Papers

Automated cough analysis offers a path to low-cost respiratory screening, but most existing work stops at binary COVID-19 detection. A practical tool needs to tell apart several respiratory conditions from one cough recording on a consumer smartphone. We present CoughSense, a system that sorts cough recordings into five classes. These are healthy, COVID-19, asthma or respiratory condition, bronchitis, and pneumonia. We aggregated 18,301 recordings from four public datasets (Coswara, CoughVID, Virufy, and the West China Hospital Pediatric Cough Dataset) and used the OpenAI Whisper encoder as a pretrained backbone for cough disease classification. The main contribution is active-frame QKV attention pooling, which restricts attention to the first 200 of 1500 encoder tokens. This avoids the silence-dilution problem that arises because a 3-second cough fills only 150 tokens of Whisper's 30-second input window. Other training parts handle the 19 to 1 class imbalance and the four-dataset domain shift. These include WeightedRandomSampler, SpecAugment, Balanced Mixup with forced minority pairing, a supervised contrastive auxiliary loss, FiLM symptom conditioning, and gradient-reversal domain adaptation. A dual-encoder model fuses Whisper with the OPERA-CT respiratory foundation model through cross-attention. CoughSense (Whisper-tiny, 8.6M parameters) reached 82.3 percent balanced accuracy on five-fold cross-validation (macro-F1 of 0.817, AUC of 0.941). It beat an ImageNet-pretrained EfficientNet-B2 by 11.1 points and a ViT trained from scratch by 29.6 points. All five classes passed 74 percent recall and four of five passed 80 percent. The dual-encoder model reached 85.4 percent balanced accuracy. Active-frame pooling is the largest single contributor across all ablation components at 5.1 points, which should help any short-audio task using Whisper as a backbone.

Large language models remain vulnerable to jailbreak backdoor attacks, where adversaries poison safety alignment data to embed hidden triggers that bypass safety mechanisms. Existing defenses often require comprehensive attack information or multiple triggered examples, making them impractical when defenders only observe a single reported failure case without knowing whether it stems from a backdoor attack or a natural alignment bug. This paper presents Patcher, a post-hoc defense framework that repairs backdoored language models using only a single reported failure case and the model parameters. Patcher operates in two stages. First, it localizes backdoor triggers by computing response-conditioned gradient-based saliency scores and applying adaptive clustering to separate triggers from benign context. Second, it patches the model through a constrained fine-tuning objective that breaks the trigger-response association while preserving benign-task utility and robustness to non-triggered jailbreak attacks through KL-divergence constraints. We conduct extensive evaluations across multiple backdoor attack strategies and demonstrate that Patcher successfully localizes triggers and neutralizes backdoors while maintaining model utility. We further show robustness against adaptive attacks designed to evade our defense. This work represents a significant step toward practical defenses against training-time attacks in deployed language models.

Understanding how structured internal structure emerges during neural network training is central to the study of deep learning. We investigate this phenomenon through the group composition task, where a two-layer neural network is trained to predict $g_1 \star g_2$ for elements of a finite group $G$. By lifting the projected gradient flow to the Fourier domain, we demonstrate that the training dynamics are governed by a Riemannian gradient ascent on a representation-theoretic energy functional. We prove that, under random initialization, this flow drives each neuron to converge almost surely toward a single irreducible representation, while the cross-layer Fourier coefficients achieve a rotational rank-one alignment. This framework provides a representation-theoretic account of feature learning and characterizes a novel low-rank compression phenomenon for matrix-valued group representations. Moreover, for Abelian groups, we provide a complete population-level description: random initialization promotes uniform diversification across nontrivial representations and induces Haar-uniform phases, jointly approximating the indicator via a majority-vote mechanism. We further prove that both phase alignment and representation competition emerge with exponential convergence rates.

The rapid growth of large language model (LLM) inference services has increased the demand for efficient multi-tenant GPU scheduling. While modern inference runtimes such as vLLM improve throughput through continuous batching and optimized memory management, accurately estimating the runtime cost of heterogeneous inference requests remains challenging. In practice, admission-time workload estimates may deviate from observed execution behavior, leading to workload misclassification, queue imbalance, increased tail latency, and degraded Quality-of-Service (QoS). This paper presents DriftSched, a QoS-aware scheduling framework for multi-tenant LLM inference serving on NVIDIA L4 GPUs. DriftSched combines workload classification, token-budget estimation, tenant-aware queue management, and an online feedback mechanism to refine workload estimates using runtime observations. The framework evaluates FIFO, Priority, Weighted, Shortest-Job-First (SJF), and Aging Priority scheduling policies under heterogeneous multi-tenant workloads. Experimental results show that adaptive calibration reduces workload estimation error by an average of 38.8% (MAE) and 40.5% (RMSE), improving workload classification stability. Among all evaluated schedulers, SJF achieves the best overall performance, reducing median end-to-end latency by approximately 42% and P99 latency by approximately 16% relative to FIFO under sustained GPU contention. The results further indicate that scheduler selection has a greater impact on latency behavior than runtime calibration alone, while accurate workload characterization largely eliminates systematic estimation drift. This work contributes a reproducible framework for studying workload-estimation fidelity and QoS-aware scheduling in multi-tenant GPU inference systems.

Human Activity Recognition (HAR) using WiFi signals has emerged as a transformative technology for smart homes, healthcare monitoring, security systems, and ambient assisted living. Unlike traditional camera-based systems that raise significant privacy concerns and fail in low-light conditions, or wearable sensors that require user compliance, WiFi-based HAR is non-intrusive, privacy-preserving, cost-effective, and works seamlessly in any lighting condition. This paper presents a comprehensive approach to recognize three distinct human activities: "No Presence" (empty room), "Walking", and "Walking + Arm-waving" using the Wallhack1.8k WiFi spectrogram dataset. We propose three key improvements to address the main challenges in WiFi-based HAR. First, to address high performance variance, we implement ensemble learning with five different CNN architectures (Deep CNN, Wide CNN, MobileNetV2, ResNet50V2, and EfficientNetB0). Second, to address the small dataset size limitation, we apply aggressive data augmentation techniques including time-warping, frequency masking, and noise addition. Third, to evaluate real-world generalization capability, we perform cross-scenario evaluation (training on Line-of-Sight and testing on Non-Line-of-Sight) and cross-antenna evaluation (training on Biquad antenna and testing on PIFA antenna). Our ensemble model achieved a test accuracy of 94.87% on the LOS scenario with Biquad antenna, outperforming the best individual model by 0.66%. Data augmentation improved Random Forest performance from 60% to 95%. Cross-scenario evaluation showed minimal accuracy drops of only 1.37% and 2.07%, demonstrating strong generalization capabilities. The results indicate that the proposed approach is robust, reliable, and suitable for real-world deployment in diverse environments with different hardware configurations.

Augmenting large language models (LLMs) with retrieved web text has become a dominant paradigm, yet the web is not natively textual: existing systems depend on complex parsing pipelines that linearize HTML and discard layout, visual structure, and formatting. We introduce PixelRAG, a new retrieval-augmented method that represents websites in their native visual form and performs retrieval and reading entirely in pixel space, enabling an end-to-end architecture that eliminates text abstraction. PixelRAG is, to our knowledge, the first pipeline to operate over a full Wikipedia corpus in this form, scaling to a datastore of 30 million screenshot images with an efficient visual retrieval index. Built on an existing visual embedding model (i.e., Qwen3-VL-Embedding), PixelRAG further fine-tunes this model on screenshot data with carefully curated contrastive training data. Retrieved screenshots are then fed directly as pixel inputs to a VLM, without intermediate text conversion. PixelRAG consistently outperforms both no-retrieval and text-based RAG baselines, most surprisingly on widely studied text-centric tasks such as NQ and SimpleQA. It also achieves strong gains on multimodal open-domain QA (e.g., MMSearch), benchmarks over noisy news corpora (e.g., LiveVQA), and agentic benchmarks (e.g., MoNaCo), improving accuracy by up to 18.1% over text-based baselines. Finally, pixel representations enable a new efficiency lever for RAG through image compression, achieving up to 3x token cost reduction at lower resolutions while maintaining accuracy. Our results challenge the necessity of text representations in web retrieval, suggesting that web RAG can operate directly in the web's native visual form while improving both performance and efficiency.

Benchmarking shortest-path algorithms is commonly based on aggregate performance over heterogeneous graph sets, which limits insight into how different search paradigms react to instance structure. We adopt an instance-landscape view of graph benchmarking by embedding graphs into a low-cost structural feature space and clustering them into regions of similar structure. Three benchmark suites are studied: weighted Erdős–Rényi graphs, random geometric (wireless) graphs, and real-world road networks. We evaluate four representative shortest-path solvers spanning uninformed exact search (Dijkstra), bidirectional exact search (bidirectional Dijkstra), heuristic-guided exact search (A$^{*}$), and deque-based strategies (DEQ). Clustering robustness is analyzed under multiple feature-selection schemes, and runtime distributions are compared across landscape regions using non-parametric tests. While generator parameters induce stable structural regions, we find that feature-space similarity does not necessarily imply performance similarity: significant runtime shifts are frequently observed even within the same landscape region. A merged-suite analysis further shows that different benchmark families occupy largely disjoint regions. These results highlight both the potential and the limits of structural landscapes for the structure-aware benchmarking of shortest-path algorithms.

Turbulent combustion simulations are crucial for many scientific and engineering systems. However, the high cost to fully resolve the complex multiscale and multiphysics behavior makes direct simulation typically infeasible. The equivalent reactor network (ERN) approach attempts to improve computational efficiency by replacing a multidimensional turbulent simulation with a series of much cheaper 0-D and 1-D chemical reactors, providing a surrogate model that retains detailed chemistry at the cost of simplified flow physics. However, their development remains a challenge, often requiring either expert analysis, or automated approaches that sacrifice accuracy. In this work, we develop an automated machine-learning-assisted framework for constructing ERNs of the Sandia-D turbulent methane/air flame. Principal component analysis is first used to reduce high-dimensional thermochemical computational fluid dynamics (CFD) data to a low-dimensional latent space, where k-means clustering identifies physically interpretable flame regions used to initialize a reactor-network graph. This initialization is then refined using finite-difference gradient descent wrapped around non-differentiable Cantera reactor simulations. Across 30 RANS simulations spanning a range of pilot temperatures and inlet methane compositions, the optimized 7-reactor ERN achieves a maximum-temperature $R^2$ score of 0.7945 while preserving a $\sim6000\times$ speedup over the CFD solver. Outlet CO prediction remains more challenging, with a final $R^2$ score of $-0.4183$, but improves substantially from the unoptimized clustering initialization. These results show that unsupervised thermochemical feature extraction can provide effective physics-informed initializations for ERN construction, while gradient-based refinement can significantly improve predictive accuracy without manual reactor-network design.

Federated Learning (FL) combined with Split Learning (SL) is a privacy preserving paradigm that enables training deep neural networks (DNNs) on resource constrained devices while reducing overall training cost. However, determining the optimal split point, meaning the layer where the model is divided still remains a critical challenge, especially when clients have heterogeneous hardware capabilities. Fixed split points can overload weak devices and increase the communication and server load, which slows convergence and reduces stability. This paper introduces QSplitFL, a novel capability-aware Deep Q-Network (DQN) framework for optimal split point selection in Split learning based Federated Learning (SFL) environments. Unlike existing approaches that rely on high-dimensional model weight representations, QSplitFL employs a lightweight state representation derived directly from client hardware metrics, including CPU utilization, memory, battery level, and network latency. The proposed framework incorporates a decayed loss-drop reward function that prioritizes early convergence, and a committee-based DQN architecture with majority voting to mitigate reward hacking. Extensive experiments on MNIST, Fashion-MNIST, CIFAR-10, and CIFAR-100 datasets using CNN, ResNet50, MobileNetV4, and ConvNeXt architectures demonstrate that our approach achieves better convergence and higher accuracy compared to existing methods, while effectively adapting to heterogeneous device resources. The source code is publicly available at https://github.com/AIPO-Lab/QSplitFL.

As Multimodal Large Language Models (MLLMs) face growing privacy risks and regulatory constraints, machine unlearning (MU) has emerged as a crucial solution for removing sensitive data while preserving model performance. However, existing MU methods typically rely on visual data of the target concepts, which is often unavailable due to strict data retention policies, thus creating a demand for source-free unlearning approaches that operate without access to the target data. In this work, we propose Source-free Proxy Anchor Concept Erasure (SPACE), the first source-free unlearning framework specialized for MLLMs. SPACE consists of two stages: (1) Text-Guided Proxy Anchor Selection (TPAS), which retrieves semantically aligned proxy anchors from the shared feature space. (2) Dual-Constraint Semantic Isolation (DCSI), which optimizes these anchors to indirectly erase target concepts. DCSI confines updates to the null space of retained knowledge, ensuring structural integrity. We theoretically prove that SPACE strictly bounds the perturbation on retained knowledge and maximizes feature spectral entropy, thereby maintaining the model's performance. Furthermore, extensive experiments across six datasets show that SPACE achieves performance comparable to that of state-of-the-art data-dependent methods, validating its effectiveness in source-free MU scenarios. The source code will be released.

Fine-tuning safety aligned large language models (LLMs) on downstream data improves adaptation but may erode learned safety behavior. Existing methods use fixed safety examples, global constraints, or one-sided task filtering. Our diagnostics show task updates expose different safety constraints, motivating joint selection of relevant references and compatible task samples. We propose DualSelect, a coupled framework for task and reference selection that refreshes task conditioned safety references before filtering whole task samples compatible with the induced reference direction. Under a minimax view, DualSelect selects safety references with high preservation loss and task conflict, together with compatible task samples, through entropy-regularized scoring surrogates, lazy reference refresh, and gradient correction. On 1B-8B LLMs, DualSelect preserves safety without losing task utility; using the REDORCA judge, it improves Safety Avg. over the strongest baseline by at least 5.10 points and remains highest in Safety Avg. across judges with moderate overhead. This view extends to retention focused continual learning.

Privacy and data sharing are often in tension. Many organizations use synthetic data to reduce privacy risk and still share useful data. For tabular data, auditing privacy remains hard. In many cases, even humans cannot easily tell if a table is real or synthetic. In this paper, we propose a method based on LLM discrimination. We ask an LLM to classify each table sample as REAL or SYNTHETIC. We test two settings: C1 with table only, and C2 with table plus distributional metadata. We use LLaMA as an open model and Gemini as a reference model. In our experiments, we run three synthesis models, CTGAN, TVAE, and Gaussian Copula, on two public datasets, UCI Adult and ACS Census. We collect 451 valid trials. Our results show clear differences between models. On Adult, LLaMA reaches DRS=0% in reported cells, while Gemini reaches DRS=100% for CTGAN and TVAE. On Census, LLaMA predicts SYNTHETIC for most samples, while Gemini stays high in C1 but drops for CTGAN and TVAE in C2. We also compare with a classifier two-sample test (C2ST) and record linkage as distributional baselines, and with a human pilot of 2 annotators and 240 trials. Our results show that LLM discrimination is a practical privacy audit signal when model choice, per provider reporting, and data encoding are handled with care. For reproducibility, code and experiment scripts are available at https://github.com/SlokomManel/LLM-as-a-Discriminator.

Key-value (KV) cache quantization is widely used to reduce LLM inference memory, yet existing evaluations focus only on perplexity and accuracy without assessing the safety impact. In this study, we explore alignment preservation under KV cache quantization. Across eleven instruction-tuned models (3.8B-72B) and five benchmarks (1,894 prompts), we find that low-bit quantization can silently destroy safety alignment: Mistral-7B loses 15.2% of its refusals at only 1.03x perplexity, and no universal safe bit-width exists, with sharp model-specific phase transitions invisible to standard metrics. We identify that the root cause is geometric: safety features occupy a low-dimensional activation subspace $10^2$-$10^3\times$ more vulnerable to quantization noise than the full representation space perplexity averages over. We propose Per-Channel Reduction (PCR), a diagnostic that classifies each model into one of three mechanistic failure modes: outlier-crushes-safety, where safety lives in non-outlier channels collaterally damaged by outlier-driven scale factors; outlier-as-safety, where safety overlaps outlier channels and finer granularity cannot rescue it; and multi-layer dilution, where safety is distributed across many layers and per-layer fixes fail. PCR predicts the correct mitigation direction on all nine primary models and one held-out model from an independent family using 20 calibration prompts. PCR generalizes across unseen prompts, models, and quantizer families, including production quantizers (KIVI: up to 97.2% recovery) and rotation-based quantization (QuaRot-style Hadamard), succeeding where attention-based allocation methods fail. The resulting training-free protocol, requiring ~35 GPU-minutes, recovers up to 97% of lost alignment at under 1% latency and minimal memory overhead, addressing vulnerabilities confirmed in production vLLM serving with FP8 KV cache on NVIDIA GPUs.

LLM agents can fail silently by asserting task completion when the environment state shows otherwise. We study this failure mode, false success, across two agent benchmarks: 9,876 tau2-bench trajectories from 8 model families and 1,879 AppWorld trajectories from 4 model families with text-independent ground truth. False success is common but varies by setting: 45–48% of failures in single-control tau2-bench domains, 3% in dual-control telecom, and 75.8% among AppWorld self-assessing coding-agent trajectories with explicit status claims. LLM judges fail reliably: no configuration across 5 judges, 5 prompt strategies, and full task specifications exceeds AUROC 0.65 on tau2-bench, and the same judges reach only 0.54 AUROC on AppWorld API-call traces. Judges rely on surface completion proxies – confident closing language in tau2-bench and coarse action-sequence volume in AppWorld – rather than verified state changes. Lightweight TF-IDF detectors achieve task-disjoint AUROC 0.83 on tau2-bench and 0.95 on AppWorld, recovering 4–8x more false successes than the best judge at the same flag rate with 3,300x lower latency. These results suggest that production monitoring should use lightweight, domain-calibrated detectors as triage signals rather than relying on LLM judges as the primary monitor for false success.

This study investigates the trade-offs between fidelity, privacy, and utility in synthetic data generation under conditions of data scarcity and privacy sensitivity. We propose an evaluation framework that jointly assesses these three dimensions and apply it to three widely used generative models, VAE, GAN, and DDPM. The evaluation spans three image datasets, MNIST, OCTMNIST, and OrganAMNIST, encompassing both general-purpose and medical imaging domains. Notable differences arise between the three models in their behaviour when differential privacy mechanisms are introduced during training. GAN and DDPM demonstrate greater robustness, maintaining higher fidelity and downstream utility across a range of noise levels, while VAE degrades more rapidly as privacy constraints increase. This study highlights the importance of a multidimensional evaluation of deep generative models, also noting that their behaviour significantly differs when privacy techniques are applied.

We present a systematic study of domain-dependent safety behavior in open-weight LLMs: 7 standardized experiments across 7 ethical domains, testing 5 models (12B–70B) in 4,200 interactions with dual-judge validation. Using a dual-condition methodology, each scenario tested in both an analytical framing (identify the harm) and an operational framing (help commit the harm), we find compliance rates vary from 14.7% (human trafficking) to 85.7% (surveillance design), a 71-percentage-point span with non-overlapping cluster-bootstrapped 95% CIs. Trustworthy deployment requires predictable safety behavior, yet we find compliance is highly context-dependent: the same model (Mistral Nemo 12B) provides surveillance designs in 100% of requests but assists with trafficking in only 26.7%. This unpredictability is opaque to deployers: the technical framing bypass, where harmful requests reframed as engineering problems override safety training without any external signal that refusal thresholds have shifted. Within-domain heterogeneity reaches 84.4pp, meaning safety behavior cannot be predicted even at the domain level. A replication on five frontier closed models (GPT-4.1/5.2, Claude Haiku/Sonnet/Opus 4.x; n=4,163 responses) accessed via the GitHub Copilot CLI deployed-product surface reproduces the same domain stratification, attenuated in absolute level but identical in shape, with the two low-codification domains (science fraud, surveillance) again the most permissive. These results show that current safety mechanisms lack the transparency and consistency required for trustworthy AI deployment.

The validation of advanced nuclear reactor designs and fuel concepts will require the design of new critical experiments with high neutronic similarity to the target technology. Neutronic similarity can be quantified by the correlation coefficient $c_k$, which captures the shared bias in $k_\text{eff}$ induced by uncertainties in nuclear data. Generally, a $c_k\geq0.9$ is needed for an experiment to be sufficiently similar to a target technology. In this work, a physics-informed deep neural network is trained to predict the neutronic sensitivity of grid-based critical experiment geometries. The differentiability of the neural network is used to enable gradient-based design optimization of new experiment geometries to maximize $c_k$ with the sensitivity profile of a target technology. This approach allows for optimization over the combinatorial design space of potential material combinations within the grid, moving beyond traditional parametric optimization approaches. The method is applied to the validation of the TN-Americas TN-LC transportation cask with HALEU fuel, for which existing critical experiment coverage is limited. This application is shown to produce experiment geometries achieving $c_k$ scores of 0.97757, 0.81324, and 0.93276 for three configurations of interest.

Transformers have become the standard solution for various AI tasks, with the query, key, and value (QKV) attention formulation playing a central role. However, the individual contribution of these three projections and the impact of omitting some remain poorly understood. We systematically evaluate three projection sharing constraints: a) Q-K=V (shared key-value), b) Q=K-V (shared query-key), and c) Q=K=V (single projection). The last two variants produce symmetric attention maps; to address this, we also explore asymmetric attention via 2D positional encodings. Through experiments spanning synthetic tasks, vision (MNIST, CIFAR, TinyImageNet, anomaly), and language modeling (300M and 1.2B parameter models on 10B tokens), we discovered that our transformers perform on par or occasionally better than the QKV transformer. In language modeling, Q-K=V projection sharing achieves 50% KV cache reduction with only 3.1% perplexity degradation. Crucially, projection sharing is complementary to head sharing (GQA/MQA): combining Q-K=V with GQA-4 yields 87.5% cache reduction, while Q-K=V + MQA achieves 96.9%, enabling practical on-device inference. We show that Q-K=V preserves quality because keys and values can occupy similar representational spaces and attention operates in a low-rank regime, whereas Q=K-V breaks attention directionality. Our results systematically characterize projection sharing as an underexplored instance of weight tying in attention, with direct, quantifiable inference memory benefits, particularly valuable for edge deployment. The code is publicly available at https://github.com/Brainchip-Inc/Do-Transformers-Need-3-Projections

Leader-follower gradient iterations, including bilevel and two-time-scale methods, can exhibit large finite-time excursions despite asymptotic stability. For cascaded iterations with block matrix J = [[A,0],[C,D]], symmetric diagonal blocks of norm at most gamma < 1, and coupling norm at most c, we prove an exact, dimension-free minimax formula: at every horizon t, the worst-case amplification equals the spectral norm of [[gamma^t,0],[c t gamma^(t-1),gamma^t]], attained by an aligned two-dimensional quadratic. Consequently, the worst transient is on the order of 1 + c/(1-gamma), and eigenvalues alone cannot predict finite-time gain. Under martingale noise, the aligned system has stationary follower variance sigma^2 c^2 (1+gamma^2)/(1-gamma^2)^3, establishing cubic inverse-gap scaling. For quadratic followers, the coupling is linked to follower sensitivity R by C = (I-D)R. When the leader is isotropic and the follower map is positive semidefinite, amplification is at most 1 plus the operator norm of R, independently of the spectral gaps. A look-ahead Newton update eliminates follower memory and is minimax-optimal among linear updates that track the leader exactly. Experiments validate the bounds and show that accelerated or inexact follower solvers can exceed the sensitivity bound when their effective maps have negative eigenvalues. Code and experiment scripts: https://github.com/Axym-Labs/transient-amplification

Reinforcement Learning (RL) has received increasing attention and adoption in real-world use cases. Most of these systems follow a train-then-fix paradigm, where trained agents do not learn while interacting with the world until performance degrades and retraining becomes necessary. In this position paper, we argue that deploying an agent that is incapable of optimality, but receives an evaluative reward signal, is inherently a continual RL problem. We identify four sources of non-stationarity after deployment that necessitate never-ending learning, and highlight why the best deployed agents never stop adapting. We analyze successful examples of continual RL in the real world, and present the community with the advantages and measures to move away from the current train-then-fix paradigm.

The IEEE P3109 draft standard defines a parameterized family of binary floating-point formats and associated operations, with a focus on facilitating machine learning. These formats allow efficient and consistent representation of values in a small number of bits. The defined formats are parameterized over width and precision in bits, signedness, and the presence of infinities. Operations are defined by decoding floating-point values to the set of closed extended reals: the reals augmented with positive and negative infinity and NaN (Not a Number). Explicit treatment of NaN and infinite operands ensures that only real arithmetic is invoked in operation definitions. Extensive rounding and saturation modes are defined; stochastic rounding is included. Operations are exception-free, accelerating throughput, with exceptional situations communicated through return values, e.g., NaN. Operations on blocks of values sharing a common scale factor are defined in terms of the underlying operations in a uniform manner. System vendors may describe approximate implementations via a novel scale-invariant measure, akin to units in the last place. Standard function definitions and various other properties are mechanically verified and generated using formal specifications.

Proteolysis-targeting chimeras (PROTACs) can selectively degrade disease-causing proteins, yet predicting which targets are amenable to degradation remains a critical bottleneck: existing computational methods require the complete PROTAC molecular structure, information unavailable before synthesis. We present DegradoMap, a graph neural network that predicts PROTAC-mediated degradability from protein structure and E3 ligase identity alone – the minimal information available at the target selection stage. The model encodes biophysical priors through lysine-weighted graph pooling with per-protein normalization, models protein-E3 compatibility via cross-attention, and integrates cellular context from the Cancer Dependency Map. On the PROTAC-8K benchmark (3,101 samples, 155 targets, 10 E3 ligases), DegradoMap achieves 0.646+-0.124 AUROC on target-unseen evaluation (best seed: 0.7449) and 0.811 AUROC on CRBN->VHL E3-unseen transfer, outperforming GNN and machine learning baselines. The model additionally recommends optimal E3 ligases with 74% Hit@3 accuracy. Two findings carry broader implications: E(3)-equivariant architectures underperform the simpler invariant design for this scalar prediction task, and ESM-2 embeddings improve peak performance only with careful regularization – naive integration fails. DegradoMap provides pre-synthesis computational guidance for degradability assessment; its well-calibrated confidence scores (ECE = 0.029, target-unseen) enable practitioners to prioritize high-confidence predictions for experimental follow-up. However, the high seed variance (std = 0.124) and limited E3 coverage require ensembling for reliable deployment.

Splice-mediated drug resistance occurs in up to 40% of patients on targeted kinase inhibitors, yet state-of-the-art druggability tools operate on single structures and cannot compare across isoforms. We introduce SpliceBind, a graph neural network framework for isoform-aware druggability prediction. Beyond improving prediction accuracy (AUROC 0.703 vs. P2Rank 0.634, p = 0.026), we address a more fundamental question: when do structural methods succeed, and when must they fail? Systematic analysis of six clinically validated variants spanning five mechanism classes reveals a two-tier resistance taxonomy. Domain deletions (AR-V7, Delta = -18.39) and pocket disruptions produce structurally detectable changes, while allosteric mechanisms (BRAF-p61) remain fundamentally invisible to any pocket-centric approach – a boundary no algorithmic improvement can cross. Notably, learned embeddings capture affinity-based resistance missed by geometry alone (ALK-L1196M: Delta_SB = -0.228 vs. Delta_P2Rank = -0.95), partially bridging the structural-biochemical gap. On 229 kinase pockets spanning 25 families, SpliceBind achieves AUROC 0.703 (p = 0.026 vs. P2Rank) with robust generalization to held-out families (AUROC 0.761). This taxonomy transforms clinical workflows: upon discovering a splice variant, clinicians can immediately determine whether computational triage suffices or biochemical validation is required – reducing time from variant discovery to therapeutic decision.

Large Language Model (LLM) inference relies on key-value (KV) caches to avoid redundant attention computation. While approximate KV cache retention techniques reduce memory usage by sacrificing model accuracy, lossless approaches instead evict KV cache blocks from GPU memory and reconstruct them on demand to preserve exact outputs. Existing lossless KV cache management systems primarily base eviction decisions on access frequency or positional heuristics, without considering how different KV cache blocks affect the execution efficiency of GPU attention kernels. In this paper, we propose AsymCache, a computation-latency-aware KV cache management system for LLM inference that explicitly aligns cache residency decisions with GPU attention kernel performance, including three key components: Multi-Segment Attention (MSA) for efficient non-contiguous KV context processing, a cache eviction policy that jointly optimizes hit rate and position-aware recomputation cost, and an adaptive chunking scheduler for high hardware utilization. Experiments show that AsymCache reduces TTFT by up to 1.90-2.03x and time-per-output-token (TPOT) by 1.62-1.71x over latest baselines, confirming the effectiveness of the method in common workloads and validating its design goal of balancing computational efficiency with cache hit rate. Moreover, the low-level design of AsymCache allows seamless integration into agent serving systems such as Continuum, where it further reduces average job latency by up to 18.1%.

Production inference increasingly targets a heterogeneous mix of accelerators. Agentic pipelines interleave reasoning, tool calls, and multi-agent coordination, each with distinct compute and memory profiles. For optimal efficiency, each stage should run on the accelerator best suited to it. This creates a systems challenge: each pipeline now requires high-performance kernels across a growing set of hardware backends and programming models. Writing these kernels by hand is time-consuming, demands deep low-level expertise, and does not scale as kernel complexity grows. Recently, Large Language Models (LLMs) have been leveraged for automatic kernel generation, but challenges in low-level code generation and cross-backend generalization persist. We present KForge, a cross-platform framework built around an iterative refinement loop driven by two collaborating LLM-based agents: a generation agent that produces and progressively refines kernels using compilation and correctness feedback, and a performance-analysis agent that interprets profiling data, from programmatic APIs to GUI-based tools, and emits recommendations that steer the next round of synthesis. The loop alternates between functional passes, which drive a candidate to correctness, and optimization passes, which close the performance gap to hand-tuned baselines. We evaluate KForge on two backends with very different baseline reference availability. On NVIDIA B200, KForge achieves a 2.12$\%$ improvement in end-to-end throughput compared to TensorRT-LLM on the gpt-oss-20b inference speed benchmark. On Intel Arc B580, KForge generates Triton kernels achieving a 5.13$\times$ geometric mean speedup over the faster of PyTorch eager and torch.compile on 37 GEMM + tail-ops workloads from KernelBench Level 2, primarily via operator fusion and mixed-precision execution.

Published evaluations of prompt-injection and jailbreak detectors for Large Language Models often suffer from two systematic weaknesses: per-dataset threshold tuning and undisclosed operating points. We describe an evaluation harness that addresses both. The detector under evaluation is scored across 16 public benchmarks (12,111 samples) using 5-fold cross-validation. StratifiedKFold (by row) is the headline pass; a parallel StratifiedGroupKFold pass over a composite key (parent-prompt id plus MinHash + LSH near-duplicate clusters at Jaccard $\gtrsim 0.8$) runs alongside it as a leakage-premium diagnostic. A single global operating point is selected on the held-out folds (max F1 subject to FPR $\leq 1\%$) and applied uniformly to every dataset, so per-dataset results reflect one threshold rather than per-benchmark optimisation. Generalisation is examined through a battery of diagnostics (leave-one-dataset-out cross-validation, a random-label control, adversarial validation, permutation feature importance, length-bias correlation, classifier-head agreement, cross-source near-duplicate detection, threshold transferability, train-vs-OOF agreement, and a paraphrase-invariance probe), most with a quantitative pass threshold and the remainder with a stated failure mode. For every external comparison, the detector's threshold is re-tuned to the competitor's published false-positive rate so head-to-head values are evaluated at matched operating points.

Existing autonomous driving datasets have enabled major progress, but fall short in sensor fidelity, map completeness, or geographic diversity. We present KITScenes Multimodal, a European dataset built around high-fidelity sensors and maps. Our fully synchronized sensor suite combines high-resolution global-shutter cameras, long-range lidar beyond 400m, 4D imaging radar, and redundant GNSS/INS localization. Our HD maps are, to our knowledge, the most complete of any sensor dataset, validated through autonomous driving trials on open-source software. For the first time in a public dataset, all driving-relevant traffic elements, such as traffic lights, are mapped in 3D to a reprojection-accurate level with full topological connectivity. Recorded in cities with irregular street layouts and mixed traffic modes, our dataset complements existing datasets by broadening the available geographic diversity. We also introduce four benchmarks, each advancing spatial learning for embodied AI: online HD map construction, long-range depth estimation, novel view synthesis, and end-to-end driving. Project page: https://kitscenes.com/

Diffusion large language models promise parallel token generation, yet inference remains bottlenecked by deciding which masked tokens can be safely committed together. Fast-dLLM addressed this with KV caching and confidence-guided parallel decoding, but its decoding theory uses a homogeneous high-confidence assumption that effectively reduces each candidate set to its weakest selected token. We argue that this leaves speed on the table because real decoding steps exhibit heterogeneous confidence profiles. We propose Fast-dLLM++, a training-free extension that introduces Fréchet profile decoding: selecting parallel commit sets from the full sorted confidence profile rather than a single worst-case confidence. The resulting rule is a heterogeneous-confidence generalization of Fast-dLLM's factor selector and it recovers the previous rule exactly in the equal-confidence case and adds a provable heterogeneity bonus when the selected tokens have uneven confidences. Fast-dLLM++ leaves the model, diffusion process, and cache implementation entirely unchanged, making it a drop-in replacement for existing Fast-dLLM decoding. Experiments on GSM8K, MATH, HumanEval, and MBPP with the LLaDA-8B model show that the theoretical improvement translates directly into empirical gains: profile-aware selection improves the accuracy–throughput frontier by exploiting safe parallelism that weakest-token rules miss, achieving up to 37% higher throughput at comparable accuracy. Our code release is at https://github.com/Ringo-Star/FastdLLM_plusplus.

Curvature adaptivity is a classical theme in online optimization: for convex Lipschitz losses, adaptive methods interpolate between the optimal $O(\sqrt{T})$ regret for general convex losses and $O(\log T)$ regret under strong convexity. Recent work has shown that Follow-the-Perturbed-Leader (FTPL) achieves optimal $O(\sqrt{T})$ regret even for online non-convex Lipschitz losses, assuming access to an approximate offline-optimization oracle, but these guarantees do not exploit curvature. We show that FTPL can be made curvature-adaptive in the non-convex setting, without knowing in advance how curvature will accumulate over time. Our algorithm replaces the fixed perturbation scale of standard FTPL with a time-varying scale chosen using only past information. We give a simple follow-the-leader tuning rule for this scale and show that it competes, up to constants, with the best choice in hindsight. The resulting method achieves $O(\sqrt{T})$ regret for arbitrary non-convex Lipschitz losses and improves as cumulative curvature grows; with sufficiently accurate oracle calls, it achieves $O(\log T)$ regret when cumulative curvature grows linearly, which includes the classical strongly convex regime. We complement these upper bounds with matching lower bounds for prescribed cumulative-curvature sequences, already for one-dimensional convex losses, showing that the tradeoff between worst-case non-convex regret and curvature-driven fast rates is intrinsic.

Supervised fine-tuning is the predominant approach for adapting autoregressive vision-language models to downstream tasks. Recent work has shown that this paradigm is highly vulnerable to backdoor attacks, and that existing defenses are ineffective in open-ended generation settings. In response, we propose BYORn, a backdoor-robust fine-tuning framework motivated by the observation that poisoned target responses are often semantically implausible given the corresponding image-text inputs and a pretrained model. BYORn identifies such misaligned responses and dynamically replaces them with alternative responses generated by the model, thereby breaking the correlation between triggers and target outputs. The resulting objective gradient corresponds to the gradient of the empirical estimate of the population risk upper bound over the clean data distribution. Empirically, BYORn consistently improves robustness to backdoor attacks while preserving clean-task performance, establishing a new trade-off frontier between generalization and attack success rate. Finally, we demonstrate that BYORn remains effective against adaptive attacks specifically designed to circumvent the proposed defense.