Papers

Trajectory-based data attribution methods estimate the influence of training samples on model predictions by unrolling the training trajectory. They are widely used in applications such as data selection, data valuation, and model diagnosis, but there is a lack of comprehensive error analysis of these methods, raising concerns about method faithfulness and hindering reliable deployment. In this work, we provide the first systematic analysis of error sources in trajectory-based data attribution, together with concrete remedies to mitigate them and practical guidelines for downstream use. We organize the total error into three categories, config-level, algorithm-level, and system-level. We make three contributions. First, we identify optimizer mismatch as the dominant config-level error: existing methods derive their attribution under the assumption of SGD, even for models trained with the modern de facto optimizer AdamW. We propose AdamW-influence to fully account for AdamW's optimization dynamics, yielding improvements from 10% to over 300% in Spearman correlation between estimated and ground-truth influence across four settings spanning MLP, CNN, GPT-2, and Llama 3.2-1B. Second, we isolate the remaining algorithm-level error arising from the first-order Taylor approximation, identify the learning rate and trajectory length as factors governing the error magnitude, and derive a closed-form error proxy that can be evaluated along the original trajectory without retraining. Third, we translate these insights into practical guidelines for data selection by unifying offline and online strategies under a K-step look-ahead framework. Under this framework, online selection with a short horizon often matches or exceeds offline, and the optimal horizon can be tuned jointly with the learning rate. Together, these results turn the framework into an actionable selection recipe for practitioners.

World models aim to predict plausible futures consistent with past observations, a capability central to planning and decision-making in reinforcement learning. Yet, existing architectures face a fundamental memory trade-off: transformers preserve local detail but are bottlenecked by quadratic attention, while recurrent and state-space models scale more efficiently but compress history at the cost of fidelity. To overcome this trade-off, we suggest decoupling future-past consistency from any single architecture and instead leveraging a set of specialized experts. We introduce a diffusion-based framework that integrates heterogeneous memory models through a contrastive product-of-experts formulation. Our approach instantiates three complementary roles: a short-term memory expert that captures fine local dynamics, a long-term memory expert that stores episodic history in external diffusion weights via lightweight test-time finetuning, and a spatial long-term memory expert that enforces geometric and spatial coherence. This compositional design avoids mode collapse and scales to long contexts without incurring a quadratic cost. Across simulated and real-world benchmarks, our method improves temporal consistency, recall of past observations, and navigation performance, establishing a novel paradigm for building and operating memory-augmented diffusion world models.

Modern NLP and LLM systems are pipelines: named entity recognition (NER) -> entity disambiguation (NED) -> entity typing, retrieval-augmented generation (retriever -> reader), and agentic chains of planner -> tool -> critic. Errors compound across stages, but existing uncertainty quantification methods either calibrate each stage independently (no joint coverage) or apply a Bonferroni union bound (joint coverage, but conservative). We present PASC (Pipeline-Aware Split Conformal), which reduces multi-stage joint coverage to a single scalar conformal prediction problem on the joint maximum nonconformity score. PASC provides a finite-sample distribution-free guarantee that all K stages are simultaneously covered with probability at least 1 - alpha, and is nearly tight up to a 1/(n+1) factor. On a three-stage NER -> NED -> entity-typing pipeline over CoNLL-2003, PASC achieves 96.4% end-to-end coverage versus 93.4% for Bonferroni and 86.5% for independent CP, at identical average prediction set size (1.083). Under distribution shift to WNUT-17 Twitter and WikiNEuRal Wikipedia data, PASC empirically maintains the target coverage in the tested shift settings while independent CP collapses to 59%. PASC requires a single quantile computation, runs 1.7x faster than Bonferroni, and scales to K = 6 stages where independent CP drops to 0.53 end-to-end coverage. The same joint-maximum-score reduction applies directly to compound LLM systems and agent pipelines.

Speculative decoding accelerates LLM inference by having a small drafter propose tokens that a larger target model verifies in parallel. Recent diffusion-based parallel drafters such as DFlash predict the full B-token block in one forward pass, enabling deeper drafters and longer accepted blocks. However, existing multi-token drafter objectives often use fixed position-dependent weighting schedules, such as head-dependent weights or block-position decays, which do not adapt as the positions limiting acceptance change during training. To address this, we derive per-position training weights from a differentiable surrogate of expected accepted draft length, matching the weight of each position to its log-probability gradient contribution. The resulting loss, D-PACE (Dynamic Position-Aware Cross-Entropy), shifts training signal toward positions that currently limit acceptance as the drafter improves. Across six benchmarks, two Qwen3-4B draft depths, two decoding temperatures, and two additional target models, D-PACE consistently improves both wall-clock speedup and average emitted length, with 2.3% measured training-time overhead and no changes to the drafter architecture or inference procedure.

General-sum multi-agent learning is often governed by a stacked update field in which each agent's policy update changes the optimization landscape faced by the others. This coupling can entangle an integrable component of collective improvement with cyclic interaction dynamics, leading to slow or unstable multi-agent learning. Existing approaches, such as regularization, credit assignment, and consensus methods, stabilize MARL through local or algorithmic modifications; HPML complements them by projecting the joint update field onto a metric-gradient component. We introduce HPML (Hodge-Projected Multi-agent Learning), which views the joint update field of a multi-agent system as an element of an $L^2$ space of vector fields and computes a Hodge-type projection onto the closest metric-gradient potential flow. HPML follows the projected component as the update direction, yielding the closest metric-gradient field under the chosen metric and sampling measure. The projection is defined variationally, characterized by a Poisson-type equation, and implemented through graph-based and amortized neural realizations that recover projected directions from samples. We show that the projected dynamics admit a Lyapunov potential and yield equilibrium-gap bounds with an explicit additive non-potentiality term. Controlled experiments validate the geometric mechanism, and CTDE benchmarks show improved stability and normalized return when HPML is used as a plug-in projection layer in MARL pipelines.

While high-capacity AI models have advanced state-of-the-art performance, their practical deployment is often hindered by high inference costs, environmental impact, and a "one-size-fits-all" approach that ignores varying sample complexity. In clinical settings for instance, the waste of computational resources on routine cases is a significant barrier to sustainable AI. In this paper, we introduce the Conformal Adaptive Decision System (CADS), a sequential multi-model algorithm designed to optimize resource allocation by efficiently sampling models based on the estimated data complexity. CADS leverages conformal prediction to quantify image uncertainty at runtime. CADS provides a mathematically grounded framework for balancing the cost-accuracy dilemma that dynamically routes samples through a model cascade, ranging from lightweight "Scout" models to high-capacity "Oracle" architectures. Validated on two datasets, CADS demonstrated superior efficiency and accuracy at a computational cost that can be up to 12 times lower than heavy-model inference. By accurately routing samples based on real-time complexity, CADS ensures high diagnostic reliability while drastically reducing the economic and environmental footprint of AI.

The inversion of diffusion models plays a central role in image editing. Algebraically reversible ODE solvers provide an appealing approach to diffusion inversion for text-guided image editing, by eliminating the inversion error inherent in DDIM-based editing pipelines. However, empirical results indicate that reversibility alone is insufficient. As edits require larger semantic or visual changes, reversible diffusion solvers often exhibit instabilities and suffer sharp drops in output quality. In this paper, we show that the trade-off between exact reversibility and numerical stability manifests empirically as a trade-off between background preservation and prompt alignment in image editing. We then investigate the use of near-reversible Runge-Kutta methods as a more stable alternative to exactly reversible diffusion schemes. When combined with a vector-field smoothing strategy, the resulting approach improves edit fidelity, remains stable under large edits, and largely retains the background-preservation benefits of reversible solvers.

Plug-and-Play (PnP) methods for imaging inverse problems replace the proximal operator of the prior, a maximum a posteriori (MAP) denoiser, with a pretrained minimum mean squared error (MMSE) denoiser. MMSE estimates are oversmoothed, motivating recent work to target the MAP with diffusion-model scores. We show that MAP-targeting iterations converge to cartoon-like images with learned scores, but to realistic ones with the exact scores of a Gaussian mixture. The failure thus seems to come from score errors, and it only appears near convergence. We therefore propose ProxiMAP, an iterative MAP approximation that queries the denoiser at the residual noise level of the current iterate. This keeps the denoiser in-distribution and slows the iterations down before they reach the cartoon regime. On deblurring, super-resolution, inpainting, HDR and phase retrieval, replacing the MMSE denoiser with ProxiMAP in DPIR, DiffPIR and DAPS improves perceptual quality in most experiments, sometimes at the expense of PSNR. To make ProxiMAP cheaper, we also propose a one-step distillation and a hybrid schedule that applies it only at the last PnP iteration. In most cases, one or the other retains much of the improvement at a fraction of the cost.

We present OrbiSim, a novel robotic simulation paradigm that redefines world models as a fully differentiable physics engine for embodied intelligence. Unlike prior world models that focus on unconstrained imagination in latent or visual domains, OrbiSim establishes a unified, physically-grounded pathway that bridges structured scene assets, neural dynamics, and downstream reinforcement learning. By enabling end-to-end differentiability throughout the entire simulation loop – spanning from explicit state transitions to visual observation generation – OrbiSim supports tasks traditionally intractable for classical simulators, such as differentiable contact modeling, gradient-based policy optimization under sparse rewards, and intuitive physical inference. Empirical results demonstrate that OrbiSim significantly outperforms state-of-the-art world models in both predictive fidelity and control performance. Furthermore, its consistent responsiveness to asset configurations and physical parameters suggests its potential as a differentiable tool for enhancing robot simulation and policy training.

Multiple instance learning (MIL) is the dominant framework for whole-slide image analysis in computational pathology, typically combining a frozen patch encoder, a projection layer, and a slide-level aggregator. While encoders and aggregators have been extensively studied, the projection layer remains a largely morphology-only bottleneck. This limits endpoints such as biomarker status and survival, which are governed by a molecular state that is not fully captured by H&E morphology. We introduce Molecularly Informed Staining Transform (MIST), a plug-in replacement for the MIL projection layer that uses paired spatial transcriptomics only during training to construct virtual molecular stains. MIST clusters gene expression profiles into cross-modal prototypes, anchors them in the frozen foundation model feature space, and uses them to reorganize H&E patch features along molecularly guided axes. It requires no transcriptomics at inference and can be inserted before standard MIL aggregators. We evaluate MIST across 23 downstream tasks and 8 MIL aggregators. MIST improves 240 of 256 configurations over the standard projection layer, with an average gain of +3.5%, observed consistently across endpoint types: +5.2% on survival prediction, +3.3% on tissue subtyping, and +2.6% on biomarker prediction. Ablations confirm that gene-derived prototypes are the primary source of the gains, while spatial, biological, and pathological analyses show that cross-modal prototype affinities capture spatially coherent molecular programs from H&E alone.

Inertial measurement units (IMUs) are fundamental sensing components in multi-source integrated navigation systems, and their performance directly determines the accuracy and reliability of solutions. However, the precision of low-cost IMUs is inherently constrained by hardware limitations. Recently, generative artificial intelligence has demonstrated remarkable capability in modeling complex data distributions and reconstructing high-fidelity signals. Motivated by this, we propose a diffusion-based generative learning framework for synthesizing high-fidelity virtual IMU data from low-cost IMU measurements. Specifically, a conditional diffusion model based on a U-Net architecture is constructed, where high-grade IMU measurements are utilized as ground-truth priors and low-cost IMU measurements are employed as conditional inputs. The virtual IMU data generated by the model is used for subsequent navigation and localization tasks. Experimental results demonstrate that the generated virtual IMU data significantly outperform the original low-cost IMU measurements in both positioning and attitude estimation. Furthermore, we transfer the model to airborne mapping experiments, where the proposed method produces thinner and more consistent point clouds. Overall, the proposed framework breaks the performance limits of low-cost IMU and demonstrates the potential of diffusion-based generative learning for virtual high-grade IMU data.

Instruction-tuned language models exhibit behavioural fairness in high-stakes decisions while retaining biased associations in their internal representations. However, whether these suppressed representations can affect model outputs - and whether such causal potency is symmetric across demographic groups - remains unknown. We investigate the use of open-weight models for mortgage underwriting using matched applications that differ only in racially-associated names and reveal a critical disconnect: models show no output-level bias, yet retain and amplify demographic representations across model layers. Through activation steering and novel cross-layer interventions, we demonstrate that this suppressed information is decision-relevant: when reinjected at critical layers, it produces near-complete decision reversals. Critically, this latent bias is asymmetric - steering interventions affect decisions in one demographic direction, while producing minimal effects in reverse - and susceptible to adversarial prompt engineering and parameter-efficient fine-tuning. These findings demonstrate that behavioural audits focused on outputs are insufficient: fair outputs can mask exploitable internal biases. They also motivate dual-layer testing frameworks combining output evaluation with representational analysis for AI governance in high-stakes decisions.

Always-on AI applications, from environmental sensors to biomedical implants, require ultra-low power consumption. Analog circuits offer a path to sub-microwatt inference, yet existing analog implementations are limited to feedforward architectures: extending them to recurrent dynamics has been considered impractical due to noise accumulation through temporal feedback. We demonstrate that this barrier can be overcome through hardware-software co-design. Specifically, we identify that Bistable Memory Recurrent Units (BMRUs), a class of Recurrent Neural Networks (RNNs) with discrete-valued outputs and hysteretic dynamics, admit an ultra-low power current-mode analog implementation which we design from first principles. The resulting circuit establishes a one-to-one correspondence between each learned parameter and a circuit element. The discrete outputs suppress analog noise by at least 20-fold at each cell boundary, breaking the noise accumulation that prevents analog recurrence. We reformulate BMRUs for first-quadrant operation with fixed thresholds, enabling the direct correspondence while preserving expressivity and trainability. Transistor-level simulations in 180 nm Complementary Metal-Oxide-Semiconductor (CMOS) show near-perfect agreement between software predictions and circuit-level behavior, with the software model thereby serving as a high-fidelity simulator of the physical hardware at low computational cost. We leverage this fidelity to conduct large-scale noise immunity and power scaling analyses: the power cost of adding recurrence scales linearly with state dimension, while the feedforward layers dominating total power scale quadratically, meaning recurrence is added at linear marginal cost relative to the feedforward backbone. End-to-end keyword spotting achieves sub-microwatt inference at the RNN core.

In large-scale multi-agent systems with shared resource constraints, an upstream planner must iteratively evaluate candidate resource plans – assessing feasibility, aggregate response, and marginal cost – before committing to one. Lagrangian relaxation separates local decisions through a broadcast cost signal, but the planner still needs the cost-to-utilization response map to explore plan space, and this map depends on population composition that changes across planning cycles. We propose population-aware coordination interfaces: learned primal and dual maps, conditioned on compact population summaries, that the planner queries inside its iterative loop. The primal map predicts aggregate utilization under a proposed cost trajectory; the dual map predicts the cost trajectory for a target plan. By encoding response-relevant population structure, these maps remain reliable across evolving populations without per-cycle retraining, and support coordination of large populations from compact subsamples. We additionally cast Sim2Real transfer as a backtestable procedure, enabling evaluation before deployment. In a supply-chain capacity-control case study, population-aware interfaces reduce forecast error by 16–19% and capacity violations by 20–51% relative to population-unaware baselines under composition shift; 20K-agent cohorts support accurate coordination of 500K-agent populations; and simulator-trained primal maps achieve 11.1% MAPE on real observations versus 13–24% for baselines.

We propose a novel multimodal deep learning framework for patient-level survival prediction, which integrates whole-slide histology features, RNA-seq expression profiles, and clinical variables. Our architecture combines an ABMIL module~\cite{ilse2018attention} for slide-level representation with feedforward encoders for RNA and clinical data. These embeddings are then integrated through low-rank bilinear cross-modal fusion~\cite{liu2018efficient} to model conditional interactions across modalities while controlling parameter growth. The model outputs continuous risk scores that are subsequently mapped to survival times using a nonparametric calibration procedure based on the Kaplan–Meier estimator~\cite{kaplan1958nonparametric}. By decomposing multimodal reasoning into independent pairwise interactions, the proposed fusion design promotes structural interpretability and parameter efficiency compared with full tensor and hierarchical fusion strategies. Experiments on the CHIMERA challenge dataset demonstrate improved predictive performance over concatenation-based baselines and competitive generalization on hidden evaluation cohorts. These results indicate that the proposed framework is a promising approach for multimodal survival prediction in HR-NMIBC. The implementation is publicly available at https://github.com/hassancpu/ChimeraChallenge2025_Task_3.

In this work, we introduce a Tropical Axial Attention neural reasoning architecture that replaces vanilla softmax dot-product attention with max-plus operators, inducing a piecewise-linear structure aligned with dynamic programming formulations. From multi-species sequence alignments, our model learns all possible pairwise distances and is trained using a combination of $\ell_1$ and tropical symmetric distance metric losses with an ultrametric violation penalty. We leverage the well known isomorphic relationship between the space of all phylogenetic trees with $n$ species and tropical Grassmannian to show that tropical attention provides a natural geometric framework for phylogenetic inference. On empirical $DS1-DS11$ alignments, where true trees are unknown, the tropical model achieves the lowest MAE to its FastME-induced tree metric on every dataset, with a MAE reductions averaging 81.5% relative to Phyloformer and 98.4% relative than Phyloformer 2. These results suggest that tropical attention is a useful geometric inductive bias for neural phylogenetic inference, especially under distribution shift and when tree-metric consistency is important.

Stain variation across hospitals degrades histopathology models at deployment. Existing augmentation methods perturb color spaces with arbitrary hyperparameters, lacking both a principled budget and coverage guarantees for unseen centers. We propose Calibrated Adversarial Stain Augmentation (CASA), which performs adversarial augmentation in the Macenko stain parameter space with a budget calibrated from multi-center statistics via the DKW inequality. On Camelyon17-WILDS (5 seeds), CASA achieves $93.9\% \pm 1.6\%$ slide-level accuracy – outperforming HED-strong ($88.4\% \pm 7.3\%$), RandStainNA ($85.2\% \pm 6.7\%$), and ERM ($63.9\% \pm 11.3\%$) – with the highest worst-group accuracy ($84.9\% \pm 0.9\%$) among all 10 compared methods.

Single-arm trials are an important study design for evaluating drug efficacy and safety without enrolling patients into a control arm. Although they do not provide the gold-standard evidence of randomized controlled trials, they are increasingly used in clinical development as they offer an efficient, ethical, and practical alternative. A wide variety of approaches can be used to construct control comparators and estimate treatment effects, from fixed comparators informed by clinical knowledge to data-based and model-based patient-level comparators, also known as synthetic controls. Powerful and flexible machine learning models can allow outcome-model-based synthetic controls to overcome key limitations of direct data-based approaches, yield more robust estimates of treatment effects, and provide a principled way to incorporate corrections or encode additional assumptions when external data are not directly comparable. In this work, we argue that outcome-model-based synthetic control arms are an important tool for single-arm trials. We focus on digital twins, personalized predictions of disease progression generated from machine learning models trained on historical datasets, which naturally leverage these flexible approaches. We review doubly robust estimators, present power and sample size formulas, and discuss trade-offs in selecting historical data for training and analysis. We also outline practical considerations for deploying digital twins within the framework of recent FDA draft guidance on the use of artificial intelligence in drug development. Finally, we reanalyze data from trials in amyotrophic lateral sclerosis and Huntington's disease to demonstrate the proposed methods.

Inverse reinforcement learning (IRL), which infers reward functions from demonstrations, is a valuable tool for modeling and understanding decision-making behavior. Many variants of IRL have been developed to capture complexities of human decision-making, such as subjective beliefs, imperfect planning, and dynamic goals. However, an often-overlooked issue in real-world behavioral datasets is that the recorded data may be missing observations that were available to the original decision-maker. In use-inspired settings such as healthcare, this can make expert actions appear suboptimal, even when they were near-optimal given the information available at the time. As a result, the rewards learned by standard IRL may be misleading. In this paper, we identify the minimal perturbations to the recorded observations needed for the expert's actions to appear optimal. We develop a practical algorithm for this problem and demonstrate its utility for quantifying the possible extent of missing observations in behavioral datasets through extensive experiments on synthetic navigation tasks, a cancer treatment simulator, and ICU treatment data.

Graph neural networks (GNNs) deployed as cloud services can be *stolen* through *model-extraction attacks*, which train a surrogate from query responses to reproduce the target's behaviour, and a growing line of ownership defenses tries to prevent or trace such theft. The title of this paper asks two questions: *how hard is it to steal a GNN?*, and *can we stop it?* Prior work cannot answer either, because experiments use inconsistent datasets, threat models, and metrics. We introduce *GraphIP-Bench*, a unified benchmark which evaluates both sides under a single black-box protocol. It integrates twelve extraction attacks, twelve defenses spanning watermarking, output-perturbation, and query-pattern-detection families, ten public graphs covering homophilic, heterophilic, and large-scale regimes, three GNN backbones, and three graph-learning tasks, and it reports fidelity, task utility, ownership verification, and computational cost on shared splits, queries, and budgets. We further add a joint attack-and-defense track which runs every attack on every defended target and measures watermark verification on the resulting surrogate, which exposes the protection that a defense retains after extraction. The empirical picture is short: stealing a GNN is easy at medium query budgets and most defenses do not change this; several watermarks verify reliably on the protected model but lose most of their verification signal on the extracted surrogate, which exposes a gap that single-model evaluations miss; and heterophilic graphs are systematically harder to steal, while a cross-architecture mismatch between target and surrogate reduces but does not prevent extraction. We release *GraphIP-Bench* with reproducible scripts and configurations, and integrate the attacks and defenses into the PyGIP library.

We introduce Orthrus, a simple and efficient dual-architecture framework that unifies the exact generation fidelity of autoregressive Large Language Models (LLMs) with the high-speed parallel token generation of diffusion models. The sequential nature of standard autoregressive decoding represents a fundamental bottleneck for high-throughput inference. While diffusion language models attempt to break this barrier via parallel generation, they suffer from significant performance degradation, high training costs, and a lack of rigorous convergence guarantees. Orthrus resolves this dichotomy natively. Designed to seamlessly integrate into existing Transformers, the framework augments a frozen LLM with a lightweight, trainable module to create a parallel diffusion view alongside the standard autoregressive view. In this unified system, both views attend to the exact same high-fidelity Key-Value (KV) cache; the autoregressive head executes context pre-filling to construct accurate KV representations, while the diffusion head executes parallel generation. By employing an exact consensus mechanism between the two views, Orthrus guarantees lossless inference, delivering up to a 7.8x speedup with only an O(1) memory cache overhead and minimal parameter additions.

Coarse-grained (CG) molecular dynamics enables simulations of atomic systems such as biomolecules at timescales inaccessible to all-atom (AA) methods, but existing CG neural potentials trained via force matching capture only the gradient of the free-energy surface, leaving its curvature unsupervised. We introduce a framework that augments force matching with stochastic Hessian-vector product (HVP) matching, instilling second-order curvature information into CG potentials without constructing the full Hessian. We derive a decomposition of the target CG Hessian into a model-independent projected AA Hessian, precomputed once before training, and a model-dependent covariance correction computed online at negligible cost. We then construct an unbiased stochastic estimator of the Hessian-matching objective by using random probe vectors. We evaluate our method by training on a benchmark set of nine fast-folding proteins and gauging agreement of the learned free-energy landscapes with reference data along the directions of slowest collective motion using time-lagged independent component analysis. HVP matching improves slow-mode accuracy over force matching for all but one of our benchmark proteins. It also sharpens local structure, reducing bond-length error on eight of nine proteins, and in several cases by an order of magnitude. Our results demonstrate that higher-order physical supervision is a practical path to more accurate CG potentials for biomolecular simulation.

Longitudinal clinical notes contain rich evidence of how patients evolve over time, but converting this signal into training supervision for clinical prediction remains challenging. We extend Foresight Learning to clinical prediction by converting time-ordered MIMIC-III notes into examples consisting of past patient context, a natural-language question about a possible future event, and a label resolved from later documentation. This process yields 6,900 prediction examples from 702 admissions across medications, procedures, organ support, microbiology, and mortality. A small LoRA adapter trained on these examples improves over the prompted base model, reducing expected calibration error from 0.1269 to 0.0398 and Brier score from 0.199 to 0.145, while slightly outperforming GPT-5 point estimates on held-out questions. The approach enables reusable clinical prediction supervision from longitudinal notes without hand-engineered structured features or endpoint-specific classifiers.

The Average Gradient Outer Product (AGOP) governs feature learning in neural networks: the Neural Feature Ansatz states that weight Gram matrices at each layer align with the corresponding AGOP matrices computed over the training distribution. We ask a complementary question: can this same quantity serve as a post-hoc attribution method for explaining individual predictions? We introduce AGOP-Weighted: a novel attribution method that multiplies the per-sample gradient by sqrt(diag(M) / max diag(M)), a training-distribution prior that suppresses gradient noise and amplifies consistently important pixels – a combination not present in any prior attribution method. We formalise two companion variants – AGOP-Local (per-sample gradient, equivalent to VanillaGrad) and AGOP-Global (diag(M) directly as a zero-cost saliency map) – and implement an efficient training-time accumulation hook; AGOP-Global then requires zero inference cost (disk lookup) while AGOP-Weighted requires only a single gradient pass. We conduct the first rigorous comparison of AGOP attribution against Integrated Gradients (IG), SmoothGrad, GradCAM, and VanillaGrad across two benchmarks with pixel-level ground truth: (i) the synthetic XAI-TRIS benchmark (four classification scenarios, 8x8 images, CNN8by8) and (ii) the photorealistic CLEVR-XAI benchmark (ResNet-18 fine-tuned from ImageNet). AGOP-Weighted achieves 44% higher mIoU than IG on linear tasks; AGOP-Global achieves 7x higher mIoU than IG on multiplicative tasks (where IG falls below random) at zero inference cost. Both findings generalise to ResNet-18 on CLEVR-XAI (+18% and +37% respectively). We further show that GradCAM fails on small-resolution images due to spatial resolution collapse, and that diag(M) quality improves monotonically throughout training even after classification accuracy has plateaued.

Large language models (LLMs) achieve strong performance across many tasks but remain vulnerable to hallucinations, making it important to systematically evaluate their reliability under realistic adversarial inputs. We formulate hallucination elicitation as a constrained optimization problem, where the goal is to find semantically coherent adversarial prompts that are equivalent to benign user prompts. Existing attack methods remain limited: discrete prompt-based attacks preserve semantic equivalence and coherence but search only over a limited set of prompt variations, while continuous latent-space attacks explore a richer space but often decode into prompts that are no longer valid rephrasings. To address these limitations, we propose REALISTA, a realistic latent-space attack framework. REALISTA constructs an input-dependent dictionary of valid editing directions, each corresponding to a semantically equivalent and coherent rephrasing, and optimizes continuous combinations of these directions in latent space. This design combines the optimization flexibility of continuous attacks with the semantic realism of discrete rephrasing-based attacks. Experiments demonstrate that REALISTA achieves superior or comparable performance to state-of-the-art realistic attacks on open-source LLMs and, crucially, succeeds in attacking large reasoning models under free-form response settings, where prior realistic attacks fail. Code is available at https://github.com/Buyun-Liang/REALISTA.

A critical step for reliable large language models (LLMs) use in healthcare is to attribute predictions to their training data, akin to a medical case study. This requires token-level precision: pinpointing not just which training examples influence a decision, but which tokens within them are responsible. While influence functions offer a principled framework for this, prior work is restricted to autoregressive settings and relies on an implicit assumption of token independence, rendering their identified influences unreliable. We introduce a flexible framework that infers token-level influence through a latent mediation approach for general prediction tasks. Our method attaches sparse autoencoders to any layer of a pretrained LLM to learn a basis of approximately independent latent features. Unlike prior methods where influence decomposes additively across tokens, influence computed over latent features is inherently non-decomposable. To address this, we introduce a novel method using Jacobian-vector products. Token-level influence is obtained by propagating latent attributions back to the input space via token activation patterns. We scale our approach using efficient inverse-Hessian approximations. Experiments on medical benchmarks show our approach identifies sparse, interpretable sets of tokens that jointly influence predictions. Our framework enhances trust and enables model auditing, generalizing to high-stakes domain requiring transparent and accountable decisions.

Neuroscience data are fragmented across labs, experimental paradigms, and storage formats, and data reuse often requires substantial manual effort. Standardized formats have been proposed in response, but a fundamental tension persists: formats flexible enough for diverse experiments are rarely self-explanatory, while fully descriptive datasets demand documentation and curation effort that few labs can sustain. Coding agents, which can read code and text quickly with apparent attention to details humans tend to skim, are a natural candidate to reduce this burden. To measure how well coding agents perform data reuse, we benchmark them on eight recent papers with large-scale mouse neural recordings spanning diverse recording modalities, behavioral paradigms, and formats. Given the data, code, and paper, the agents must reformat the data for a common downstream task: training a decoder from neural activity to task and behavioral variables. Evaluating the resulting solutions, however, is challenging, since many conversion decisions are under-specified and several choices are defensible. Yet a subtle but scientifically significant error can invalidate the dataset in ways aggregate metrics can miss. We therefore pair outcome-based metrics with a process-based rubric, scored by two human evaluators against manual reference solutions. We find that agents succeed on most subtasks but rarely produce a fully error-free end-to-end solution; we characterize their mistakes and suggest data-sharing practices for agent-assisted reuse. Finally, we find that agent judges are unreliable at catching errors on their own, but can automate process-based evaluation when given reference solutions.

MeanFlow enables one-step generation in continuous spaces by learning an average velocity over a time interval rather than the instantaneous velocity field of flow matching. However, discrete state spaces lack smooth trajectories or spatial derivatives, so the continuous formulation does not directly apply. We introduce Discrete MeanFlow, a theoretical finite-state formulation that replaces point motion with probability-mass transport. Our key object is the conditional transition kernel of a continuous-time Markov chain (CTMC), from which we define a mean discrete rate that measures average transition-probability change over a finite interval. We prove a Discrete MeanFlow identity that relates this finite-interval rate to the instantaneous CTMC generator at the endpoint, using the Kolmogorov forward equation in place of the spatial chain rule in continuous MeanFlow. We parameterize the transition kernel directly using a boundary-by-construction design that guarantees valid probability outputs and exact boundary conditions without auxiliary losses. Because the learned kernel defines a probability distribution, generation requires only a single forward pass followed by direct sampling, eliminating the need for iterative denoising, ODE integration, or multi-step refinement. We validate the identity on exact finite-state Markov chains, demonstrating that the learned kernel recovers the analytical ground truth with high precision. We also investigate the training objective and the limits of factorization on synthetic sequence tasks across varying alphabet sizes and sequence lengths.

Detecting concept drift in high-speed data streams remains challenging, particularly when models must operate on unlabeled data and avoid false alarms caused by benign shifts. While disagreement-based uncertainty has shown promise in neural networks, its adaptation to ensembles of incremental decision trees (IDTs) remains largely unexplored. We investigate this approach by constructing batch-specific disagreement measures via label flipping in ensemble members and evaluating their effectiveness for drift detection in tabular data streams. Our experiments show that, although this method performs well in ensembles of multi-layer perceptrons (MLPs), it consistently underperforms loss-based detectors when applied to IDTs. We attribute this behavior to the intrinsic rigidity of IDTs: learning primarily through structural expansion, with limited parameter adaptation, restricts model plasticity and prevents disagreement from reliably reflecting learning potential. Recent work on restructuring IDTs using their intrinsic decomposition into non-overlapping rules offers a promising direction for improving adaptability.

Post-training a language model can induce surprising behavioral changes far beyond the intended training target. We study two particularly striking examples: emergent misalignment (EM), where fine-tuning on a narrow dataset leads to misaligned behavior in unrelated contexts, and subliminal learning (SL), where a student model inherits a teacher's misalignment through seemingly benign training data. We examine how the training data, evaluation data, and training method shape where and how strongly misalignment transfers between data and models. We characterize the data by the domain of a prompt (its topical context) and its task (the function the model performs), and we build natural-language and synthetic datasets organized as domain-task grids. We find that EM transfers more strongly to prompts that share the task of the fine-tuning data than to prompts that share its domain. In the synthetic setting, tasks that the model has learned well spread misalignment more broadly. For SL, we go beyond SFT (supervised fine-tuning) to off-policy distillation on the teacher's full next-token distribution and, to our knowledge for the first time, on-policy distillation, where the teacher only scores tokens sampled by the student. Misalignment transfers even under on-policy distillation, more strongly than under SFT and at a level comparable to off-policy distillation. Unlike SFT, where the teacher both generates the data and supervises the student, distillation lets us separate the two. With the teacher fixed, the data change how much misalignment the student inherits, but the teacher determines to which domains and tasks it transfers most strongly. Together, these results show how data and training influence the extent and patterns of emergent and subliminal misalignment.