Vipin Chaudhary

Jiayi Yuan, Hongyi Liu, Shaochen Zhong et al.

Long context capability is a crucial competency for large language models (LLMs) as it mitigates the human struggle to digest long-form texts. This capability enables complex task-solving scenarios such as book summarization, code assistance, and many more tasks that are traditionally manpower-intensive. However, transformer-based LLMs face significant challenges with long context input due to the growing size of the KV cache and the intrinsic complexity of attending to extended inputs; where multiple schools of efficiency-driven approaches - such as KV cache quantization, token dropping, prompt compression, linear-time sequence models, and hybrid architectures - have been proposed to produce efficient yet long context-capable models. Despite these advancements, no existing work has comprehensively benchmarked these methods in a reasonably aligned environment. In this work, we fill this gap by providing a taxonomy of current methods and evaluating 10+ state-of-the-art approaches across seven categories of long context tasks. Our work reveals numerous previously unknown phenomena and offers insights - as well as a friendly workbench - for the future development of long context-capable LLMs. The source code is available at https://github.com/henryzhongsc/longctx_bench.

Wang Yang, Debargha Ganguly, Xinpeng Li et al.

Hybrid reasoning language models are commonly controlled through high-level Think/No-think instructions to regulate reasoning behavior, yet we found that such mode switching is largely driven by a small set of trigger tokens rather than the instructions themselves. Through attention analysis and controlled prompting experiments, we show that a leading ``Okay'' token induces reasoning behavior, while the newline pattern following ``</think>'' suppresses it. Based on this observation, we propose Mid-Think, a simple training-free prompting format that combines these triggers to achieve intermediate-budget reasoning, consistently outperforming fixed-token and prompt-based baselines in terms of the accuracy-length trade-off. Furthermore, applying Mid-Think to RL training after SFT reduces training time by approximately 15% while improving final performance of Qwen3-8B on AIME from 69.8% to 72.4% and on GPQA from 58.5% to 61.1%, demonstrating its effectiveness for both inference-time control and RL-based reasoning training.

Debargha Ganguly, Srinivasan Iyengar, Vipin Chaudhary et al.

Large Language Models (LLMs) have revolutionized natural language processing, yet they struggle with inconsistent reasoning, particularly in novel domains and complex logical sequences. This research introduces Proof of Thought, a framework that enhances the reliability and transparency of LLM outputs. Our approach bridges LLM-generated ideas with formal logic verification, employing a custom interpreter to convert LLM outputs into First Order Logic constructs for theorem prover scrutiny. Central to our method is an intermediary JSON-based Domain-Specific Language, which by design balances precise logical structures with intuitive human concepts. This hybrid representation enables both rigorous validation and accessible human comprehension of LLM reasoning processes. Key contributions include a robust type system with sort management for enhanced logical integrity, explicit representation of rules for clear distinction between factual and inferential knowledge, and a flexible architecture that allows for easy extension to various domain-specific applications. We demonstrate Proof of Thought's effectiveness through benchmarking on StrategyQA and a novel multimodal reasoning task, showing improved performance in open-ended scenarios. By providing verifiable and interpretable results, our technique addresses critical needs for AI system accountability and sets a foundation for human-in-the-loop oversight in high-stakes domains.

Vinooth Kulkarni, Jaehyun Lee, Adam Hutchings et al.

Dynamic quantum circuits with mid-circuit measurement and classical feedforward are essential for near-term algorithms such as error mitigation, adaptive phase estimation, and Variational Quantum Eigensolvers (VQE), yet transpiling these programs across frameworks remains challenging due to inconsistent support for control flow and measurement semantics. We present a transpilation pipeline that converts OpenQASM 3.0 programs with classical control structures (conditionals and bounded loops) into optimized CUDA-Q C++ kernels, leveraging CUDA-Q's native mid-circuit measurement and host-language control flow to translate dynamic patterns without static circuit expansion. Our open-source framework is validated on comprehensive test suites derived from IBM Quantum's classical feedforward guide, including conditional reset, if-else branching, multi-bit predicates, and sequential feedforward, and on VQE-style parameterized circuits with runtime parameter optimization. Experiments show that the resulting CUDA-Q kernels reduce circuit depth by avoiding branch duplication, improve execution efficiency via low-latency classical feedback, and enhance code readability by directly mapping OpenQASM 3.0 control structures to C++ control flow, thereby bridging OpenQASM 3.0's portable circuit specification with CUDA-Q's performance-oriented execution model for NISQ-era applications requiring dynamic circuit capabilities.

Caleb Tung, Abhinav Goel, Xiao Hu et al.

Computer vision is often performed using Convolutional Neural Networks (CNNs). CNNs are compute-intensive and challenging to deploy on power-contrained systems such as mobile and Internet-of-Things (IoT) devices. CNNs are compute-intensive because they indiscriminately compute many features on all pixels of the input image. We observe that, given a computer vision task, images often contain pixels that are irrelevant to the task. For example, if the task is looking for cars, pixels in the sky are not very useful. Therefore, we propose that a CNN be modified to only operate on relevant pixels to save computation and energy. We propose a method to study three popular computer vision datasets, finding that 48% of pixels are irrelevant. We also propose the focused convolution to modify a CNN's convolutional layers to reject the pixels that are marked irrelevant. On an embedded device, we observe no loss in accuracy, while inference latency, energy consumption, and multiply-add count are all reduced by about 45%.

Vinooth Kulkarni, Aaron Orenstein, Xinpeng Li et al.

The quantum computing community is increasingly positioning quantum processors as accelerators within classical HPC workflows, analogous to GPUs and TPUs. However, many real-world applications require scaling to hundreds or thousands of physical qubits to realize logical qubits via error correction. To reach these scales, hardware vendors employing diverse technologies -- such as trapped ions, photonics, neutral atoms, and superconducting circuits -- are moving beyond single, monolithic QPUs toward modular architectures connected via interconnects. For example, IonQ has proposed photonic links for scaling, while IBM has demonstrated a modular QPU architecture by classically linking two 127-qubit devices. Using dynamic circuits, Bell-pair-based teleportation, and circuit cutting, they have shown how to execute a large quantum circuit that cannot fit on a single QPU. As interest in quantum computing grows, cloud providers must ensure fair and efficient resource allocation for multiple users sharing such modular systems. Classical interconnection of QPUs introduces new scheduling challenges, particularly when multiple jobs execute in parallel. In this work, we develop a multi-programmable scheduler for modular quantum systems that jointly considers qubit mapping, parallel circuit execution, measurement synchronization across subcircuits, and teleportation operations between QPUs using dynamic circuits.

Mohsen Hariri, Michael Hinczewski, Jing Ma et al.

Test-time scaling evaluates reasoning LLMs by sampling multiple outputs per prompt, but ranking models in this regime remains underexplored. We formalize dense benchmark ranking under test-time scaling and introduce Scorio, a library that implements statistical ranking methods such as paired-comparison models, item response theory (IRT) models, voting rules, and graph- and spectral-based methods. Across $20$ reasoning models on four Olympiad-style math benchmarks (AIME'24, AIME'25, HMMT'25, and BrUMO'25; up to $N=80$ trials), most full-trial rankings agree closely with the Bayesian gold standard $\mathrm{Bayes}_{\mathcal{U}}@80$ (mean Kendall's $τ_b = 0.93$--$0.95$), and $19$--$34$ methods recover exactly the same ordering. In the single-trial regime, the best methods reach $τ_b \approx 0.86$. Using greedy decoding as an empirical prior ($\mathrm{Bayes}_{\mathbf{R}_0}@N$) reduces variance at $N=1$ by $16$--$52\%$, but can bias rankings when greedy and stochastic sampling disagree. These results identify reliable ranking methods for both high- and low-budget test-time scaling. We release Scorio as an open-source library at https://github.com/mohsenhariri/scorio.

Debargha Ganguly, Sreehari Sankar, Biyao Zhang et al.

Current approaches to LLM safety fundamentally rely on a brittle cat-and-mouse game of identifying and blocking known threats via guardrails. We argue for a fresh approach: robust safety comes not from enumerating what is harmful, but from deeply understanding what is safe. We introduce Trust The Typical (T3), a framework that operationalizes this principle by treating safety as an out-of-distribution (OOD) detection problem. T3 learns the distribution of acceptable prompts in a semantic space and flags any significant deviation as a potential threat. Unlike prior methods, it requires no training on harmful examples, yet achieves state-of-the-art performance across 18 benchmarks spanning toxicity, hate speech, jailbreaking, multilingual harms, and over-refusal, reducing false positive rates by up to 40x relative to specialized safety models. A single model trained only on safe English text transfers effectively to diverse domains and over 14 languages without retraining. Finally, we demonstrate production readiness by integrating a GPU-optimized version into vLLM, enabling continuous guardrailing during token generation with less than 6% overhead even under dense evaluation intervals on large-scale workloads.

Andrew Seohwan Yu, Mohsen Hariri, Xuecen Zhang et al.

Intelligent medical image segmentation methods are rapidly evolving and being increasingly applied, yet they face the challenge of domain transfer, where algorithm performance degrades due to different data distributions between source and target domains. To address this, we introduce a method for zero-shot, single-prompt segmentation of 3D knee MRI by adapting Segment Anything Model 2 (SAM2), a general-purpose segmentation model designed to accept prompts and retain memory across frames of a video. By treating slices from 3D medical volumes as individual video frames, we leverage SAM2's advanced capabilities to generate motion- and spatially-aware predictions. We demonstrate that SAM2 can efficiently perform segmentation tasks in a zero-shot manner with no additional training or fine-tuning, accurately delineating structures in knee MRI scans using only a single prompt. Our experiments on the Osteoarthritis Initiative Zuse Institute Berlin (OAI-ZIB) dataset reveal that SAM2 achieves high accuracy on 3D knee bone segmentation, with a testing Dice similarity coefficient of 0.9643 on tibia. We also present results generated using different SAM2 model sizes, different prompt schemes, as well as comparative results from the SAM1 model deployed on the same dataset. This breakthrough has the potential to revolutionize medical image analysis by providing a scalable, cost-effective solution for automated segmentation, paving the way for broader clinical applications and streamlined workflows.

Vikash Singh, Weicong Chen, Debargha Ganguly et al.

Estimating treatment effects from observational data requires choosing an adjustment set, but valid adjustment depends on an unknown causal graph. Graph misspecification can cause under-coverage, while graph-agnostic conformal wrappers may regain nominal coverage only through large padding. We introduce CausalGuard, a structure-weighted conformal framework that calibrates after aggregating graph-conditional doubly robust pseudo-outcomes. Candidate DAGs are proposed from an LLM-derived edge prior, pruned by conditional-independence tests, and reweighted by Bayesian Information Criterion. A composite nonconformity score then calibrates the posterior-weighted pseudo-outcome. CausalGuard provides distribution-free finite-sample marginal coverage for this aggregated pseudo-outcome; under causal identification, overlap, conditional-mean nuisance stability, and concentration on target-aligned valid adjustment strategies, its conditional mean converges to the true Conditional Average Treatment Effect. Across five benchmarks, CausalGuard attains mean coverage above the nominal 90% level for the directly evaluable target and reduces width when graph-agnostic conformal baselines require large padding. Stress tests show that CausalGuard suppresses invalid collider adjustment and remains stable under misspecified priors when the retained candidate set is data-supported.

Caleb Tung, Nicholas Eliopoulos, Purvish Jajal et al.

Computer vision often uses highly accurate Convolutional Neural Networks (CNNs), but these deep learning models are associated with ever-increasing energy and computation requirements. Producing more energy-efficient CNNs often requires model training which can be cost-prohibitive. We propose a novel, automated method to make a pretrained CNN more energy-efficient without re-training. Given a pretrained CNN, we insert a threshold layer that filters activations from the preceding layers to identify regions of the image that are irrelevant, i.e. can be ignored by the following layers while maintaining accuracy. Our modified focused convolution operation saves inference latency (by up to 25%) and energy costs (by up to 22%) on various popular pretrained CNNs, with little to no loss in accuracy.

Wanru Guo, Juan Xie, Binbin Wang et al.

Background: High-dimensional genomic data exhibit strong group correlation structures that challenge conventional feature selection methods, which often assume feature independence or rely on pre-defined pathways and are sensitive to outliers and model misspecification. Methods: We propose the Dorfman screening framework, a multi-stage procedure that forms data-driven variable groups via hierarchical clustering, performs group and within-group hypothesis testing, and refines selection using elastic net or adaptive elastic net. Robust variants incorporate OGK-based covariance estimation, rank-based correlation, and Huber-weighted regression to handle contaminated and non-normal data. Results: In simulations, Dorfman-Sparse-Adaptive-EN performed best under normal conditions, while Robust-OGK-Dorfman-Adaptive-EN showed clear advantages under data contamination, outperforming classical Dorfman and competing methods. Applied to NSCLC gene expression data for trametinib response, robust Dorfman methods achieved the lowest prediction errors and enriched recovery of clinically relevant genes. Conclusions: The Dorfman framework provides an efficient and robust approach to genomic feature selection. Robust-OGK-Dorfman-Adaptive-EN offers strong performance under both ideal and contaminated conditions and scales to ultra-high-dimensional settings, making it well suited for modern genomic biomarker discovery.

Andrew Seohwan Yu, Mohsen Hariri, Kunio Nakamura et al.

Vision language models (VLMs) have shown significant promise in visual grounding for images as well as videos. In medical imaging research, VLMs represent a bridge between object detection and segmentation, and report understanding and generation. However, spatial grounding of anatomical structures in the three-dimensional space of medical images poses many unique challenges. In this study, we examine image modalities, slice directions, and coordinate systems as differentiating factors for vision components of VLMs, and the use of anatomical, directional, and relational terminology as factors for the language components. We then demonstrate that visual and textual prompting systems such as labels, bounding boxes, and mask overlays have varying effects on the spatial grounding ability of VLMs. To enable measurement and reproducibility, we introduce \textbf{MIS-Ground}, a benchmark that comprehensively tests a VLM for vulnerabilities against specific modes of \textbf{M}edical \textbf{I}mage \textbf{S}patial \textbf{Ground}ing. We release MIS-Ground to the public at \href{https://anonymous.4open.science/r/mis-ground}{\texttt{anonymous.4open.science/r/mis-ground}}. In addition, we present \textbf{MIS-SemSam}, a low-cost, inference-time, and model-agnostic optimization of VLMs that improve their spatial grounding ability with the use of \textbf{Sem}antic \textbf{Sam}pling. We find that MIS-SemSam improves the accuracy of Qwen3-VL-32B on MIS-Ground by 13.06\%.

Tianyi Zhang, Mohsen Hariri, Shaochen Zhong et al.

Large-scale AI models, such as Large Language Models (LLMs) and Diffusion Models (DMs), have grown rapidly in size, creating significant challenges for efficient deployment on resource-constrained hardware. In this paper, we introduce Dynamic-Length Float (DFloat11), a lossless compression framework that reduces LLM and DM size by 30% while preserving outputs that are bit-for-bit identical to the original model. DFloat11 is motivated by the low entropy in the BFloat16 weight representation of LLMs, which reveals significant inefficiency in the existing storage format. By applying entropy coding, DFloat11 assigns dynamic-length encodings to weights based on frequency, achieving near information-optimal compression without any loss of precision. To facilitate efficient inference with dynamic-length encodings, we develop a custom GPU kernel for fast online decompression. Our design incorporates the following: (i) compact, hierarchical lookup tables (LUTs) that fit within GPU SRAM for efficient decoding, (ii) a two-phase GPU kernel for coordinating thread read/write positions using lightweight auxiliary variables, and (iii) transformer-block-level decompression to minimize latency. Experiments on Llama 3.3, Qwen 3, Mistral 3, FLUX.1, and others validate our hypothesis that DFloat11 achieves around 30% model size reduction while preserving bit-for-bit identical outputs. Compared to a potential alternative of offloading parts of an uncompressed model to the CPU to meet memory constraints, DFloat11 achieves 2.3--46.2x higher throughput in token generation. With a fixed GPU memory budget, DFloat11 enables 5.7--14.9x longer generation lengths than uncompressed models. Notably, our method enables lossless inference of Llama 3.1 405B, an 810GB model, on a single node equipped with 8x80GB GPUs. Our code is available at https://github.com/LeanModels/DFloat11.

Osama Zafar, Alexander Nemecek, Yiqian Zhang et al.

Standard PII filters often miss contextual data leakage in RAG systems, such as non-regulated attribute clusters that collectively identify individuals. We introduce a Privacy Policy Enforcement (PPE) framework using dual one-class density estimators with fused text embeddings and a calibrated abstain region for out-of-distribution inputs. Using an axis-stratified, multi-LLM synthetic data pipeline across medicine, finance, and law, we found that traditional Gaussian Mixture baselines fail on borderline-safe stress tests by focusing on linguistic register rather than content. Our proposed T3+OCSVM detector, trained on safe and borderline-safe data, achieves a borderline AUROC of 0.93+ while reducing false positives by 44-55 percentage points and maintaining millisecond latency. Compared to supervised MLP classifiers or 14B-parameter LLM judges, our framework offers superior operational suitability, as the former suffers from high abstention rates and the latter from latency and calibration issues. This methodology provides a robust stress-testing standard for any synthetic-data-trained classifier.

Debargha Ganguly, Debayan Gupta, Vipin Chaudhary

Deep neural networks (DNNs), while increasingly deployed in many applications, struggle with robustness against anomalous and out-of-distribution (OOD) data. Current OOD benchmarks often oversimplify, focusing on single-object tasks and not fully representing complex real-world anomalies. This paper introduces a new, straightforward method employing graph structures and topological features to effectively detect both far-OOD and near-OOD data. We convert images into networks of interconnected human understandable features or visual concepts. Through extensive testing on two novel tasks, including ablation studies with large vocabularies and diverse tasks, we demonstrate the method's effectiveness. This approach enhances DNN resilience to OOD data and promises improved performance in various applications.

Vikash Singh, Debargha Ganguly, Weicong Chen et al.

Continual test-time adaptation (CTTA) updates a pretrained model online on an unlabeled, non-stationary stream while anchoring it to a frozen source checkpoint. This anchor is useful only when the source remains reliable. On CCC-Hard, however, a ResNet-50 source falls to approximately $1.3\%$ top-$1$ accuracy, while existing source-anchored CTTA methods continue applying the same anchor strength. We call this failure mode blind anchoring and propose RMemSafe, a reliability-gated extension of ROID that uses the frozen source's normalized predictive entropy to attenuate all explicit source-coupled uses in the objective. When the source posterior approaches uniformity, the gate closes: the source anchor and agreement filter vanish, and the objective reduces to a source-agnostic fallback comprising ROID's base losses plus marginal calibration. Combined with ASR, RMemSafe achieves the lowest error on $8$ of $9$ matched-split continual-corruption cells and is the best reset-based method on all $9$, improving ROID+ASR by $1.05$~pp on ResNet-50 and $0.48$~pp on ViT-B/16. A controlled source-degradation sweep shows a $1.13{\times}$ shallower harm slope than ROID+ASR, consistent with the graceful-decay prediction. The entropy gate detects high-entropy source collapse, not confidently wrong low-entropy sources; this scope is explicitly evaluated and discussed.

Xuecen Zhang, Vipin Chaudhary

The rapid advancement of diffusion-based generative models has made face forgery detection a critical challenge in digital forensics. Current detection methods face two fundamental limitations: poor cross-domain generalization when encountering unseen forgery types, and substantial computational overhead that hinders deployment on resource-constrained devices. We propose LRD-Net (Lightweight Real-centered Detection Network), a novel framework that addresses both challenges simultaneously. Unlike existing dual-branch approaches that process spatial and frequency information independently, LRD-Net adopts a sequential frequency-guided architecture where a lightweight Multi-Scale Wavelet Guidance Module generates attention signals that condition a MobileNetV3-based spatial backbone. This design enables effective exploitation of frequency-domain cues while avoiding the redundancy of parallel feature extraction. Furthermore, LRD-Net employs a real-centered learning strategy with exponential moving average prototype updates and drift regularization, anchoring representations around authentic facial images rather than modeling diverse forgery patterns. Extensive experiments on the DiFF benchmark demonstrate that LRD-Net achieves state-of-the-art cross-domain detection accuracy, consistently outperforming existing methods. Critically, LRD-Net accomplishes this with only 2.63M parameters - approximately 9x fewer than conventional approaches - while achieving over 8x faster training and nearly 10x faster inference. These results demonstrate that robust cross-domain face forgery detection can be achieved without sacrificing computational efficiency, making LRD-Net suitable for real-time deployment in mobile authentication systems and resource-constrained environments.

Yanyan Zhang, Chaoda Song, Vikash Singh et al.

Vision-Language-Action (VLA) models achieve remarkable flexibility and generalization beyond classical control paradigms. However, most prevailing VLAs are trained under a single-frame observation paradigm, which leaves them structurally blind to temporal dynamics. Consequently, these models degrade severely in non-stationary scenarios, even when trained or finetuned on dynamic datasets. Existing approaches either require expensive retraining or suffer from latency bottlenecks and poor temporal consistency across action chunks. We propose Pace-and-Path Correction, a training-free, closed-form inference-time operator that wraps any chunked-action VLA. From a single quadratic cost, joint minimization yields a unified solution that decomposes orthogonally into two distinct channels. The pace channel compresses execution along the planned direction, while the path channel applies an orthogonal spatial offset, jointly absorbing the perceived dynamics within the chunk window. We evaluate our approach on a comprehensive diagnostic benchmark MoveBench designed to isolate motion as the sole controlled variable. Empirical results demonstrate that our framework consistently outperforms state-of-the-art training-free wrappers and dynamic-adaptive methods and improves success rates by up to 28.8% and 25.9% in absolute terms over foundational VLA models in dynamic-only and static-dynamic mixed environments, respectively.

Mingjian Lu, Pawan K. Tripathi, Mark Shteyn et al.

Segmentation architectures are typically benchmarked on single imaging modalities, obscuring deployment-relevant performance variations: an architecture optimal for one modality may underperform on another. We present a cross-modal evaluation framework for materials image segmentation spanning SEM, AFM, XCT, and optical microscopy. Our evaluation of six encoder-decoder combinations across seven datasets reveals that optimal architectures vary systematically by context: UNet excels for high-contrast 2D imaging while DeepLabv3+ is preferred for the hardest cases. The framework also provides deployment feedback via out-of-distribution detection and counterfactual explanations that reveal which microstructural features drive predictions. Together, the architecture guidance, reliability signals, and interpretability tools address a practical gap in materials characterization, where researchers lack tools to select architectures for their specific imaging setup or assess when models can be trusted on new samples.

Mohsen Hariri, Amirhossein Samandar, Michael Hinczewski et al.

Pass$@k$ is widely used to report performance for LLM reasoning, but it often yields unstable, misleading rankings, especially when the number of trials (samples) is limited and compute is constrained. We present a principled Bayesian evaluation framework that replaces Pass$@k$ and average accuracy over $N$ trials (avg$@N$) with posterior estimates of a model's underlying success probability and credible intervals, yielding stable rankings and a transparent decision rule for differences. Evaluation outcomes are modeled as categorical (not just 0/1) with a Dirichlet prior, giving closed-form expressions for the posterior mean and uncertainty of any weighted rubric and enabling the use of prior evidence when appropriate. Theoretically, under a uniform prior, the Bayesian posterior mean is order-equivalent to average accuracy (Pass$@1$), explaining its empirical robustness while adding principled uncertainty. Empirically, in simulations with known ground-truth success rates and on AIME'24/'25, HMMT'25, and BrUMO'25, the Bayesian/avg procedure achieves faster convergence and greater rank stability than Pass$@k$ and recent variants, enabling reliable comparisons at far smaller sample counts. The framework clarifies when observed gaps are statistically meaningful (non-overlapping credible intervals) versus noise, and it naturally extends to graded, rubric-based evaluations. Together, these results recommend replacing Pass$@k$ for LLM evaluation and ranking with a posterior-based, compute-efficient protocol that unifies binary and non-binary evaluation while making uncertainty explicit. Code is available at https://mohsenhariri.github.io/bayes-kit

Phat Thanh Dang, Saahil Thoppay, Wang Yang et al.

Large language models suffer issues when operated on long contexts that are larger than their training context length due to the standard position encoding for tokens in the attention layer. Tokens a long distance apart will rarely have an effect on each other and long prompts yield unexpected results. To solve this problem, we propose SELF (Self-Extend the Context Length With Logistic Growth Function): a solution of grouping consecutive tokens at varying group sizes using a logistic capacity equation combined with a constant group size at smaller relative distances. Our model had an increase in performance of up to 12% compared to the LongLM extension method in LEval (specifically on the Qwen model). On summarization related tasks in LongBench, our model performed up to 6.4% better than LongLM (specifically on the Llama-2-7b model). On reading comprehension tasks from LEval, our model performed up to 5.4% better than the LongLM. Our code is available at https://github.com/alexeipc/SELF-LLM.

Mohsen Hariri, Alan Luo, Weicong Chen et al.

Large Language Models (LLMs) suffer inference-time memory bottlenecks dominated by the attention Key-Value (KV) cache, which scales with model size and context length. While KV-cache quantization alleviates this cost, bit allocation between keys and values is often tuned heuristically, lacking theoretical grounding and generalizability. This paper proposes two theorems that anchor mixed-precision KV quantization in the intrinsic geometry of Transformer models. First, key projections systematically have larger spectral and Frobenius norms than value matrices, implying higher information density along the key path. Second, for any given memory budget, prioritizing precision for keys over values strictly reduces quantization error and better preserves accuracy. Empirical evaluations across various prominent LLMs and benchmarks show that key-favored allocations (e.g., 4-bit keys, 2-bit values) retain up to 98.3\% accuracy compared to uniform allocations (e.g., 4-bit for both), while conserving memory. These results transform bit allocation from ad hoc tuning into a theoretically grounded, geometry-driven design principle for efficient LLM inference. Source code is available at https://github.com/mohsenhariri/spectral-kv.

Mohsen Hariri, Michael Hinczewski, Vipin Chaudhary

Scorio.jl is a Julia package for evaluating and ranking systems from repeated responses to shared tasks. It provides a common tensor-based interface for direct score-based, pairwise, psychometric, voting, graph, and listwise methods, so the same benchmark can be analyzed under multiple ranking assumptions. We describe the package design, position it relative to existing Julia tools, and report pilot experiments on synthetic rank recovery, stability under limited trials, and runtime scaling.

Wang Yang, Xiang Yue, Vipin Chaudhary et al.

Recent advances leverage post-training to enhance model reasoning performance, which typically requires costly training pipelines and still suffers from inefficient, overly lengthy outputs. We introduce Speculative Thinking, a training-free framework that enables large reasoning models to guide smaller ones during inference at the reasoning level, distinct from speculative decoding, which operates at the token level. Our approach is based on two observations: (1) reasoning-supportive tokens such as "wait" frequently appear after structural delimiters like "\n\n", serving as signals for reflection or continuation; and (2) larger models exhibit stronger control over reflective behavior, reducing unnecessary backtracking while improving reasoning quality. By strategically delegating reflective steps to a more capable model, our method significantly boosts the reasoning accuracy of reasoning models while shortening their output. With the assistance of the 32B reasoning model, the 1.5B model's accuracy on MATH500 increases from 83.2% to 89.4%, marking a substantial improvement of 6.2%. Simultaneously, the average output length is reduced from 5439 tokens to 4583 tokens, representing a 15.7% decrease. Moreover, when applied to a non-reasoning model (Qwen-2.5-7B-Instruct), our framework boosts its accuracy from 74.0% to 81.8% on the same benchmark, achieving a relative improvement of 7.8%.

Shouren Wang, Wang Yang, Chuang Ma et al.

Hybrid-thinking language models expose explicit think and no-think modes, but current designs do not separate them cleanly. Even in no-think mode, models often emit long and self-reflective responses, causing reasoning leakage. Existing work reduces this issue through better data curation and multi-stage training, yet leakage remains because both modes are still encoded in the same feed-forward parameters. We propose Path-Lock Expert (PLE), an architecture-level solution that replaces the single MLP in each decoder layer with two semantically locked experts, one for think and one for no-think, while keeping attention, embeddings, normalization, and the language-model head shared. A deterministic control-token router selects exactly one expert path for the entire sequence, so inference preserves the dense model's per-token computation pattern and each expert receives mode-pure updates during supervised fine-tuning. Across math and science reasoning benchmarks, PLE maintains strong think performance while producing a substantially stronger no-think mode that is more accurate, more concise, and far less prone to reasoning leakage. On Qwen3-4B, for example, PLE reduces no-think reflective tokens on AIME24 from 2.54 to 0.39 and improves no-think accuracy from 20.67% to 40.00%, all while preserving think-mode performance. These results suggest that controllable hybrid thinking is fundamentally an architectural problem, and separating mode-specific feed-forward pathways is a simple and effective solution.

Hongyi Liu, Shaochen Zhong, Xintong Sun et al.

Finetuning LLMs with LoRA has gained significant popularity due to its simplicity and effectiveness. Often, users may even find pluggable, community-shared LoRAs to enhance their base models for a specific downstream task of interest; enjoying a powerful, efficient, yet customized LLM experience with negligible investment. However, this convenient share-and-play ecosystem also introduces a new attack surface, where attackers can distribute malicious LoRAs to a community eager to try out shared assets. Despite the high-risk potential, no prior art has comprehensively explored LoRA's attack surface under the downstream-enhancing share-and-play context. In this paper, we investigate how backdoors can be injected into task-enhancing LoRAs and examine the mechanisms of such infections. We find that with a simple, efficient, yet specific recipe, a backdoor LoRA can be trained once and then seamlessly merged (in a training-free fashion) with multiple task-enhancing LoRAs, retaining both its malicious backdoor and benign downstream capabilities. This allows attackers to scale the distribution of compromised LoRAs with minimal effort by leveraging the rich pool of existing shared LoRA assets. We note that such merged LoRAs are particularly infectious -- because their malicious intent is cleverly concealed behind improved downstream capabilities, creating a strong incentive for voluntary download -- and dangerous -- because under local deployment, no safety measures exist to intervene when things go wrong. Our work is among the first to study this new threat model of training-free distribution of downstream-capable-yet-backdoor-injected LoRAs, highlighting the urgent need for heightened security awareness in the LoRA ecosystem. Warning: This paper contains offensive content and involves a real-life tragedy.

Feng Luo, Yu-Neng Chuang, Guanchu Wang et al. · tencent-ai, tsinghua

The reasoning-capable large language models (LLMs) demonstrate strong performance on complex reasoning tasks but often suffer from overthinking, generating unnecessarily long chain-of-thought (CoT) reasoning paths for easy reasoning questions, thereby increasing inference cost and latency. Recent approaches attempt to address this challenge by manually deciding when to apply long or short reasoning. However, they lack the flexibility to adapt CoT length dynamically based on question complexity. In this paper, we propose Auto Long-Short Reasoning (AutoL2S), a dynamic and model-agnostic framework that enables LLMs to dynamically compress their generated reasoning path based on the complexity of the reasoning question. AutoL2S enables a learned paradigm, in which LLMs themselves can decide when longer reasoning is necessary and when shorter reasoning suffices, by training on data annotated with our proposed method, which includes both long and short CoT paths and a special <EASY> token. We then use <EASY> token to indicate when the model can skip generating lengthy CoT reasoning. This proposed annotation strategy can enhance the LLMs' ability to generate shorter CoT reasoning paths with improved quality after training. Extensive evaluation results show that AutoL2S reduces the length of reasoning generation by up to 57% without compromising performance, demonstrating the effectiveness of AutoL2S for scalable and efficient LLM reasoning.

Nengbo Wang, Xiaotian Han, Jagdip Singh et al.

Large language models (LLMs) have revolutionized natural language processing (NLP), particularly through Retrieval-Augmented Generation (RAG), which enhances LLM capabilities by integrating external knowledge. However, traditional RAG systems face critical limitations, including disrupted contextual integrity due to text chunking, and over-reliance on semantic similarity for retrieval. To address these issues, we propose CausalRAG, a novel framework that incorporates causal graphs into the retrieval process. By constructing and tracing causal relationships, CausalRAG preserves contextual continuity and improves retrieval precision, leading to more accurate and interpretable responses. We evaluate CausalRAG against regular RAG and graph-based RAG approaches, demonstrating its superiority across several metrics. Our findings suggest that grounding retrieval in causal reasoning provides a promising approach to knowledge-intensive tasks.

Wang Yang, Hongye Jin, Jingfeng Yang et al.

Supervised Fine-Tuning (SFT) has been a go-to and effective method for enhancing long chain-of-thought (CoT) reasoning in relatively small LLMs by fine-tuning them with long CoT responses from larger LLMs. To continually improve reasoning abilities, we can either collect new high-quality long CoT reasoning SFT data or repeatedly train on existing SFT datasets. However, acquiring new long CoT SFT data is costly and limited, while repeated training often results in a performance plateau or decline. To further boost the performance with the SFT data, we propose Thinking Preference Optimization (ThinkPO), a simple yet effective post-SFT method that enhances long CoT reasoning without requiring new long CoT responses. Instead, ThinkPO utilizes readily available or easily obtainable short CoT reasoning responses as rejected answers and long CoT responses as chosen answers for the same question. It then applies direct preference optimization to encourage the model to favor longer reasoning outputs. Experiments show that ThinkPO further improves the reasoning performance of SFT-ed models, e.g. it increases math reasoning accuracy of SFT-ed models by 8.6% and output length by 25.9%. Notably, ThinkPO is capable of continually boosting the performance of the publicly distilled SFT model, e.g., increasing the official DeepSeek-R1-Distill-Qwen-7B's performance on MATH500 from 87.4% to 91.2%.

Wang Yang, Zirui Liu, Hongye Jin et al.

Recent language models exhibit strong reasoning capabilities, yet the influence of long-context capacity on reasoning remains underexplored. In this work, we hypothesize that current limitations in reasoning stem, in part, from insufficient long-context capacity, motivated by empirical observations such as (1) higher context window length often leads to stronger reasoning performance, and (2) failed reasoning cases resemble failed long-context cases. To test this hypothesis, we examine whether enhancing a model's long-context ability before Supervised Fine-Tuning (SFT) leads to improved reasoning performance. Specifically, we compared models with identical architectures and fine-tuning data but varying levels of long-context capacity. Our results reveal a consistent trend: models with stronger long-context capacity achieve significantly higher accuracy on reasoning benchmarks after SFT. Notably, these gains persist even on tasks with short input lengths, indicating that long-context training offers generalizable benefits for reasoning performance. These findings suggest that long-context modeling is not just essential for processing lengthy inputs, but also serves as a critical foundation for reasoning. We advocate for treating long-context capacity as a first-class objective in the design of future language models.

Wang Yang, Chaoda Song, Xinpeng Li et al.

Existing Agent benchmarks suffer from two critical limitations: high environment interaction overhead (up to 41\% of total evaluation time) and imbalanced task horizon and difficulty distributions that make aggregate scores unreliable. To address these issues, we propose ACE-Bench built around a unified grid-based planning task, where agents must fill hidden slots in a partially completed schedule subject to both local slot constraints and global constraints. Our benchmark offers fine-grained control through two orthogonal axes: Scalable Horizons, controlled by the number of hidden slots $H$, and Controllable Difficulty, governed by a decoy budget $B$ that determines the number of globally misleading decoy candidates. Crucially, all tool calls are resolved via static JSON files under a Lightweight Environment design, eliminating setup overhead and enabling fast, reproducible evaluation suitable for training-time validation. We first validate that H and B provide reliable control over task horizon and difficulty, and that ACE-Bench exhibits strong domain consistency and model discriminability. We then conduct comprehensive experiments across 13 models of diverse sizes and families over 6 domains, revealing significant cross-model performance variation and confirming that ACE-Bench provides interpretable and controllable evaluation of agent reasoning.

Wang Yang, Hongye Jin, Shaochen Zhong et al.

Long-context capability is considered one of the most important abilities of LLMs, as a truly long context-capable LLM enables users to effortlessly process many originally exhausting tasks -- e.g., digesting a long-form document to find answers vs. directly asking an LLM about it. However, existing real-task-based long-context evaluation benchmarks have two major shortcomings. First, benchmarks like LongBench often do not provide proper metrics to separate long-context performance from the model's baseline ability, making cross-model comparison unclear. Second, such benchmarks are usually constructed with fixed input lengths, which limits their applicability across different models and fails to reveal when a model begins to break down. To address these issues, we introduce a length-controllable long-context benchmark and a novel metric that disentangles baseline knowledge from true long-context capabilities. Experiments demonstrate the superiority of our approach in effectively evaluating LLMs.

Andrew Ye, James Xu, Vidyut Veedgav et al.

We propose and study the integration of sentiment analysis and deep reinforcement learning ensemble algorithms for stock trading by evaluating strategies capable of dynamically altering their active agent given the concurrent market environment. In particular, we design a simple-yet-effective method for extracting financial sentiment and combine this with improvements on existing trading agents, resulting in a strategy that effectively considers both qualitative market factors and quantitative stock data. We show that our approach results in a strategy that is profitable, robust, and risk-minimal - outperforming the traditional ensemble strategy as well as single agent algorithms and market metrics. Our findings suggest that the conventional practice of switching and reevaluating agents in ensemble every fixed-number of months is sub-optimal, and that a dynamic sentiment-based framework greatly unlocks additional performance. Furthermore, as we have designed our algorithm with simplicity and efficiency in mind, we hypothesize that the transition of our method from historical evaluation towards real-time trading with live data to be relatively simple.

Wang Yang, Shouren Wang, Chaoda Song et al.

Group Relative Policy Optimization (GRPO) has become a key technique for improving reasoning abilities in large language models, yet its behavior under different domain sequencing strategies is poorly understood. In particular, the impact of sequential (one domain at a time) versus mixed-domain (multiple domain at a time) training in GRPO has not been systematically studied. We provide the first systematic analysis of training-order effects across math, science, logic, and puzzle reasoning tasks. We found (1) single-domain generalization is highly asymmetric: training on other domains improves math reasoning by approximately 25\% accuracy, while yielding negligible transfer to logic and puzzle; (2) cross-domain interactions are highly order-dependent: training in the order math$\rightarrow$science achieves 83\% / 41\% accuracy on math / science, while reversing the order to science$\rightarrow$math degrades performance to 77\% / 25\%; (3) no single strategy is universally optimal in multi-domain training: sequential training favors math (up to 84\%), mixed training favors science and logic, and poor ordering can incur large performance gaps (from 70\% to 56\%). Overall, our findings demonstrate that GRPO under multi-domain settings exhibits pronounced asymmetry, order sensitivity, and strategy dependence, highlighting the necessity of domain-aware and order-aware training design.

Chengwei Zhou, Vipin Chaudhary, Gourav Datta

The computational overhead of Vision Transformers in practice stems fundamentally from their deep architectures, yet existing acceleration strategies have primarily targeted algorithmic-level optimizations such as token pruning and attention speedup. This leaves an underexplored research question: can we reduce the number of stacked transformer layers while maintaining comparable representational capacity? To answer this, we propose a branch-based structural reparameterization technique that operates during the training phase. Our approach leverages parallel branches within transformer blocks that can be systematically consolidated into streamlined single-path models suitable for inference deployment. The consolidation mechanism works by gradually merging branches at the entry points of nonlinear components, enabling both feed-forward networks (FFN) and multi-head self-attention (MHSA) modules to undergo exact mathematical reparameterization without inducing approximation errors at test time. When applied to ViT-Tiny, the framework successfully reduces the original 12-layer architecture to 6, 4, or as few as 3 layers while maintaining classification accuracy on ImageNet-1K. The resulting compressed models achieve inference speedups of up to 37% on mobile CPU platforms. Our findings suggest that the conventional wisdom favoring extremely deep transformer stacks may be unnecessarily restrictive, and point toward new opportunities for constructing efficient vision transformers.

Alexander Nemecek, Zebin Yun, Zahra Rahmani et al.

As large language models (LLMs) become progressively more embedded in clinical decision-support, documentation, and patient-information systems, ensuring their privacy and trustworthiness has emerged as an imperative challenge for the healthcare sector. Fine-tuning LLMs on sensitive electronic health record (EHR) data improves domain alignment but also raises the risk of exposing patient information through model behaviors. In this work-in-progress, we present an exploratory empirical study on membership inference vulnerabilities in clinical LLMs, focusing on whether adversaries can infer if specific patient records were used during model training. Using a state-of-the-art clinical question-answering model, Llemr, we evaluate both canonical loss-based attacks and a domain-motivated paraphrasing-based perturbation strategy that more realistically reflects clinical adversarial conditions. Our preliminary findings reveal limited but measurable membership leakage, suggesting that current clinical LLMs provide partial resistance yet remain susceptible to subtle privacy risks that could undermine trust in clinical AI adoption. These results motivate continued development of context-aware, domain-specific privacy evaluations and defenses such as differential privacy fine-tuning and paraphrase-aware training, to strengthen the security and trustworthiness of healthcare AI systems.

Jierui Peng, Yanyan Zhang, Yicheng Duan et al.

The evaluation of Vision-Language-Action (VLA) agents is hindered by the coarse, end-task success metric that fails to provide precise skill diagnosis or measure robustness to real-world perturbations. This challenge is exacerbated by a fragmented data landscape that impedes reproducible research and the development of generalist models. To address these limitations, we introduce NEBULA, a unified ecosystem for single-arm manipulation that enables diagnostic and reproducible evaluation. NEBULA features a novel dual-axis evaluation protocol that combines fine-grained capability tests for precise skill diagnosis with systematic stress tests that measure robustness. A standardized API and a large-scale, aggregated dataset are provided to reduce fragmentation and support cross-dataset training and fair comparison. Using NEBULA, we demonstrate that top-performing VLAs struggle with key capabilities such as spatial reasoning and dynamic adaptation, which are consistently obscured by conventional end-task success metrics. By measuring both what an agent can do and when it does so reliably, NEBULA provides a practical foundation for robust, general-purpose embodied agents.

Shouren Wang, Wang Yang, Xianxuan Long et al.

Hybrid thinking enables LLMs to switch between reasoning and direct answering, offering a balance between efficiency and reasoning capability. Yet our experiments reveal that current hybrid thinking LLMs only achieve partial mode separation: reasoning behaviors often leak into the no-think mode. To understand and mitigate this, we analyze the factors influencing controllability and identify four that matter most: (1) larger data scale, (2) using think and no-think answers from different questions rather than the same question, (3) a moderate increase in no-think data number, and (4) a two-phase strategy that first trains reasoning ability and then applies hybrid think training. Building on these findings, we propose a practical recipe that, compared to standard training, can maintain accuracy in both modes while significantly reducing no-think output length (from $1085$ to $585$ on MATH500) and occurrences of reasoning-supportive tokens such as ``\texttt{wait}'' (from $5917$ to $522$ on MATH500). Our findings highlight the limitations of current hybrid thinking and offer directions for strengthening its controllability.

Biyao Zhang, Mingkai Zheng, Debargha Ganguly et al.

Training Large Language Models(LLMs) is one of the most compute-intensive tasks in high-performance computing. Predicting end-to-end training time for multi-billion parameter models distributed across hundreds of GPUs remains challenging due to complex interactions between transformer components, parallelism strategies(data, model, pipeline, tensor), and multi-tier communication. Learned models require costly sampling, while analytical models often struggle with real-world network and hardware complexities. We address this by decomposing LLMs into core computational primitives and modeling them with: (1) operator-level decomposition for fine-grained analysis; (2) lightweight sampling based hardware-aware prediction models for key operations; (3) an end-to-end prediction system integrating these components across complex parallelization strategies. Crucially, our methodology has been validated on two large-scale HPC systems. Our framework achieves low average prediction errors-4.98\% on Perlmutter(A100) and 9.38\% on Vista(GH200)-for models up to 20B parameters across 128 GPUs. Importantly, it runs entirely on CPUs, enabling rapid iteration over hardware configurations and training strategies without costly on-cluster experimentation.

Debargha Ganguly, Sumit Kumar, Ishwar Balappanawar et al.

Curating high-quality, domain-specific datasets is a major bottleneck for deploying robust vision systems, requiring complex trade-offs between data quality, diversity, and cost when researching vast, unlabeled data lakes. We introduce Labeling Copilot, the first data curation deep research agent for computer vision. A central orchestrator agent, powered by a large multimodal language model, uses multi-step reasoning to execute specialized tools across three core capabilities: (1) Calibrated Discovery sources relevant, in-distribution data from large repositories; (2) Controllable Synthesis generates novel data for rare scenarios with robust filtering; and (3) Consensus Annotation produces accurate labels by orchestrating multiple foundation models via a novel consensus mechanism incorporating non-maximum suppression and voting. Our large-scale validation proves the effectiveness of Labeling Copilot's components. The Consensus Annotation module excels at object discovery: on the dense COCO dataset, it averages 14.2 candidate proposals per image-nearly double the 7.4 ground-truth objects-achieving a final annotation mAP of 37.1%. On the web-scale Open Images dataset, it navigated extreme class imbalance to discover 903 new bounding box categories, expanding its capability to over 1500 total. Concurrently, our Calibrated Discovery tool, tested at a 10-million sample scale, features an active learning strategy that is up to 40x more computationally efficient than alternatives with equivalent sample efficiency. These experiments validate that an agentic workflow with optimized, scalable tools provides a robust foundation for curating industrial-scale datasets.

Weicong Chen, Vikash Singh, Zahra Rahmani et al.

Existing Log Anomaly Detection (LogAD) methods are often slow, dependent on error-prone parsing, and use unrealistic evaluation protocols. We introduce $K^4$, an unsupervised and parser-independent framework for high-performance online detection. $K^4$ transforms arbitrary log embeddings into compact four-dimensional descriptors (Precision, Recall, Density, Coverage) using efficient k-nearest neighbor (k-NN) statistics. These descriptors enable lightweight detectors to accurately score anomalies without retraining. Using a more realistic online evaluation protocol, $K^4$ sets a new state-of-the-art (AUROC: 0.995-0.999), outperforming baselines by large margins while being orders of magnitude faster, with training under 4 seconds and inference as low as 4 $μ$s.

Zirui Liu, Guanchu Wang, Shaochen Zhong et al.

With the rapid growth in model size, fine-tuning the large pre-trained language model has become increasingly difficult due to its extensive memory usage. Previous works usually focus on reducing the number of trainable parameters in the network. While the model parameters do contribute to memory usage, the primary memory bottleneck during training arises from storing feature maps, also known as activations, as they are crucial for gradient calculation. Notably, neural networks are usually trained using stochastic gradient descent. We argue that in stochastic optimization, models can handle noisy gradients as long as the gradient estimator is unbiased with reasonable variance. Following this motivation, we propose a new family of unbiased estimators called WTA-CRS, for matrix production with reduced variance, which only requires storing the sub-sampled activations for calculating the gradient. Our work provides both theoretical and experimental evidence that, in the context of tuning transformers, our proposed estimators exhibit lower variance compared to existing ones. By replacing the linear operation with our approximated one in transformers, we can achieve up to 2.7$\times$ peak memory reduction with almost no accuracy drop and enables up to $6.4\times$ larger batch size. Under the same hardware, WTA-CRS enables better down-streaming task performance by applying larger models and/or faster training speed with larger batch sizes.

Shi Yan, Taghi Ramazanian, Elham Sagheb et al.

Knee pain is undoubtedly the most common musculoskeletal symptom that impairs quality of life, confines mobility and functionality across all ages. Knee pain is clinically evaluated by routine radiographs, where the widespread adoption of radiographic images and their availability at low cost, make them the principle component in the assessment of knee pain and knee pathologies, such as arthritis, trauma, and sport injuries. However, interpretation of the knee radiographs is still highly subjective, and overlapping structures within the radiographs and the large volume of images needing to be analyzed on a daily basis, make interpretation challenging for both naive and experienced practitioners. There is thus a need to implement an artificial intelligence strategy to objectively and automatically interpret knee radiographs, facilitating triage of abnormal radiographs in a timely fashion. The current work proposes an accurate and effective pipeline for autonomous detection, localization, and classification of knee joint area in plain radiographs combining the You Only Look Once (YOLO v3) deep convolutional neural network with a large and fully-annotated knee radiographs dataset. The present work is expected to stimulate more interest from the deep learning computer vision community to this pragmatic and clinical application.