Papers

Counterfactual fairness audits of clinical language-model agents report a flip rate: how often an action changes when only the patient's demographic descriptor changes. Part of that rate is not demographic. A stochastic agent also changes its own action when nothing changes, and a flip rate cannot be interpreted without knowing how often. We measured it. Re-running one condition ten times over sixteen synthetic vignettes at default sampling changed a clinical agent's action in 8.7 percent of replicate pairs, from 2.2 percent for intensive-care escalation to 17.9 percent for controlled-substance caution, an output given no operational criteria. Across six models from five vendors, pooled floors ranged from 2.5 to 23.7 percent; in this panel neither disclosed size, vendor, nor hosting ordered them. The floor depends on the decoding configuration: majority voting over five draws removed 39 percent of it (95 percent confidence interval (CI) 18 to 64); at temperature 0 three of four locally served models showed no disagreement, but a hosted model still did. We also show that, for a binary action, the flip rate expected under no demographic effect equals the floor and a real effect adds only its square, so a flip rate inside the floor is not evidence of fairness, and direction must be tested with a signed paired test. We give a four-step reporting procedure and release FairMedAgent, the harness, with its protocol, vignettes and analysis scripts, so any team can measure the floor for its own agent.

Writing proficiency manifests in how students develop content, organize ideas, choose words, and use language. Despite growing interest in LLM-based student simulation, whether LLMs can reproduce such multidimensional variation in extended writing remains largely unexplored. In this work, we explore if language models can realistically simulate student writing, and introduce SWIM, a task that formulates Student Writing sIMulation as proficiency-conditioned essay generation. We evaluate prompting, supervised fine-tuning (SFT), and reinforcement learning (RL) methods for writing simulation using automated essay scoring as a measure of profile alignment. Extensive experiments reveal that prompting provides limited proficiency control, even for strong proprietary LLMs with rubric-grounded strategies. In particular, while models can adjust content-oriented traits, they struggle to reproduce the lexical, grammatical, and organizational variation in different proficiency levels. SFT substantially improves alignment, while RL with the proposed proficiency-alignment reward yields further gains across all writing traits and essay prompts. Our findings suggest that explicit supervision enables substantially stronger profile alignment than prompting alone, while authentic low-proficiency writing remains challenging to reproduce.

Precipitation forecasts shape decision-making across the global economy, particularly in sectors such as agriculture. However, unlike variables such as temperature, precipitation is highly intermittent and localized, making it difficult to forecast. While recent advances in AI weather prediction systems have enabled them to surpass physical models on globally averaged metrics, improvements in mean error rarely translate to actionable forecasts of severe flooding or dry crop fields. Furthermore, most of these models are trained and evaluated against a reanalysis data product, ERA5, which has well-known biases. Here we retrain AIFS, ECMWF's widely-used, open-source operational 0.25{\deg} probabilistic graph-transformer weather model, on satellite-based precipitation observations. Our model, Laxmi, improves global probabilistic accuracy by 19% and resolves systematic distributional biases in ERA5. Specifically, Laxmi reduces drizzle overprediction by 33% for amounts less than 3 mm per day. It also mitigates extreme rainfall underprediction, improving the global 95th percentile Brier skill score by 57%. Across a case study of 10 Indian tropical storms, Laxmi delivered the most accurate forecast of 150 mm event-total precipitation in 7 events, compared to 1 for AIFS and 2 for the leading physical model, IFS. Our results demonstrate that incorporating observation-based precipitation data directly into training can substantially improve forecasts.

Plan accuracy alone cannot show whether a speech-delivery decision follows its source: a fixed prior may match the original label yet fail to respond appropriately when a cue changes. VoxReason provides a 100-case verifier benchmark that holds each utterance fixed, edits one designated source-label cue, and scores cited evidence, eight plan fields, and the permitted response. On a source-key-disjoint test of 24 cases, a source-emotion prior reaches plan-slot accuracy 0.958, but none of the 24 edited neutral targets appears in its training labels; its required-change accuracy is 0.000. This diagnoses the support boundary of this prior, not its performance on supported edits. In a complementary 32-case emotion-disjoint test, the prior has seen all edited neutral targets but neither original test emotion; its plan-slot accuracy is 0.219 and required-change accuracy is 1.000. The partitions reuse and overlap the same 100 cases, so these deterministic diagnostics are not independent cohorts or learned-planner results. The benchmark evaluates derived labels and structured plans, not audio input, generated speech, or listener judgments.

Long-term LLM agents must preserve information across interactions while distinguishing repeated evidence, historical states, updates, and unresolved contradictions. Existing textual memory systems retrieve semantically relevant memories efficiently but often leave these relationships implicit, whereas richer structured approaches model them through global graphs, hierarchical abstractions, or reflection at greater complexity. We introduce MemoryLACE (MemLACE), a lightweight memory framework that explicitly models the lifecycle of textual evidence through sparse merge, supersession, and contradiction relations while preserving atomic natural-language memories and their provenance. Rather than retrieving memories independently, MemLACE reconstructs relation-aware evidence units that expose current, historical, supporting, and conflicting evidence for downstream reasoning. Across BEAM and StructMemEval, using open-weight and proprietary LLM backbones, MemLACE achieves the highest overall performance in same-backbone comparisons while reducing end-to-end runtime on BEAM by 66.6% relative to Hindsight, the strongest reported reflective-memory baseline. Ablation studies identify lifecycle expansion and temporal awareness as the principal contributors to these gains. Together, the results demonstrate that explicitly modeling the local lifecycle of textual evidence is sufficient to substantially improve long-term memory reasoning without requiring comprehensive knowledge graphs or global reflection.

Nested sampling (NS) resolves the thermodynamics of an atomistic system from a single simulation, but its practical reach is limited by the Markov-chain updates needed to decorrelate walkers within each likelihood-constrained ensemble. Flow-based NS has reduced this bottleneck for gravitational-wave (GW) inference, yet its transfer to atomistic systems is not merely a change of application. Comparing a GW150914-like binary-black-hole likelihood with an eight-particle two-dimensional Lennard-Jones (LJ) system of comparable dimensionality, we show that the two landscapes differ fundamentally: atomistic multimodality is discrete and combinatorial, generated by particle permutations separated by hard collision walls, and its coordinate coupling is dense and collective, whereas the GW posterior exhibits smooth degeneracies and localized parameter coupling. Guided by this diagnosis, we introduce NS-Flows: a single conditional normalizing flow, conditioned on the NS energy bound and trained on a sliding window of recent live sets, that replaces MCMC by direct parallel draws corrected by importance-weighted rejection resampling. Live sets supply the data self-consistently, allowing flow training without structured priors or a pre-existing dataset. For LJ disks under periodic boundary conditions, the algorithm can reduce energy evaluations by two orders of magnitude and wall-clock time by about 30%, an advantage that becomes more favorable as the cost of the potential grows. The flow's generative efficiency further acts as a physical diagnostic: it varies non-monotonically along the annealing trajectory, is lowest in the dense disordered regime, and is quantitatively captured by the constrained ensemble's internal mode complexity together with target drift across the training window, identifying liquid-like ensembles, not prior-target separation, as the hard case for current flow architectures.

Many biomedical challenges can be posed as tensor completion problems where the observed entries of a multidimensional array (a tensor) are used to impute the missing values. In such settings, incorporating side information about the modes of the tensor, such as gene-gene similarity, can significantly enhance the solutions of the completion problem. Most existing tensor completion methods can only incorporate side information in the form of matrices. In this study, we introduce a novel framework to incorporate side information in the form of tensors. Our new approach, called Coupled Tensor-Tensor Completion (CTTC), leverages the hidden connections among multimodal tensors to improve tensor completion performance. In addition to practical utility, CTTC has theoretical foundations in distance metric learning and group theory. We derive an alternating algorithm to solve the CTTC optimization problem and establish its convergence to a stationary point. Finally, we show that CTTC outperforms state-of-the-art tensor completion methods at predicting drug effects. Results: Compared with other tensor completion methods, including HaLRTC, CTRC, Cell, and NTDDR, CTTC demonstrates superior run-time and RSE tensor completion accuracy on two benchmark datasets, DTD and LINCS.

When a statistical agency or regulator releases synthetic data in place of sensitive records, it chooses the generator that produces the table, and can shape that generator so unfair pathways are absent. DECAF made this concrete on one non-private GAN: three fairness definitions become three sets of edge cuts on the generator's causal graph. Whether the mechanism belongs to DECAF, or to causal factorisation itself, was untested. We port all three definitions to nine generators from three unrelated families (marginals-based, GAN, and diffusion, each with differentially private variants), across three levels of formal privacy guarantee, over 2,520 matched-pair runs on Adult and COMPAS datasets. The mechanism transfers everywhere, and our new causal diffusion backbone yields the fairest release of any family we tested, at fidelity close to the marginals tier. Applying the cut barely moves fidelity, only costs a downstream classifier about $0.07$ to $0.15$ AUC on average, and adding privacy guarantees don't make the data less fair.

Frontier large language models (LLMs) have become attractive priors for optimization due to their large-scale pretraining that enables them to navigate a variety of optimization settings. However, the effectiveness of modern reasoning LLMs in batch optimization settings remains underexplored. Here we investigate the performance of the current generation of frontier LLMs as batch optimizers in both continuous and discrete settings. We find that while LLMs are competitive zero-shot batch optimizers for numerical test functions, their performance is brittle compared to classical non-LLM optimization approaches. However, LLM priors are significantly better in semantically rich settings, indicating that their batch optimization behavior is highly effective when navigating and reasoning over the discrete spaces most similar in structure to their pretraining data.

Visual token pruning reduces the inference cost of vision-language models (VLMs), but most methods only ask which tokens to keep. This retained-token view can keep redundant high-scoring tokens while leaving discarded evidence without a close representative. We propose CoverPruner, a training-free pruner that asks the complementary demand-side question: after a token is removed, which surviving original token represents it for the target VLM? CoverPruner formulates pruning as Representational Coverage Maximization (RCM), covering the full projected visual-token set with query-weighted demand. It instantiates RCM with projector-space coverage and a lightweight first-layer attention probe. Across multiple VLM architectures and compression rates, CoverPruner achieves the best average accuracy among all compared methods, with the largest gains usually appearing under aggressive compression.

Long-context reasoning for large language models (LLMs) is becoming increasingly important, but training over long sequences remains challenging due to massive memory and communication requirements. Sequence parallelism has emerged as an essential technique for addressing bottlenecks in long sequence LLM training. However, we observe that existing sequence parallelism methods are batch-agnostic and apply uniform sequence partitioning across all batch sizes, resulting in inefficient communication. In this paper, we introduce Batch- Aware Sequence Parallelism (BASP), a sequence parallelism approach that leverages batch structure to reduce communication overhead. BASP exploits batch structure by partitioning GPUs into disjoint sequence-parallel groups according to the micro- batch size. This design reduces the all-to-all communication group size, thereby localizing communication and improving training efficiency. Experimental results on an NVIDIA A100 cluster show that BASP improves end-to-end training time by up to 1.17 - 1.31x in Llama and Qwen models compared to standard sequence parallel baselines, while preserving identical model accuracy and memory usage.

Multi-domain fine-tuning often combines MoE routing with LoRA, assuming that token-level routing separates domain-specific updates. We test this assumption in MoE+LoRA using Python code paired with biomedical text and mathematical reasoning. Although these domains show near-disjoint expert routing, adding biomedical data substantially increases code perplexity, indicating that routing separation alone may not prevent negative transfer. To localize the failure, we introduce Jaccard routing overlap and adapter-gradient cosine similarity, which measure expert sharing and update compatibility, respectively. These diagnostics indicate that interference arises mostly from nearly orthogonal domain gradients competing within the same low-rank adapter subspace. We address this issue with SpawnLoRA, which dynamically adds gated sub-adapters inside MoE experts when adapter-level contention is detected, while keeping the router fixed. We evaluate SpawnLoRA on Phi-tiny-MoE-instruct and OLMoE-1B-7B across multiple mixture settings and find that it effectively reduces negative transfer compared with standard and rank-adaptive LoRA. These results demonstrate that structural separation inside experts provides benefits beyond routing or rank expansion alone.

Analog in-memory computing (AIMC) speeds up neural-network inference by doing the arithmetic directly inside a memory array, instead of shuttling weights back and forth between memory and a processor. This saves energy, but the physical devices that store the weights are imperfect: programming errors, electrical noise, limited-resolution converters, and outright broken cells all distort the computation, and every physical chip is distorted in its own way. A designer with several such chips available faces an uncomfortable choice: run all of them and combine the answers (safe, but wasteful of energy), or trust a single chip blindly (cheap, but with no guarantee on how often it is wrong). This paper introduces RACE-AIMC (Risk-Aware Certified Ensemble for AIMC), a framework that resolves this choice with statistics rather than guesswork. Offline, RACE-AIMC studies a pool of physical accelerators, picks the single best one for a given energy budget, and computes a mathematically exact upper bound on how often that accelerator will be wrong when it chooses to answer. Online, only that one accelerator is switched on; a lightweight check decides whether to accept its answer or defer to a fallback. In our simulations using a noisy weight mapping and multiple independent test runs, every certified bound stayed under a 10% error target (mean bound 7.83% +- 0.89%, with 70.88% +- 0.98% of inputs answered directly). The resulting system matches the accuracy of a clean digital baseline while cutting modeled energy use by 69.02% relative to always running every accelerator in the pool.

Learning-based schedulers have been proposed to provide real-time user, target, and access point (AP) association in distributed cell-free integrated sensing and communication systems. In a typical approach, a graph neural network (GNN), trained on labels from a mixed-integer linear program, maps lightweight per-AP statistics to decisions on AP clustering, user and target scheduling, and mode selection in one forward pass. Such solutions assume that hard constraints, enforced only as soft training penalties, hold at inference, and that the self-reported statistics are truthful. Using our ASSENT algorithm as an example, we find that despite high $F_1$ scores, many solutions violate at least one hard constraint, demonstrating that high prediction accuracy does not ensure joint feasibility. Projecting the GNN output onto a feasible solution restores constraint satisfaction with low utility loss, even with a simple greedy repair procedure. We further show that feasibility alone does not guarantee robustness to false data injection attacks. A single malicious AP that reports false information cannot substantially increase its user associations, but can greatly increase the rate of infeasible solutions. The effect of such attacks depends on the type of information being falsified. Misreporting information that affects the objective can largely be mitigated through feasibility projection, whereas falsifying information that affects the constraints cannot. The latter can, however, be detected using a low-complexity cross-AP consistency check. These results show that learned ISAC schedulers should be evaluated using constraint-aware feasibility metrics in addition to conventional accuracy measures.

Vision-Language-Action (VLA) policies fuse multimodal sensory inputs, but training on limited and homogeneous robot demonstrations encourages spurious inter-sensor correlations rather than task-relevant signal, a failure we term modality entanglement. Under real-world occlusions and distractors, this manifests as nuisance sensitivity to corruption of uninformative sensors and single-modality insufficiency when only one informative sensor remains intact. We propose Evidence-Gated Regularization (EGR), a modality-agnostic training objective that introduces zero inference-time overhead. EGR derives a per-frame and per-sensor task-relevance signal to gate two state-conditional consistency objectives: invariance on low-evidence sensors, and single-sensor sufficiency on high-evidence ones. We introduce a benchmark based on BEHAVIOR-1K, comprising a fast inference-only diagnostic suite and 47 rollout-based skills targeting modality entanglement. We validate EGR on this benchmark and on two real-robot setups with fundamentally different embodiments: a bi-manual setup with two Kinova arms and three RGB cameras, and a single-arm MELFA ASSISTA setup combining vision and GelSight tactile sensors. EGR improves simulation success rates (SR) from 12.5% to 16.4% under full modalities (+31%), from 9.4% to 16.5% under uninformative-sensor corruption (+75%), and from 2.8% to 6.1% under single-sensor fallback (+120%). Under physical-object distractors, EGR boosts SR from 30% to 85% on the bi-manual setup (+183%) and from 55% to 70% on the tactile setup (+27%).

Several classical machine-learning methods, such as KRRs and SVRs, are both computationally and analytically tractable since their estimators either admit closed-form expressions or are obtained by minimizing convex training objectives; neither feature is generally available for deep neural networks. We address this by introducing a simple closed-form “two-stage” compositional formula $\hat{f}$ for reconstructing an unknown Lipschitz function $f:\mathcal{X}\to \mathbb{R}$ on a metric space $(\mathcal X,\rho)$ from $N$ i.i.d. noisy observations. Our main result is a high-probability uniform ($L^{\infty}$) recovery guarantee that jointly controls approximation and statistical errors while enjoying an optimization error of zero; in particular, we do not assume oracle access to an approximate ERM. Our secondary main results establish the optimality of our formula in three complementary senses. 1) Function space: On Ahlfors-regular metric spaces, the hypothesis class parameterized by our formula attains the optimal fat-shattering dimension. 2) Parameter space: Its dependence on the parameters is maximally numerically stable, in the sense that a smaller approximation error cannot be achieved with a smaller Lipschitz dependence on the model parameters. 3) Forward pass: Its dependence on the input is maximally regular, matching the Lipschitz constant of the target function $f$. When $\mathcal X=[0,1]^d$ is equipped with the $\ell^\infty$ norm, $\hat{f}$ admits algorithmic ReLU-MLP and exact ReLU-multi-head transformer realizations of depth $\mathcal{O}(\log(N))$ with $\mathcal{O}(N)$ nonzero parameters.

Neural fields (NFs) map continuous coordinates to signals such as color or density, but fast high-quality reconstruction from sparse observations remains difficult. Classical Neural Tangent Kernel (NTK) regression gives closed-form fits, yet it is fundamentally linear and cannot accumulate reusable task priors. We develop three algorithms that address these gaps. NTK-KIP learns a distilled support set of coordinates (and optional labels) so that a finite NTK can inpaint large missing regions from little observed data, yielding a compact non-linear representation instead of a raw kernel solve. MetaQuill meta-learns a shared initialization for an INR so that new scenes can be adapted by updating only a small task-specific weight offset, which provides true feature learning and a reusable prior. Finally, MetaQuill-KIP fuses both ideas: it seeds the task with a KIP-style non-linear warm start, then refines only that small offset around the meta-learned initialization. MetaQuill-KIP achieves high-PSNR reconstructions and semantically plausible inpainting under very sparse observations, while requiring only lightweight per-instance adaptation, whereas diffusion-style baselines typically depend on large pretrained generative priors and costly per-image tuning. This shows that NTK-driven neural fields can be made both non-linear and meta-learnable, narrowing the gap between analytic kernels and practical few-shot reconstruction.

Recent advances in generative speech models have made it increasingly difficult to distinguish authentic from synthetic audio, enabling new forms of fraud and misinformation. Audio watermarking offers a promising defense by embedding an imperceptible signal into generated speech that can later be detected to verify its provenance. However, recent studies have shown that existing post-hoc watermarking methods fail under neural codecs and denoisers, transformations routinely applied during real-world storage, transmission, and processing, severely limiting their practical utility. Here we introduce CRAW, a codec-robust audio watermarking framework that jointly improves robustness against neural re-synthesis while maintaining high perceptual quality. CRAW combines distortion-aware training with an attention-based pooling mechanism, inference-time perceptual mask- ing, and an error-correcting code to recover the fidelity lost during robust training. Experiments demonstrate that CRAW achieves state-of-the-art robustness against neural codecs, denoisers, and vocoders while maintaining perceptual quality comparable to existing post-hoc watermarking methods. The code is available at https://github.com/DavidC1212/craw.

Scaling laws in modern deep learning describe how held-out loss improves as model capacity, training data, and compute increase, often following power-law trends. We investigate whether analogous scaling regularities arise in actuarial ratemaking, where data are tabular, heterogeneous, and noisy, and where classical models such as GLMs remain strong baselines. Using a real-world motor insurance portfolio, we train models from different families across increasing fractions of the training data and multiple random seeds, evaluating out-of-sample Poisson deviance, a likelihood-based loss for Poisson count predictions in which lower values indicate better held-out fit. We find that all model families improve with additional data, but scaling exponents differ substantially: TabM exhibits markedly stronger data scaling than purely supervised tabular Transformers and standard MLP baselines. Transformer variants show weak parameter scaling unless augmented with additional inductive biases (TabM-style adaptation or self-supervision). These results provide quantitative guidance on model selection by data regime and suggest that effective scaling on actuarial tabular tasks depends on architecture and loss function objective design, with simple increases in Transformer size providing limited gains.

Conformal risk control is an emerging framework for the safe deployment of machine learning models with finite-sample guarantees. To accommodate a broader class of risk notions, quantile risk control extends this framework to quantile-based risk measures. However, existing methods either suffer from excessive conservatism or lack rigorous finite-sample guarantees. To address these limitations, we introduce Occupancy-based Quantile Risk Control (OQRC), a novel method that provides tight risk control bounds with finite-sample validity. Our key idea is to formulate risk control as a finite-occupancy problem by partitioning the loss space with the ordered calibration losses. Specifically, we estimate the distribution of test losses across the resulting bins and upper-bound the risk by the maximum loss attained within each bin. We then select the parameter $\lambda$ such that this upper bound does not exceed a predefined threshold $\alpha$ with high probability $1-\delta$. Theoretically, we establish a finite-sample guarantee showing that OQRC yields tight risk control bounds that converge to the optimal bounds at a provable rate of $\mathcal{O}_ p(n^{-1/2})$. Extensive experiments demonstrate the effectiveness of our method, reducing the risk gap by up to 78.64% on common benchmarks.

Geostationary atmospheric motion vectors (AMVs) provide the dense horizontal wind vectors (u,v) and heights ingested into data assimilation systems. Traditional AMVs track features using window-based cross-correlation and estimate heights via infrared brightness temperatures paired with numerical weather prediction (NWP) background states, creating a circular dependency that yields inaccurate heights, high computational cost, and sparse retrievals. Stereo winds from GEO-GEO and GEO-LEO geometrically resolve heights from parallax shifts across different poses, eliminating NWP dependence and improving accuracy, but they remain computationally heavy with limited coverage. In this work, we replace window-based tracking in stereo matching with deep optical flow for efficient, improved retrieval. Fine-tuning balances a self-supervised geometric residual loss with supervised radiosonde reconstruction. To eliminate multi-satellite overlap requirements, we distill the stereo teacher into a single-satellite student model. Chi-square and height uncertainties from the teacher are emulated by the student for quality assurance. The student generates winds across full-disk GEO imagery globally. Validation compares stereo and student models against radiosondes, operational AMVs, ERA5 reanalysis, and EarthCARE cloud profiles. Results through triple collocation show that stereo winds improve performance beyond operational AMVs for water vapor bands (6.2, 6.9, and 7.3 {\mu}m), wit degradation in the long-wave infrared (11.2 {\mu}m) band.

Smartphone skin photographs are indispensable to teledermatology, yet assessing the diagnostic suitability of submitted cases (gradability) remains a critical bottleneck in mobile care workflows. Dermatologists routinely review multiple photographic views (regional, angled, and close-up) to identify consistent textural detail rather than relying on a single image. We present the Semantic Tri-view Pipeline, an interpretable architecture for automated teledermatology gradability screening that formalizes epidermal micro-relief as a computable biomarker of image quality. Using an expert-annotated subset of the public SCIN dataset, we train a lightweight DeepLabV3+ model to segment micro-relief fidelity. These spatial masks are then aggregated across up to three case views with a logistic regression classifier, leveraging viewpoint redundancy to support robustness under uncontrolled smartphone acquisition. This approach learns context-aware, clinically intelligible heuristics, such as penalizing high-fidelity texture in regional distance views. Evaluated at a predefined 90% sensitivity operating point, the system's apparent errors largely reflect subjective clinical variance on borderline cases where clinicians rely on non-visual metadata. On SCIN, performance improves from an AUC of 0.81 (80.6% PPV) on variance-heavy majority-consensus cases to 0.96 (97.7% PPV) on optically unambiguous unanimous cases. Overall, this work delivers an interpretable, privacy-by-design, edge-ready system that can provide real-time feedback during case submission to filter ungradable photo sets before review.

Continuous chain-of-thought models compress reasoning into latent tokens. Matrix-valued variants, which route each latent token through a d x d matrix bottleneck, introduce rank as a single-sample structural observable on the latent matrix Z. If matrix latents carry parallel reasoning paths via superposition, rank should track them, and truncating Z to low rank should hurt accuracy on tasks whose solutions plausibly require multiple components. Across four training regimes of a matrix-CODI model (three on ProsQA, one on GSM8K-Aug below the learning threshold), the rank-k projection ablation curve is flat to within 0.6 percentage points. A three-seed replication yields 81.0 +/- 2.0 percentage points accuracy while the final effective rank of Z spans {4, 12, 13}; the loss does not reward any particular rank. To test whether rank-blindness arises from the flatten-then-project readout alone, we trained four readouts: a bilinear reparametrization, a bilinear-plus-GELU readout nonlinear in Z, an SVD-augmented readout feeding singular values through an MLP, and a quadratic readout in Z Z^T. All four rank-k curves remain flat (Spearman p-values 0.63, 0.14, 0.82, 0.46). The flat curves persist for readouts nonlinear in Z. A linear probe on Z underperforms a raw pretrained hidden state at target prediction (AUC 0.673 vs. 0.846). A negative control on vanilla GPT-2 SFT (no matrix bottleneck, no Z, three seeds, n=500) reproduces a flat rank-k curve under the same intervention paradigm with pooled-mean range 0.20pp, and a random-h sensitivity floor lands at the same accuracy: the rank-k ablation alone conflates rank-blindness with position-irrelevance.

On-device LLM inference is attractive for privacy and responsiveness, but remains challenging on mobile and embedded devices because model weights far exceed available DRAM. Prior systems exploit activation sparsity and offload weights to SSD or flash storage, but face a fundamental systems trade-off: accurate sparse execution decisions require the latest context, whereas efficient computation-I/O overlap requires early prediction. As a result, existing designs either serialize execution or incur redundant weight fetches, extra computation, and large cache overheads. We present LeanStream, a streaming speculate-and-refine framework for efficient on-device LLM inference. LeanStream progressively refines computation, loading, and cache-retention priorities using partial GPU results, enabling fine-grained overlap between GPU execution and storage I/O. We implement LeanStream on both mobile and embedded platforms. Compared with prior on-device LLM inference systems, LeanStream reduces memory usage by 4.8$\times$ to 7.5$\times$ at the best throughput achieved by prior work, while further improving token generation throughput by 1.6$\times$ to 2.1$\times$.

This position paper argues that the absence of labels does not imply the absence of human supervision in visual learning, and urges the research community to identify sources of supervision more explicitly. Many recent methods in computer vision build upon representations learned from large-scale unlabeled data, and are therefore grouped under the same umbrella term “unsupervised.” However, different data curation schemes and training objectives embed substantially different human priors on which models rely, and we argue that one “unsupervised” umbrella term is no longer capturing these distinctions. This ambiguity makes it harder to compare unsupervised learning research conducted under different assumptions, coinciding with a sharp decline in papers titled with “unsupervised” in flagship computer vision conferences since 2021, despite continued growth of the field. While we fully embrace pre-training as a strong foundation for modern computer vision, we advocate for a community-level effort toward greater conceptual clarity: authors are encouraged to disclose priors in data selection and learning objectives, and to specify which components of a learning pipeline depend on which assumptions. Standardized disclosure practices can improve academic communication, ensure fairer comparisons, and preserve methodological diversity in unsupervised learning.

Neural operators are fast, differentiable surrogates for physical simulation, but their accuracy often degrades when domain geometry, size, or operating conditions differ from training. Supervised adaptation can recover accuracy, but even a small target set requires costly high-fidelity simulations. We therefore ask how pretraining and transfer can be designed together to reduce this deployment cost. LatentDDM first pretrains a neural operator to predict fields on small subdomains. For a new setting, it freezes this operator and trains only a lightweight module that composes the local predictions. We evaluate our method on two complementary problems: steady Darcy flow, where long-range pressure coupling must extend across increasingly large porous domains, and unsteady incompressible flow around a pitching airfoil, where rollout errors compound as target pitching frequencies exceed the training range. Compared with the capacity-matched models that process the full domain at once, LatentDDM's error is 36-56% lower on larger Darcy domains after adaptation with 16 target simulations. It also improves 20-step field rollouts in fast-pitching airfoil flow, both zero-shot and after few-shot calibration. These results identify the co-designed local pretraining and composition-level transfer as a promising design principle for physical foundation models.

Federated learning allows banks, hospitals, and other regulated organizations to train a shared model without moving raw records off their own servers, which is attractive wherever data protection law or competitive sensitivity rules out pooling data centrally. Two problems limit how far this promise can be trusted in practice. First, the parameter updates that clients exchange still leak information about local records through gradient inversion and membership inference attacks. Second, an honest averaging rule such as FedAvg has no defense against a subset of clients that submit corrupted or adversarial updates, so a small number of malicious or compromised participants can quietly steer the shared model off course. This paper presents a federated learning framework, DP-BR-FedAvg, that combines a Gaussian-mechanism differential privacy layer with a coordinate-wise trimmed-mean Byzantine-robust aggregation rule, evaluated on a simulated cross-institutional classification task resembling fraud and clinical-risk scoring. Across sixty communication rounds with twenty clients, a quarter of them Byzantine, plain FedAvg collapses on the minority class (F1-score 0.030) while the proposed framework recovers substantially more of the signal (F1-score 0.119) while bounding the privacy loss of any single client's contribution. A Byzantine-robust aggregator with no privacy layer performs best in raw accuracy, quantifying the cost privacy imposes on robustness. The results show that privacy and robustness mechanisms interact rather than simply add, and that system design for regulated, adversarial, cross-institutional settings needs to budget for that interaction.

As services move online, trust institutions such as banks, lenders, and governments must verify the identity of remote users. Fraud detection tools are widely available, but evaluating and fine-tuning them remains difficult because identity documents are sensitive and therefore scarce. Synthetic data generation offers a path forward, and demand is clear: our prior work in this area has been downloaded over $11{,}000$ times (aggregated from eight parts). We introduce IDSpace, extending this line of research in three directions. First, we propose model-guided Bayesian optimization, which tunes generation parameters to maximize both visual similarity and prediction consistency with target-domain models given only a few samples from a target domain. Second, we decouple user-specified metadata (demographics, fraud patterns, capture device) from automatically tuned control parameters (font styles, noise levels, image quality), allowing users to configure evaluations without low-level expertise. Third, we expand beyond template images to support scanned and mobile-captured documents. Experiments show IDSpace improves evaluation consistency by $15-45\%$ over baselines including CycleGAN, diffusion inpainting, and non-guided optimization, using only a few real samples, while improving training accuracy by up to $9\%$ and SSIM similarity with the target domain by $10\%$. We also released a new dataset consisting of $359{,}240$ high-quality synthetic documents across ten European ID types.

The last five-plus years have seen many protein engineering disciplines transformed by advances in machine learning (ML), but the same cannot be said for directed evolution. Reflecting on a previously co-authored perspective, I discuss why I believe this to be the case, arguing that a disconnect between the goals of machine-learning-assisted directed evolution (MLDE) researchers–"identify an optimal protein"–and the goals of directed evolution more broadly–"identify a sufficient protein given time and resource constraints"–is a principal culprit. As an example, I highlight how nearly all current MLDE methods neglect to account for the cost of DNA synthesis, resulting in strategies that have limited practical applicability regardless of the underlying models' capabilities. I close by discussing recent works that are exceptions to this overarching trend, and emphasize that the last five years of efforts in ML-assisted protein engineering and the prescribed reframe of MLDE objectives need not be mutually exclusive.

Compliance screening of blockchain addresses is, in practice, a lookup against sanctions registries plus clustering heuristics; it fails on unlabelled addresses and on chains with no label coverage at all. We describe a deployed system that scores an address by its position in a multi-chain transaction graph rather than by its presence in a list. The substrate is a single graph of 835,330,427 addresses and 15,826,261,934 edges across five EVM chains; a shared inductive encoder with per-chain normalisation feeds two scoring heads. Decision thresholds are exact quantiles of the score distribution over the full population, scanned per chain segment, so the alert volume is known in advance. We report: label-free transfer: heads trained on two chains recall 0.8598 / 0.8182 / 0.9967 of held-out positives on Base, Arbitrum and Gnosis at a $10^{-3}$ population alert rate, with no target-chain labels in head training; a static lead-time replay over 68 external registry events: 40 of 68 (58.8%) flagged at the 0.1% budget, $\times$152 over an event-level random-flagging baseline, with first on-chain appearance a median of 528.8 days (Ethereum) / 647.8 days (Tron) before public designation; a serving path whose score is bit-identical to the offline artefact at end-to-end p50 151 ms, gated by a 2,882-address drift panel; and an adversarial harness of eight recurrent reinforcement-learned archetypes that passes an 8-criterion degeneracy audit and, on a detector-independent snapshot, exposes a measured blind spot of the deployed heads against synthesised behaviour.