells optimized for specific operations. Finally, we provide a heterogeneous design space exploration framework which reduces the simulation burden by a factor of 10^4 or more and allows us to characterize optimal design points. We use these techniques to design superconducting quantum modules for entanglement distillation, error correction, and code teleportation, reducing error rates by 2.6x, 10.7x, and 3.0x compared to homogeneous systems.

quant-ph↗

arXiv2023

Dalorex: A Data-Local Program Execution and Architecture for Memory-bound Applications

Applications with low data reuse and frequent irregular memory accesses, such as graph or sparse linear algebra workloads, fail to scale well due to memory bottlenecks and poor core utilization. While prior work with prefetching, decoupling, or pipelining can mitigate memory latency and improve core utilization, memory bottlenecks persist due to limited off-chip bandwidth. Approaches doing processing in-memory (PIM) with Hybrid Memory Cube (HMC) overcome bandwidth limitations but fail to achieve high core utilization due to poor task scheduling and synchronization overheads. Moreover, the high memory-per-core ratio available with HMC limits strong scaling. We introduce Dalorex, a hardware-software co-design that achieves high parallelism and energy efficiency, demonstrating strong scaling with >16,000 cores when processing graph and sparse linear algebra workloads. Over the prior work in PIM, both using 256 cores, Dalorex improves performance and energy consumption by two orders of magnitude through (1) a tile-based distributed-memory architecture where each processing tile holds an equal amount of data, and all memory operations are local; (2) a task-based parallel programming model where tasks are executed by the processing unit that is co-located with the target data; (3) a network design optimized for irregular traffic, where all communication is one-way, and messages do not contain routing metadata; (4) novel traffic-aware task scheduling hardware that maintains high core utilization; and (5) a data placement strategy that improves work balance. This work proposes architectural and software innovations to provide the greatest scalability to date for running graph algorithms while still being programmable for other domains.

cs.AR↗

arXiv2020

The MosaicSim Simulator (Full Technical Report)

As Moore's Law has slowed and Dennard Scaling has ended, architects are increasingly turning to heterogeneous parallelism and domain-specific hardware-software co-designs. These trends present new challenges for simulation-based performance assessments that are central to early-stage architectural exploration. Simulators must be lightweight to support rich heterogeneous combinations of general purpose cores and specialized processing units. They must also support agile exploration of hardware-software co-design, i.e. changes in the programming model, compiler, ISA, and specialized hardware. To meet these challenges, we introduce MosaicSim, a lightweight, modular simulator for heterogeneous systems, offering accuracy and agility designed specifically for hardware-software co-design explorations. By integrating the LLVM toolchain, MosaicSim enables efficient modeling of instruction dependencies and flexible additions across the stack. Its modularity also allows the composition and integration of different hardware components. We first demonstrate that MosaicSim captures architectural bottlenecks in applications, and accurately models both scaling trends in a multicore setting and accelerator behavior. We then present two case-studies where MosaicSim enables straightforward design space explorations for emerging systems, i.e. data science application acceleration and heterogeneous parallel architectures.

cs.AR↗