Papers

Large language models can memorize private or harmful information, motivating machine unlearning methods that remove targeted knowledge while preserving other capabilities. However, existing evaluations rely primarily on behavioral benchmarks, which assess whether a model appears to forget but provide limited insight into how unlearning changes the model or affects related knowledge. We introduce TRIAGE (Tripartite Representation-internal Introspection for Adjacency Gap Evaluation), a benchmark-agnostic evaluation framework for characterizing these changes. TRIAGE uses diagonal approximations of the Fisher information and Hessian to measure changes in parameter sensitivity and local curvature, and utilizes a Forget / Adjacent-Retain / Generic-Retain partition to quantify an adjacency gap in semantically related knowledge. Based on the magnitude and distribution of these changes, TRIAGE further classifies each algorithm's update as no-op, partially localized, collateral dominant, or globally destructive. Across 12 unlearning methods, four language models, and the WMDP, TOFU, and MUSE benchmarks, we find that methods with similar behavioral forgetting can produce substantially different internal changes and patterns of collateral damage. These signatures also vary across models and benchmarks, indicating that the effects of unlearning are not determined solely by the unlearning algorithm. TRIAGE can be applied alongside existing unlearning benchmarks to complement behavioral evaluation with a model-internal view of how unlearning reshapes the model's parameter space and affects retained knowledge.

The neural scaling law relating longer training to better performance through a power law is central to today's large language models (LLMs), yet its origin remains debated. One recent proposal is that power laws can emerge from the strong non-linearity of a single softmax head learning peaked distributions. What happens with multiple softmax functions, as in LLMs, is unclear. Here, we show through toy models that any softmax learning peaked distributions, regardless of its position in the model, can develop logit magnitudes that grow in a power law with exponent $1/3$, becoming a training bottleneck whose loss contribution decays as a power law with the same exponent $1/3$. The overall loss therefore obeys $1/3$ scaling whenever at least one softmax learns peaked distributions. We confirm that many softmax functions in LLMs learn peaked distributions and that LLM loss scaling matches this $1/3$ prediction. Moreover, logit growth dynamics reveal that attention heads, rather than the language modeling head, are the bottleneck likely driving the $1/3$ loss scaling in LLMs. Attention trying to concentrate on specific information, which is the heart of Transformers, may therefore also be the heart of the neural scaling law of training.

In this article, we propose algebraic integer number formats (AINF) as a framework for highly configurable and accurate compressed computation. A key property of AINF is its straightforward reduction to purely integer arithmetic, enabling error-free multiply-accumulate operations. We develop a matrix multiplication engine that exploits this property, obtaining performance within a predictable factor of the most efficient numerical counterparts on modern hardware accelerators. Another key advantage of AIQ is the freedom to select algebraic integers tailored to the numerical characteristics of a given use case, offering optimization opportunities that are fundamentally unavailable in floating-point or fixed-point formats. Motivated by these properties, we introduce algebraic integer quantization (AIQ), a lossy compression format for large language model (LLM) inference, and provide results demonstrating that AIQ achieves competitive accuracy compared to the industry-standard MX format across a diverse set of models and benchmarks at equivalent bit-per-parameter budgets. The experiments presented in this article employ algebraic integers of degree two (quadratic integers); generalizations to higher-order and, in particular, multidimensional algebraic integers promise even greater computational advantages and remain a promising direction for future work.

Comparing parameter-efficient fine-tuning recipes under a single, shared learning rate is a common but flawed practice: when the arms being compared have very different trainable-parameter counts, a shared rate can simultaneously depress the larger arms' means and inflate their variance, manufacturing a large, seemingly multi-seed-significant advantage for the smallest arm that is not a real effect. We document this confound in a concrete setting: post-hoc SVD-based KV-cache compression, where an already-pretrained model is converted to a low-rank (multi-head-latent-attention-style) cache by factorizing its key/value weights into a down-projection ("encoder") and an up-projection ("decoder"), after which a short fine-tune ("healing") recovers the accuracy lost to truncation. Under a shared learning rate, freezing the decoder and healing only the encoder looks like a clear win over healing the decoder or both factors; once every arm is given its own tuned learning rate, that apparent advantage disappears, and encoder-only healing instead reaches parity with the alternatives, at a real, measured saving of 3x fewer trainable parameters and 3x less optimizer-state memory. We verify this parity with per-arm learning-rate tuning and three seeds per configuration on a vision-language model (Qwen2.5-VL-3B-Instruct), at the one compression ratio this protocol covers, and replicate it on a text-only testbed across two backbones. Encoder-only healing is therefore a lower-memory drop-in recipe for retrofitting low-rank KV-cache compression at training time, and the shared-learning-rate pitfall we document and correct is a cautionary result for comparing any fine-tuning recipes whose arms differ in trainable-parameter count.

Automatic Speech Recognition (ASR) systems often underperform for children and non-native speakers, while adapting adult ASR models to child speech can cause adult-speech forgetting. We study child ASR adaptation with adult retention across Arabic and English. We compare full fine-tuning, LoRA, and post-hoc weight-space merging across encoder–decoder, encoder–CTC, and AudioLLM-based ASR systems. Experiments use Arabic native and non-native child speech, English MyST child speech, and adult benchmarks from MGB-2 and LibriSpeech test-clean. We evaluate recognition quality with WER and quantify the adaptation–retention trade-off using Retention Index, Child Adaptation Gain, and Adaptation Recovery. Results show that child adaptation is necessary, especially for non-native Arabic and English child speech, but direct adaptation often reduces adult ASR performance. Bilingual adaptation is more stable than language-specific adaptation. Weight-space merging often improves the trade-off, especially for encoder–CTC, Whisper, and AudioLLM-based ASR, with LERP favoring adult retention and TIES recovering stronger child gains. For the encoder–decoder model, direct bilingual fine-tuning remains strongest in raw WER.\footnote{Code, and models are available at https://github.com/qcri/Child-ASR-Adaptation.

Continual Graph Learning (CGL) aims to incrementally learn from graph-structured data while preserving knowledge acquired from previous tasks. A major challenge in this setting is catastrophic forgetting, where learning new tasks degrades performance on previously learned ones. Hierarchical Prototype Networks (HPNs) address this problem through a prototype-based memory mechanism that avoids storing historical data, but their reliance on linear feature extractors limits their ability to exploit graph topology, while point-based prototypes often lead to inefficient prototype growth on structurally diverse graphs. In this work, we propose HCPN-GCN, a graph-aware extension of HPN that replaces the original linear feature extractors with Graph Convolutional Networks (GCNs) and introduces cone-based prototypes with a diversity regularization objective. The proposed design produces richer graph-aware representations while compactly modeling the embedding space, reducing prototype proliferation without sacrificing discriminability. Experimental results on six continual graph learning benchmarks demonstrate that HCPN-GCN consistently improves average classification accuracy over the original HPN and representative continual learning baselines while maintaining near-zero forgetting. Furthermore, our analysis shows that the proposed model learns substantially richer class-level prototype hierarchies using approximately $30\times$ fewer atomic prototypes than the original HPN, providing a more compact and effective memory representation for continual graph learning.

Ageing bridge infrastructure is a growing global concern, yet conventional Structural Health Monitoring (SHM) systems are costly and difficult to scale, and routine visual inspections remain subjective. Drive-by, or indirect, bridge inspection, in which a sensorised vehicle recovers structural information from vehicle-bridge interaction (VBI) and vehicle-road interaction (VRI) responses, offers a scalable alternative. However, key challenges remain unresolved, including separating bridge responses from road roughness, detecting damage under normal traffic, and generalising across diverse bridge types. This paper presents a vehicle-integrated digital twin framework that unifies physics-based modelling and machine learning for continuous monitoring of bridge and road conditions. The framework comprises three pillars. First, surrogate models of VBI and VRI are constructed using a Fourier Neural Operator that learns function-to-function mappings from operating conditions to vehicle responses. Trained on both simulated and field data, these surrogates deliver millisecond-scale inference, replacing computationally intensive full-order analyses. Second, the design of a custom electric inspection vehicle, its sensor layout, and signal processing chain are optimised through Bayesian optimisation to maximise bridge information yield while suppressing road and vehicle noise. Unsupervised damage-assessment pipelines based on adversarial autoencoders, matrix profiles, and transformer architectures have been developed and validated to process the resulting vehicle data. Third, the complete workflow is validated through coordinated multi-site field trials in Australia and Japan, covering a range of bridge types, traffic conditions, and environmental settings.

The partial area under the receiver operating characteristic curve (pAUC) is an important performance metric for binary classification that summarizes true positive rates within a specific range of false positive rates (FPRs). Classifiers that achieve high pAUC need to be obtained in many real-world applications such as cybersecurity, medical care, and advertising. Although many methods for maximizing the pAUC have been proposed, they typically require both labeled positive and negative data for training. However, in practice, labeled negative data are often difficult to collect due to privacy concerns or the need for high expertise to annotate them. In this paper, we propose a method for maximizing the pAUC from positive and unlabeled (PU) data without negative data. Within an empirical risk minimization framework, we show that the pAUC, including its FPR-dependent thresholds, can be represented using only the positive and marginal densities, and derive an empirical estimator from PU data. A classifier is then trained by maximizing the derived smoothed empirical pAUC estimator. We experimentally demonstrate the effectiveness of the proposed method with ten real-world datasets.

We introduce a dataset of approximately 10,000 Reynolds-Averaged Navier-Stokes (RANS) simulations of steady, incompressible, two-dimensional subsonic flow around downforce-generating NACA 4-digit airfoils, targeting aerodynamic regimes relevant to automotive and motorsport applications (openly available on https://huggingface.co/datasets/ratiolabs/downforce-airfoils). Using this resource, we study geometry-conditioned neural surrogates for fast flow prediction, comparing neural fields with neural ODEs, MLPs with graph-based models, and several spectral geometry-conditioning methods. We further propose Airfoil2Vec, an airfoil-specific spectral geometry encoder that combines the joint contour spectrum with separate spectral representations of camber and thickness, for predicting continuous pressure and velocity fields. We evaluate generalization through angle-of-attack interpolation, interpolation and extrapolation to unseen NACA 4-digit geometries, and generalization to unseen non-NACA airfoils. The resulting surrogate accurately captures aerodynamic quantities and qualitative flow features while providing orders-of-magnitude speedups over conventional computational fluid dynamics.

Behavioral tests measure how a language model reads evidence. We ask whether those measurements help choose a recommendation interface. We evaluate six small instruction-tuned checkpoints across four recommendation domains with chronological evaluation and 3,426 evaluation users. Each request ranks eight candidates. A baseline selector chooses among history-only prompting, prompting with collaborative evidence, and score fusion. It uses observable features and six stability prompts that vary wording and candidate order. An augmented selector adds features from six evidence-reading prompts that ask the model to compare support counts. An interface chosen once on development (validation) data for each domain and checkpoint scores 0.5524 NDCG@5, compared with 0.5447 for the baseline selector and 0.5428 for the augmented selector. Adding the diagnostic features changes NDCG@5 by -0.0019 (95% interval [-0.0046, 0.0004]). The interval includes zero, and its upper bound is below the analysis plan's 0.005 improvement target. Matching the selectors' hyperparameters also leaves the interval upper bound below that target. Evidence from retrieved similar users improves prompting by 0.0999 NDCG@5 over a control using randomly selected users matched for activity. The evidence-reading tests also reveal answer-position and tie-response biases. These results concern the tested selectors and candidate sets. They illustrate why diagnostic measurements should be evaluated by whether they improve recommendation choices beyond existing features and a fixed interface.

Modern bioacoustic foundation models like Perch and BirdNET can identify species with high discriminative accuracy, yet their confidence scores are often uncalibrated and difficult to interpret as probabilities of real-world occurrence. This limits their use for ecological inference beyond threshold-based detection. We leverage a global annotated acoustic dataset (WABAD) to produce calibration priors for an acoustic model, optionally incorporating species-level information. We introduce new methods of fusing the acoustic predictions with geopriors, which empirically improves calibration while preserving discrimination. Together, these results suggest a path to simpler and more reliable acoustic monitoring for broad biodiversity.

This research describes the adaptation of Large Language Models (LLMs) for economic monitoring in the public sector to automatically determine whether an article discusses Economic Policy Uncertanity (EPU) and to identify its specific type. Previous studies either rely on keywords, which often result in a high count of false positives, or use machine learning approaches that require a large number of quality human labeled data that is costly and time consuming to acquire. In this study, we propose approaches based on weak supervision techniques, using generative LLMs to create synthetic labels through prompting, making the approach both cost-effective and scalable. Additionally, we propose methods for for multi-label and hierarchical classification of articles related to EPU.

Optimizing deployed AI systems increasingly amounts to editing prompts, skills, harnesses, and code rather than model weights. Existing approaches commonly optimize these artifacts through propose-evaluate-select procedures, where candidate configurations are evaluated and only those meeting an acceptance criterion are selected. Yet our analysis shows that discarded candidates often contain information critical for subsequent optimization. Discarding them causes later proposals to revisit the same failure modes. We introduce Mara Chain, a refinement procedure that turns rejected candidates into stepping stones. Rather than discarding a rejected candidate, Mara Chain retains and iteratively refines it using evidence accumulated across preceding attempts. The procedure limits each refinement chain to a fixed depth and applies Pareto-filtered Top-N selection to bound the candidate pool. Across AppWorld skill optimization, TerminalBench 2.1 harness optimization, and MuSiQue retrieval-pipeline optimization, Mara Chain delivers greater task-performance gains with fewer rollouts. It outperforms GEPA, ACE, and SkillOpt-Lite by up to 20.5% in relative performance on AppWorld, reaching the target score with 65.5% fewer rollouts than GEPA. It improves the pass rate by 20.2 and 22.5 percentage points over AHE and Meta-Harness on TerminalBench 2.1, respectively, and improves MuSiQue test nDCG@10 and Recall@10 by 0.104 and 0.131 over a hand-written retrieval pipeline.

In the field of Machine Learning, many papers contain empirical results supporting claimed statements or illustrating the performance of a proposed method. However, most practitioners know that (1) results are generally hard to reproduce, and increasingly so, (2) code is not often available to do so, and (3) it hinders the development of research. In this position paper, we analyze and quantify these issues, and make concrete proposals to improve result checkability, if not reproducibility. Code and supporting materials are available at https://github.com/giddyyupp/position-enforce-verifiability.

Pooled statistics of Transformer weights obscure how magnitude is distributed across functional channels, while individual weights are too numerous to compare directly. We study the mesoscopic level between them: row and column scale fields, the median-centred log-RMS profiles of a weight matrix over its channels, which together with a global scale and a full balanced core represent the matrix exactly. Across public Pythia checkpoints at four sizes and controlled runs from three initialization families, balancing reveals similar measured core magnitude profiles. A mixture bridge, with its form fixed before the analysis and its coefficients fitted, predicts the pooled-shape departure from field width on held-out runs and data arms of the controlled grid. The indexed fields retain further structure: they align across projections that share a functional channel, and query/key profiles follow reassigned RoPE frequencies rather than fixed matrix coordinates. Training trajectories show early field formation followed by component-dependent broadening or recession. Extending the channel-based analysis to AdamW's second moment reveals related functional organization in its log-space row and column factors. Finally, edits of a frozen checkpoint separate reciprocal scale balance, which preserves the forward computation, from relative channel gain: flattening the gain increases in-distribution loss while preserving matrix norms and the balanced core. Row and column scale fields thus connect pooled magnitude statistics to channel organization and provide coordinates for tracking and testing trained weight structure.

On-board compression of synthetic aperture radar (SAR) phase history is bandwidth-critical, and block-adaptive quantization (BAQ) remains the operational standard. We test whether a small convolutional autoencoder, with its encoder on the sensor, can compete with BAQ on complex phase-history patches from the AFRL GOTCHA collection. Every method is charged for all transmitted bits, rates are reported in bits per complex sample (b/cs), and detection is scored by one-to-one matching of CA-CFAR detections. The autoencoder (28,656 encoder parameters) loses at every rate. At 16 b/cs it reaches -2.87 dB NMSE, against -35.5 dB for 8-bit BAQ with $\pm 3\sigma$ clipping and -41.0 dB with a tuned clipping range. It also loses to a $16 \times 16$ block Karhunen-Loève transform (KLT), a local linear coder with a tenth of its encoder cost (-5.39 dB). Running the network in a companded Fourier domain helps, but its detection F1 remains bounded at 33%. The evidence points to this model, its normalization, and its objective, not to a fundamental limit of learned coding. Per patch, the data have modest lag-1 coherence ($|\rho| \approx 0.3$) and patch-specific spectral concentration. Two findings concern evaluation itself. First, 97% of CFAR crossings on raw $64 \times 64$ patches are border artifacts of the zero-padded detector. Second, on interior cells BAQ's clipping range decides detection: 8-bit BAQ keeps 69% F1 with tuned clipping but 17% at $\pm 3\sigma$, and at 8 b/cs or less adaptive FFT thresholding preserves more detections than BAQ. We close with an evaluation protocol for learned radar compression.

Macroeconomic productivity metrics, such as Total Factor Productivity, register technological breakthroughs with multi-year reporting lags due to administrative survey intervals and national accounting conventions. This paper introduces Hyperspherical Semantic Trajectory Analysis (HSTA), an unsupervised quantitative methodology that tracks technology diffusion directly from unstructured scientific and commercial text streams. We analyze 30,000 filtered document records spanning academic preprints from arXiv and patent application records from the USPTO. By projecting high-dimensional Transformer sentence embeddings onto unit hyperspheres using Spherical K-Means clustering across eight primary sub-topics and UMAP manifold reductions, HSTA formalizes two quantitative metrics: (1) Semantic Centroid Vector Drift, which tracks vocabulary shifts between temporal sub-corpora to identify structural paradigm transformations; and (2) Commercialization Offset, which evaluates cross-corpus peak density alignments between scientific discovery and intellectual property filings. Linking quarterly topic volume velocity with physical hardware metrics from the Epoch AI database, Vector Autoregressive F-tests demonstrate that quarterly paper volume velocity alone does not Granger-cause frontier compute allocation surges at conventional statistical significance levels, highlighting the necessity of conditioning textual signals on physical capital constraints. Empirical results reveal that sub-topics covering Large Language Models (with a drift metric of 0.332) and Artificial Intelligence Systems (with a drift metric of 0.234) undergo the highest rate of semantic evolution, offering an objective, real-time mechanism to complement traditional economic statistics.

Temporal dynamics in learning behavior can be revealed through period segmentation in Transition Network Analysis (TNA). Cristea et al. demonstrated that segmenting courses into halves and quarters reveals how learning strategies evolve and relate to academic performance. Building on this approach, we investigate whether Topological Data Analysis (TDA), specifically connected components ($\beta_0$) change points from Zigzag Persistent Homology, can provide data-driven period boundaries that identify intervention windows. Analyzing 22 courses from the Open University Learning Analytics Dataset, we find that $\beta_0$-based segmentation captures greater between-period variation than time-based segmentation (median variance ratio (VR) = 4.10$\times$; 21/22 courses show VR $>$ 1.1). Permutation tests confirmed significance ($p < 0.05$) in 18% of individual courses, with 73% showing positive improvement over random breakpoints. Only one course showed better performance with time-based segmentation. However, analysis of academic outcomes reveals a key insight: final-period behavior shows lower correlation with outcomes in $\beta_0$-based segmentation than in time-based segmentation, reflecting a marked behavioral collapse where engaged learners rapidly disengage. We interpret this as evidence that final-period behavior reflects rather than causes outcomes: students who will pass maintain engagement, while those who will fail disengage. This reframes the value of $\beta_0$: rather than improving prediction, change points identify intervention windows. These are periods where behavioral structure shifts and targeted support may be most effective.

Deploying AI systems requires deciding when to delegate tasks and when humans should intervene to monitor and mitigate risk induced by AI operations. These decisions are challenging when failures cluster: a hallucination or harmful output can trigger further errors, creating periods of elevated risk. We introduce a continuous-time framework for learning adaptive human oversight under such turbulent AI risk. Existing oversight and delegation formulations condition on history but do not model incident clustering, or its suppression by supervision effort, jointly with the delegation decision and this study fixes this gap. The self-exciting dynamics capture how risk events increase the likelihood of subsequent events, making their timing and history central to decision-making. We formulate a stochastic control problem combining human actions, monitoring effort, and switching between human-AI-assisted operation and full AI delegation, balancing operational rewards against oversight costs, and cascading AI-failures and induced uncertainty. Human participation is an endogenous component of risk management: the policy determines both when oversight is needed and how much effort to allocate. We study a relaxed switching formulation and propose Hawkes-PPO, a policy-gradient method that uses a bank of exponential filters of observed incident times. In a synthetic environment it attains a higher risk-adjusted objective than either fixed regime and approaches an approximate full-information oracle. We illustrate our results with numerical simulations by examining how cascade risks influence intervention and delegation, connecting reinforcement learning with adaptive human oversight of AI systems. In particular, we illustrate the benefit of our switching strategy and Hawkes-PPO algorithm to monitor the project efficiently along time, reducing turbulent risks occurrences and costs.

Although $\mathbb{E}[\mathrm{dropout}(x)] = x$, here we show that $\mathbb{E}[\mathrm{LayerNorm}(\mathrm{dropout}(x))]$ is not equal to $\mathrm{LayerNorm}(x)$. Accordingly, the pattern of a Dropout layer followed by a LayerNorm, which is common to many AlphaFold2-based protein structure predictors, produces a systematic bias at evaluation time that can hamper performance. To address this, we derive a closed-form, first-order correction for this gap, which we call a Dropout-LayerNorm Correction (DLC). DLC empirically matches the performance boost of large Monte Carlo dropout ensembles. We evaluate its effect across nine protein structure models (ESMFold, OpenFold, ABB3, FlashABB, Ibex, NbForge, Genie1, Genie2, Genie3) on both paired and single-chain antibody structures as well as one protein-ligand docking model (QuickBind). The correction is computationally negligible and improves accuracy in all ten models tested ($\sim0.3\%-13\%$), with a modest but consistent improvement in ESMFold and OpenFold and a substantial improvement in antibody-specific models. This work identifies the mathematical consequence of chaining together Dropout and LayerNorm and provides a free, principled adjustment to improve the evaluation performance of many pretrained models. Code to reproduce the experiments is available at https://github.com/oxpig/DLC.

How many bits does a network need before its accuracy collapses, and how does this grow with depth? We study the precision floor, the perturbation level or bit-width at which accuracy falls halfway to chance, in MLPs, CNNs, Vision Transformers and nine pretrained language models, under post-training quantization (PTQ) and quantization- or noise-aware training (QAT). (i) A first-order theory sets the floor through one full-precision quantity, the predictive amplification $G$: $\eta_c=\Lambda/G$, and $G^2$ grows linearly in depth at a rate proportional to the squared residual branch scale. (ii) The predicted equality $\alpha_{PTQ}=\rho$ of depth exponents holds within 95% intervals in twelve of thirteen trained architectures and in GPT-2 from 12 to 48 layers, with $\Lambda=1.45\pm14\%$ across trained architectures. (iii) Residual branches scaled by $1/\sqrt{D}$ and pre-normalisation remove the depth penalty, and each quantizer turns noise into bits at a rate fixed by its step rule, giving $b_c=(\alpha/\gamma)\log_2 D+C$. (iv) A QAT paradox: noise-aware training roughly doubles the tolerable noise of shallow networks, but the gain decays with depth, so the depth law steepens ($\alpha_{QAT}/\alpha_{PTQ}=1.45$-$1.47$ on two datasets, ten seeds each). Decision margins, cross-layer error cancellation and heavy tails do not set the floor.

Achieving precise control over multiple properties without sacrificing chemical validity remains a central challenge in molecular inverse design. Existing reinforcement learning (RL) methods fine-tune graph diffusion models by treating **reverse** sampling as a sequential policy, using a single terminal reward to optimize hundreds of coupled decisions. They often suffer from instability, validity collapse, and limited property gains. We introduce GraphFDM (Graph Forward Distribution Matching), a new online RL paradigm for graph diffusion that performs optimization through the **forward** process. GraphFDM uses valid generations to define a reward-tilted target distribution jointly optimized over graph size and molecular structure for each property condition, incorporating reinforcement signals into supervised learning without storing reverse trajectories. We derive the unique optimal target, prove a condition-wise improvement guarantee, and show that the fixed graph-size prior of standard graph diffusion leaves an irreducible matching gap. In multi-conditional polymer and small-molecule generation, GraphFDM achieves the lowest MAE on every target property, with reductions of up to 53.0% relative to the strongest baselines and chemical validity above 0.99. It further generalizes to out-of-distribution property combinations.

Model misspecification poses a fundamental challenge in multi-agent reinforcement learning, where transition uncertainty can be amplified by strategic interactions among agents. Distributionally robust Markov games (DRMGs) provide a principled framework for addressing such uncertainty, yet existing methods often rely on restrictive assumptions or scale poorly to large state and joint action spaces. We study online learning in general-sum DRMGs with general function approximation and $\phi$-divergence uncertainty sets. We propose RoMEX-$\phi$, a model-free framework that integrates equilibrium-based exploration with dual fitted learning. Through a functional dual representation of the robust multi-agent Bellman operator, RoMEX-$\phi$ enables tractable worst-case value estimation from nominal interaction data using a centered empirical robust discrepancy. We introduce the robust Multi-Agent Decoupling Coefficient (robust MADC) to characterize the intrinsic exploration complexity arising from strategic interactions and adversarial transition uncertainty. We establish sublinear robust regret guarantees governed by the robust MADC rather than explicitly by the state and joint action space sizes, replacing tabular dependence with intrinsic function-class complexity. Numerical experiments on a scalable general-sum DRMG under total variation uncertainty show that RoMEX-$\phi$ is substantially more resilient to transition shifts than its non-robust counterpart while remaining competitive with an exact tabular robust baseline. Our results provide a scalable framework for distributionally robust multi-agent reinforcement learning with general function approximation.

Modern Document AI reads contracts, extracts fields, reasons over tables, and grounds answers to page regions, then forgets everything. Processing an amendment the next day begins from scratch: no schema retained, no contradiction detected, no experience carried forward. This is a structural choice, not a scale failure: current systems are stateless functions. We call this the statelessness bottleneck. This bottleneck lies beyond parameter scaling, context extension, and retrieval augmentation: storage provides persistence and retrieval provides access, but neither consolidates observations into knowledge that improves future processing. This survey formalizes persistent evidence-grounded document state as a unifying framework, specifying the operations and invariants required to convert multimodal evidence into durable, provenance-linked state. We introduce a statefulness audit showing that ten representative benchmarks, coded against eight statefulness criteria, leave cross-session state evolution untested, and derive a longitudinal benchmark harness with five counterfactual metrics: Experience Gain, Cost Efficiency, Memory Harm, Forgetting Fidelity, Coverage Retention, to characterize the benefit, cost, risk, and governability of persistent document state. Document AI lacks mechanisms coupling persistent state to document-native structure, provenance, and temporal validity. The next era of Document AI will be defined by what systems retain across documents, sessions, and time.

Physical systems evolve continuously in time, yet their states are typically observed only at discrete times. Generating trajectories consistent with their probability densities from such observations therefore requires capturing the continuous-time evolution rather than only learning transition mappings between consecutive observations. We propose PhiBE-Flow, a framework that directly estimates the probability velocity field induced by the stochastic differential equation (SDE) which governs this continuous-time distributional evolution. PhiBE-Flow learns from discrete observations using a model-free approach requiring neither known SDE coefficients nor score estimation. We establish convergence guarantees for the method, accounting for both time-discretization and finite-sample errors. We evaluate PhiBE-Flow on systems of increasing complexity, from controlled stochastic numerical systems to Navier–Stokes dynamics and real-world videos. Our results show that PhiBE-Flow accurately recovers probability flows of stochastic dynamics, preserves multiscale physical statistics, and improves video generation performance over representative baselines. The code is available at https://github.com/R1fe/PhiBE-Flow.

Modern neural predictors can model highly nonlinear maps, but many scientific and engineering tasks require reasoning in the opposite direction: given a performance or safety level, the goal is to characterize the preimage, that is, the complete set of inputs which meet the desired target level and optimize over that set. For expressive neural predictors, however, such preimages typically have no explicit representation and are expensive to recover or optimize over. This creates a fundamental three-way challenge between expressive forward prediction, accurate preimage approximation, and tractable optimization over the preimage for downstream tasks. We introduce TRIO (tractable representations for preimage learning and inverse optimization), a framework for learning representations that make these objectives compatible by construction. Our key contribution is a preimage factorization: the forward model remains expressive through nonlinear radial transformations (including neural networks), while, under inversion, each transformation reduces to a single scalar radius, which yields simple geometric level sets. This yields an explicit geometric representation that is reusable for downstream optimization over the preimage, and, for linear objectives, we show that this admits a closed-form global solution. We finally prove a universal approximation theorem which shows that TRIO can approximate any continuous forward map and its entire family of potentially disconnected, nonconvex preimages arbitrarily well. Hence, TRIO combines expressive forward modeling, exact preimage recovery, and tractable global downstream optimization over preimages by design.

Recent advances in sensor technology have enabled more effective human activity recognition (HAR), particularly in real-time systems with limited computational resources. However, Ultra-Wideband (UWB) radar data remain challenging due to high dimensionality, noise, complexity, and nonlinear characteristics. This research proposes a framework to efficiently reduce data size, uncover significant patterns, and classify six activity types (Standff, Liedown, Noactivity, Sit, Stand, and Walk) from UWB signals with high accuracy. Two novel dimensionality reduction techniques are introduced in this paper. The first, Clustered Polynomial Expansion with Incremental PCA (CPE-IPCA), combines clustering and polynomial feature expansion with Incremental PCA, preserving 100% of the variance in only 50 components. The second, Post-PCA Standardization Approach (PPSA), standardizes data after PCA and retains 99.1% of the variance in 80 components, achieving superior compression and computational efficiency compared to conventional nonlinear methods. Frequent patterns are identified using Apriori and FP-Growth, which are then classified with Random Forest and a Vector Space Model (VSM). The framework achieves 100% accuracy with Random Forest on CPE-IPCA and 99% on PPSA, while VSM attains 100% precision, recall, and F1 on PPSA and near-perfect performance on CPE-IPCA (precision 1.00, recall 0.98-1.00, F1 0.99-1.00), demonstrating a fast, interpretable, and robust HAR system suitable for healthcare, assisted living, and smart environments.

Encoded demographic information in internal model representations is a commonly assumed risk factor for algorithmic bias, with demographic representation invariance often being touted as the ideal state. However, while demographic shortcut learning is a genuine threat, some degree of encoding is necessary when demographics correlate with target labels. Here, we show, mathematically and empirically, that enforcing demographic invariance can actually hamper bias mitigation and even create new biases. We distinguish marginal from class-conditional representation invariance, and show that they imply the standard group fairness notions of demographic parity and equalized odds, respectively. We evaluate the effects on predictive performance and fairness of enforcing both invariance types, both theoretically and empirically across five tabular and two chest X-ray imaging datasets. Our findings support our mathematical argument that demographic representation invariance is neither desirable nor sufficient for fairness.

The rank $r$ in LoRA is widely treated as a capacity control: a smaller rank is assumed to yield a simpler model that generalizes better. We show that, under hard per-factor norm budgets—the idealization of the weight decay and norm control used in practice—this intuition breaks down. The reason is structural: under such budgets, the updates LoRA can reach are exactly the matrices of rank at most $r$ inside a nuclear-norm ball, and every complexity and displacement functional we analyze is maximized over this set by a rank-one update—so the rank cap never binds. The consequences follow directly. The linear-readout model class we study is identical for every $r \ge 1$, its Rademacher complexity carries no dependence on $r$, and the distance the adaptation can move the source distribution obeys a rank-independent upper bound that we show is sharp. If rank does not control capacity, where does it act? We identify two places. Statistically, replacing the per-factor budgets with a joint budget on the product restores a data-dependent, rank-sensitive complexity bound—though the gain appears only for well-spread feature distributions, and the worst case remains rank-free. Spectrally, rank sets the price of adaptation: canceling the leading singular directions of the pretrained weight requires both sufficient rank and sufficient budget. We bound the smallest rank achieving a desired source–target alignment, with upper and lower bounds that match under two-sided spectral decay. Together, these results recast rank as governing which updates are reachable and what cancellation costs—not how much capacity the model has.

Whereas previous results on abilities and limitations of transformers have restricted the definition of transformers in various ways, here we study softmax-attention, multi-layer transformers operating on real values, with very few additional assumptions. Applying results from real geometry, we obtain upper bounds on the VC dimension and split VC dimension of such transformers ($O(n^4)$ and $O(n^6)$, respectively, where $n$ is the input length). Conversely, we also construct specific transformers witnessing lower bounds on these quantities ($\Omega(n)$ in each case). These results have some notable consequences. For example, within the class of symmetric (permutation-invariant) functions, we show that transformers can uniformly express all functions over an alphabet of one symbol and non-uniformly express all functions over an alphabet of two symbols, but cannot (even non-uniformly) express some functions over an alphabet of six symbols. We also prove limitations on how many bits of a real number a transformer can access.