Papers

We study switching regret in adversarial multi-armed bandits, where the learner competes with an arm sequence that changes at most $S$ times. When $S$ is known, an optimal expected regret of $\widetilde{\mathcal{O}}(\sqrt{(S+1)KT})$ is obtainable [Auer et al., 2002]. However, when $S$ is unknown, Marinov and Zimmert [2021] show that this guarantee is impossible under an adaptive adversary. In this paper, we show that a single algorithm achieves $\widetilde{\mathcal{O}}(\sqrt{(S+1)KT})$ expected regret for every $S$ against an oblivious adversary, resolving an open problem of Auer et al. [2019b]. Our algorithm combines a fixed-share learner initialized with a small learning rate and dyadic-interval subroutines that search for local improvements using randomized learning rates and implicit exploration. Importantly, a non-uniform prior favors following the main learner, keeping the cost of maintaining many subroutines small. When the subroutines accumulate sufficient improvement over the main learner, its learning rate doubles, allowing adaptation to the unknown number of comparator switches $S$.

Learning-based adaptive control of robotic manipulators with non-observable friction memory has been addressed by attention- based meta-controllers whose number of attention heads is fixed before training and is tuned by costly offline search. At long memory horizons, such fixed-capacity controllers are prone to catastrophic failures on a sizeable fraction of training seeds. The present paper introduces a runtime mechanism that grows and prunes the heads of the attention block during reinforcement learning, governed by two signals: the effective rank of the on-policy context distribution, which triggers growth when representational capacity becomes insufficient, and the per-head output magnitude, which flags redundant heads for removal. Policy continuity at growth events and a quantitative bound at prune events are established analytically. On a two- link manipulator with Stribeck friction, the proposed mechanism attains full success across all memory regimes, eliminating the long-horizon failure mode and removing the need for offline tuning of the head count.

Most hyperparameter configurations for Evolvable-Substrate HyperNEAT (ES-HyperNEAT) produce networks that stagnate at random-guessing performance, wasting computational resources. We frame early stopping as binary classification on early fitness trajectories: for each trial, we compute the cumulative median of best-per-generation fitness and test it against a threshold derived by maximizing the F1 score on an initial 90-trial dataset. The resulting rule (generation G* = 3, threshold T* = 0.140) achieves F1 = 0.872 on 180 independent validation trials, retaining over 90% of successful trials while cutting computational cost by 41.6%. Compared to Hyperband, our domain-specific rule is 64% more efficient with higher mean fitness, though Hyperband occasionally discovers higher peak solutions. On a converged search population the rule becomes too aggressive (recall 31.1%), motivating adaptive thresholds. The specific thresholds are ES-HyperNEAT-specific, but the methodology, deriving stopping criteria from fitness dynamics classification, is applicable to other evolutionary algorithms with stagnation-prone hyperparameter spaces.

Can attention move a shock across several cells in one update without breaking it? We develop a conservative, fixed grid finite-volume scheme in which a CFL-conditioned attention flux selects upstream information according to the transport required by the current time step. One-dimensional inviscid Burgers transport is used as the central mechanism test: the same learned flux remains reliable in the conventional small-step regime and, with a time step four times larger, preserves sharp shocks while using one stage per update. A standard fifth-order WENO scheme with third-order strong-stability-preserving Runge-Kutta time integration (WENO-5+SSP-RK3) is included alongside controlled Forward Euler comparisons to separate flux selection from time integration. The learned attention shifts upstream with the local transport reach and becomes more selective near shocks; inference-time interventions and retrained ablations show that transport-scale information and state-dependent selection contribute directly to performance. Directional two-dimensional scalar Burgers transport and the one-dimensional shallow-water system then test whether the conservation-scale-selection principle transfers beyond the original scalar setting. The results support attention as a learnable information stencil for conservative large-step shock transport, while identifying finite candidate reach and problem-dependent robustness as the present limits.

As the use of Large Language Models moves from chatbots into agentic systems, where outputs become actions with irreversible consequences on reality, the existing paradigm on AI Interpretability research, post-hoc interpretability, is structurally inadequate for safe and trustworthy model deployment: it explains behavior after the fact but cannot audit or intervene in an inference computation before it commits to an output. We therefore argue for a shift toward generative interpretability, an architectural property under which a model's inference pass natively exposes semantically meaningful checkpoints that are human-understandable and amenable to causal intervention. We show the merits of generative interpretability as comparison to other interpretability research paradigms, and propose Neuro-Symbolic Models as a concrete instantiation.

We develop a spectral three-term modification of the classic Hestenes–Stiefel conjugate gradient algorithm, preserving its anti-jamming characteristic and, simultaneously, taking care of the sufficient descent property. We discuss how a modified secant equation can be extracted from our modification scheme, yielding a memoryless BFGS updating formula. Then, the spectral parameter of our method is obtained by steering its direction toward the given BFGS direction within a least-squares context. Using our technical improvements, we outline the general framework of our algorithm and discuss its theoretical features, including the descent and convergence properties, without the convexity assumption. We put our algorithm to the test in comparison with the three other conjugate gradient algorithms on a set of CUTEr unconstrained optimization test models, comparing the outputs using the Dolan–More measure. Next, we provide a concise evaluation of the results, highlighting the practical advantages of our algorithm. As a real-world case study, we introduce a reduced quadratic surface SVM with the rescaled loss for robust binary classification and apply the proposed algorithm to assess its accuracy and training time against several other SVMs.

Critically ill patients with acute kidney injury (AKI) on continuous renal replacement therapy (CRRT) face high mortality, yet current risk assessment relies primarily on clinical parameters from electronic health records (EHR) and ignores minute-level circuit pressure waveforms generated by CRRT machines that track the extracorporeal circuit's interaction with the patient. Clinicians therefore cannot see deterioration as it develops. Risk is reassessed only when labs are drawn, while this continuous record is discarded because it is contaminated by shared-device records, non-physiological minutes, and sensor artifacts. To make the stream usable, we aligned machine records to charted therapy intervals to prevent cross-patient leakage, removed priming and downtime minutes, tuned denoising on a synthetic spike-injection benchmark, and masked unobserved intervals rather than imputing them. On this cleaned stream, we define a rolling day-wise task and a transformer-based stacked ensemble that late-fuses a window-reduced sequence transformer with classical models using circuit-instability features and clinical EHR variables. In a leak-safe benchmark on the multi-center CRRTnet cohort (976 patients, 4,585 treatment days), the machine-only model had the lowest standalone prognostic value (AUROC 0.625), followed by the EHR-only model (0.717). Integrating EHR and machine streams reached a one-day mortality AUROC of 0.766. SHAP attribution showed that circuit-instability descriptors raised the machine share of the top 15 combined-model features from 3 to 7 (20.0% to 46.7%), highlighting filter pressure, transmembrane pressure (TMP), and access-to-return difference (ARD). To our knowledge, this is the first patient-level mortality prediction incorporating CRRT machine data, turning a discarded bedside stream into a continuous risk signal.

We present GeoTTER, a novel framework that redefines optimal transport in the realm of zero-shot classification. Conventional methods often suffer from miscalibration and a lack of adaptability, as they rely on fixed cost matrices derived solely from pre-trained model embeddings. In contrast, GeoTTER addresses these limitations by incorporating two key techniques. First, to alleviate high-frequency label jaggedness (sample-level manifold jitter that assigns neighboring embeddings to different classes), GeoTTER integrates local geometric structure into the optimal transport formulation via graph-Laplacian smoothing, a technique grounded in spectral graph theory that enforces neighborhood consistency. Second, to correct coherent angular drift (a low-frequency orientation bias in which large groups of samples share the same angular offset from their true label prototypes), we fuse clustering-guided cost components with a globally adjusted transport cost, achieving a multi-objective optimization that respects both global distribution constraints and latent data structure. With a median improvement of +6.82% compared to zero-shot and +2.13% compared to OTTER, GeoTTER shows robust improvements across a diverse set of benchmarks.

Ensuring the reliability of black-box machine learning models in safety-critical space missions remains a significant challenge, particularly when ground-truth is unavailable for validation. Although machine learning models offer a powerful means to augment standard pipelines by extracting transmission spectra from complex exoplanetary light curves, their susceptibility to unmodelled instrument anomalies, stellar activity, and domain shifts introduces unquantified risks. This study evaluates a modular safety cage architecture that operates as a parallel monitoring layer to assess the validity of a prediction without modifying the underlying estimator. By monitoring different runtime indicators, including uncertainty quantification, out-of-domain detection, and influence functions, the framework constrains the model's operational domain to a verified region. A controlled evaluation is conducted under both in-domain and cross-domain conditions, using datasets from the 2019 and 2021 editions of the Ariel Data Challenges. The results reveal that model failure is multifaceted and that no single indicator captures all failure modes, demonstrating the need for indicator fusion. The application of safety-driven rejection strategies shows that a modest 20% reduction in data coverage results in error reductions between 45% and 65% across different domains and evaluation metrics. Using a formalised coverage-risk framework, a systematic analysis of indicator combinations is performed to identify configurations that maximise risk-ranking accuracy and optimise the trade-off between data coverage and scientific performance. Safety cages provide a transparent mechanism for detecting unreliable predictions and represent a critical step towards the safe deployment of data-driven models in scientific applications, such as astrophysics, where ground truth is seldom available.

Federated fine-tuning of large language models with low-rank adaptation (LoRA) reduces the number of trainable parameters, but communication remains the dominant cost, and protocols are usually compared by parameter-count ratios rather than by measured bytes. This paper measures per-round upload and download bytes for five federated LoRA protocols, three of them from prior work, and places them on a single communication-quality frontier scored by held-out instruction-following loss. The frontier has a knee. ReverseAdaptive, which learns both LoRA factors before freezing one once the relative improvement in training loss falls below a dimensionless threshold, sits at that knee: it cuts measured round-trip communication by 40.5% relative to FLoRA at a held-out loss cost of 0.0063, and beats FFA-LoRA, which freezes that factor at initialization, by 0.0182 in held-out loss, more than twenty times the largest per-method seed standard deviation. The same threshold carries to LLaMA-3.2-3B without retuning, where it saves 30.0%.

Large language models are increasingly benchmarked against classical machine learning for network intrusion detection (NIDS), almost always using same-dataset evaluation, and that protocol turns out to be incomplete. Evaluating XGBoost and RoBERTa-LoRA on two independently collected NetFlow v2 networks across three axes (same-dataset performance, cross-dataset transfer, and adversarial evasion) reveals no universal winner. The two models are statistically tied same-dataset. XGBoost wins decisively under cross-dataset distribution shift, by 15 points of F1 and 25 points of balanced accuracy; on the target network RoBERTa-LoRA's false positive rate reaches 0.78, leaving it barely above chance despite a superficially moderate F1. RoBERTa-LoRA wins decisively under adversarial evasion, by roughly 17 points of F1 at a representative mid-range perturbation strength, while both models hold false positive rates below 0.01 throughout. The model an evaluator would recommend therefore depends entirely on which axis is tested, not on same-dataset accuracy alone. A staged feature-leakage ablation improves cross-dataset transfer non-monotonically, indicating the leakage signal is distributed across the feature representation rather than confined to a few columns, and cross-dataset transfer between our two networks is strongly directional. These results argue for evaluating NIDS models along multiple independent robustness axes, and with more than one metric per axis.

Brain-computer interfaces (BCIs) decode neural activity to restore lost function. Typically, training a high-performance neural decoder requires a large labeled dataset to be collected from every new subject. One way to reduce the labeled data cost is self-supervised pretraining, which learns general neural representations from unlabeled recordings that accumulate across subjects. However, for intracranial electroencephalography (iEEG) recordings, self-supervised learning has been challenging due to differences in contact placement and neuroanatomy between subjects. We propose MAPA, an otherwise vanilla masked autoencoder with two spatial encodings, an anatomical region embedding and a relative positional encoding, which together enable it to learn neural representations that transfer to unseen subjects and across various tasks. MAPA sets a new state of the art across all three regimes of the Neuroprobe benchmark without fine-tuning: within-session, cross-session, and cross-subject. In the cross-subject regime, a linear probe on MAPA's features needs only ${\sim}164$ labeled trials to reach the accuracy that takes 3,500 without pretraining. Our results show that self-supervised pretraining can scale across heterogeneous iEEG recordings and reduce the labeled data needed for accurate decoding in new subjects.

How much learned memory is needed to benefit from more data? We show that the two resources are governed by one predictive-energy spectrum in a positive-entropy autoregressive retrieval source. Each coordinate contributes its query probability times the squared radius of its unknown logit. Writing $\mu$ for the resulting energy spectrum, we prove the minimax law $\mathfrak R^*_{\rm value}(n,B)\asymp_R \Phi_\mu(n^{-1})+\Phi_\mu(\tau_B), \Phi_\mu(t)=\int\min\{x,t\}\,\mu(\mathrm dx),$ for $n$ prediction blocks and a learned state with at most $2^B$ values. Data set the resolution $1/n$; memory sets the level $\tau_B$ reached by optimal bit allocation. The complete curve also recovers the positive spectrum. Energy-dimension pairing is essential: two causal sources with identical block-energy and block-dimension marginals have different data and memory exponents. A masked query-key attention head learns the route and values, realizing the law with explicit routing, format, and arithmetic errors. Further results give exponent-adaptive allocation, finite-precision realization, and compute-precision laws under two-sided arithmetic assumptions. Experiments recover the data-memory collapse and coupling exponents, explain the routing and allocation mechanisms, and examine weight-only quantization across six pretrained-model scales.

Private Evolution (PE) generates high-fidelity synthetic data in federated settings without exposing users' raw data. It aggregates clipped user votes over a shared candidate bank into a differentially private histogram, with noise calibrated to the worst-case user contribution. However, it is unclear whether an adversary can realize this worst-case privacy loss while following the PE protocol. We introduce a protocol-aware empirical audit in which the server commits to a single shared candidate bank and replaces roughly 1% of its entries with probes derived from a known, non-private canary. We evaluate eight attacks, including an unchanged-bank baseline, exact copies, plausible paraphrases, and high-entropy synthetic nonces. Experiments on Yelp and Sentiment140 show that natural-text attacks remain substantially below the theoretical DP bound, while nonce-based attacks yield considerably stronger bounds and come closest to the mechanism's privacy ceiling. These results quantify the gap between formal worst-case privacy and leakage achievable through protocol-valid candidate-bank manipulation.

Predictive Process Monitoring (PPM) forecasts how ongoing organizational processes unfold, enabling information systems to move beyond execution support toward proactive analysis and monitoring. Although deep learning has improved prediction accuracy in PPM, most approaches follow a single-task learning (STL) setup, training a separate model per task. This increases maintenance effort and overlooks potential synergies. Multi-task learning (MTL), which jointly learns multiple prediction targets in one model, offers a promising alternative, yet its effectiveness in PPM remains underexplored. It remains unclear whether and under which settings MTL improves upon STL, which prediction tasks benefit most from joint learning, which task combinations are particularly synergistic, and if and how tasks should be balanced. To fill this gap, we present the first comprehensive empirical study of MTL for PPM, evaluating a variety of task combinations, neural architectures, and optimization methods. Overall, our results position MTL as a strong paradigm for PPM: we see substantial improvements in next-activity prediction and inherent mitigation of class imbalance using MTL, while task balancing is especially critical under low-capacity models.

Understanding how humans respond to incentives, both at the individual and collective levels, is crucial to the design of effective policies. Within the continuous-time stochastic framework for large interacting populations, mean field games (MFGs) model populations of non-cooperative agents, whereas mean field control (MFC) describes the fully cooperative benchmark, interpreted in our setting as fully altruistic behavior. Mixed-individual MFGs interpolate between these two extremes through a parameter governing the degree of altruism. A central challenge for regulators and policymakers, however, is that intrinsic altruism levels and other private structural parameters, such as individual labor costs, are typically unobservable. To address this challenge, we develop an inverse learning framework for mixed-individual MFGs. Our approach enables the recovery of (latent) altruism and labor cost levels from noisy observations, with experiments demonstrating the feasibility and accuracy of our method. These findings underscore the promise of inverse MFG methodologies for uncovering latent preference structures in large populations, with important implications for incentive design, empirical behavioral modeling, and data-driven policy analysis.

The increasing deployment of AI agents in long-horizon tasks yields massive execution logs. Diagnosing failures within these records is crucial for reliability, as it transforms outcome-level signals into actionable interventions. The sheer scale of the data renders human review impractical, driving the need for automated root-cause attribution (RCA). However, automated RCA methods using LLMs suffer from low diagnostic accuracy, especially as execution traces grow larger. They struggle because relevant information is often sparse, distributed across distant actions, and disconnected from the visible failure, reducing root-cause attribution to a massive search problem. Existing RCA methods typically rely on one-shot LLM judgments to diagnose failures from execution traces. While effective for shorter trajectories, these judges tend to settle on a plausible diagnosis early, leaving critical evidence in longer traces unexamined. We introduce Continual Search, an iterative framework that nudges the judge, over successive turns, to keep searching for unresolved diagnostic evidence. We evaluate Continual Search across four existing RCA benchmarks. Recognizing the lack of massive execution traces in current benchmarks, we introduce MegaRCA-Mix to evaluate RCA at scale. MegaRCA-Mix provides a challenging testbed of 50 human-annotated failure trials spanning long-horizon, execution-heavy tasks. Across multiple benchmark suites and model families, Continual Search consistently improves attribution performance. On MegaRCA-Mix, for example, it improves Opus-4.8's F1 score by 29%, from $0.471$ to $0.608$. More interestingly, within the same model family, lower-tier models can even surpass their higher-tier counterparts, demonstrating that effective search supersedes raw model scale.

Earth Observation (EO) data are increasingly organized as spatio-temporal data cubes, while machine learning (ML) methods operate on tabular feature matrices or structured tensor inputs. This mismatch forces platform-specific transformations that are difficult to reproduce or transfer across cloud infrastructures. The openEO specification provides a unified interface for EO data access and processing across heterogeneous backends, but lacks a standardized approach for ML integration. We propose a process-level ML specification for openEO structured into three stages: model initialization, model actions (training, tuning, inference, validation), and model management. It supports classical algorithms such as Random Forest and SVM, as well as deep learning architectures for time-series and spatial patch-based modeling, including TempCNN, Temporal Attention Encoders, and foundation model inference. Three prototype implementations in R and Python demonstrate feasibility across diverse technology stacks. A crop type mapping use case demonstrates cross-backend interoperability by submitting an identical process graph to independent R and Python backends and comparing predictions and evaluation metrics. Two further use cases demonstrate deep learning on time series and foundation model inference, each executed on a dedicated backend. The prototypes reveal, however, that full cross-backend portability requires deeper harmonization of serialization formats and execution semantics than the process level alone can enforce; backend library versions and preprocessing conventions outside the specification's boundary also affect reproducibility. Addressing both through explicit backend conformance profiles represents the most important near-term direction. The specification advances the reproducibility, portability, and accessibility of ML workflows on EO data cubes across cloud platforms.

We demonstrate that training LLMs with RL does not improve performance equally across a dataset. RL shows large improvements on easy problems that an LLM is already good at solving, but small improvements on hard problems. We call this the Matthew Effect in RL for LLMs, after the phenomenon of cumulative advantage from economics and network science summarized as "the rich get richer". The naive explanation is that hard problems require more compute to find a solution. We argue that modern RL methods are exacerbating the issue by wasting too much compute on easy problems and instead should dynamically reallocate how they use compute. We introduce Never Give Up (NGU), a simple adaptive sampling method that keeps generating samples for a problem until one is correct. By leveraging asynchronous RL, this naturally uses fewer samples to filter out easy problems and allocates more compute to solving harder problems. We investigate the design choices that affect NGU, such as off-policy robustness, and develop a set of best practices. On the math benchmark Deepscaler, NGU improves performance per compute, especially on harder problems. On a recent coding task, Manufactoria, standard GRPO with a per-test reward fails to fully solve problems that have a range of easy and difficult tests. NGU iteratively improves, solving harder and harder tests, until it learns to fully solve coding problems.

We give a deterministic algorithm for online inverse linear optimization with regret $O(d)$, uniform in the horizon and $O(d^{2})$ time per round. A bound of this order was obtained recently by Dewasurendra, settling a question of Gollapudi et al.\ and of Oki and Sakaue, but by an improper rule that enumerates covers at every scale and costs $T^{\Theta(d)}$ a round; ours is the first efficient such bound and the first proper one. We build on the variable-metric framework of Sakaue et al., adding a self-normalized rank-one update, and we replace the $\log\det$ potential by the trace power $\tr(H^{-1/2})$, which is bounded outright and removes the $\ln T$. The bound also holds against an expert that does not optimize, and we give corruption-robust and rank-adaptive variants, and an application to convex minimization.

As compute scales, models evolve, and training algorithms advance, our ability to explain the increasingly powerful AI systems they enable is eroding. To help safeguard interpretability, we introduce a specialized training algorithm (MACCHIATO) that jointly constructs (i) an explicitly structured $\operatorname{ReLU}$-MLP from partial truth-table observations and (ii) an explicit Boolean circuit over signed literals with $\{\operatorname{AND},\operatorname{OR},\operatorname{XOR}\}$ gates certifying what its subnetworks compute and how they compose. Intuitively, we iteratively project the residuals of a Boolean function onto low-dimensional $\{\operatorname{AND},\operatorname{OR},\operatorname{XOR}\}$-circuit classes and exactly compile the resulting circuit into a $\operatorname{ReLU}$-MLP; we combine $\operatorname{ReLU}$-MLP circuit compilation, ESPRESSO logic minimization, and influence-based variable selection. Roughly speaking, our interpretability certificate is complemented by a statistical guarantee: under the theorem's influence-recovery conditions, if each of the $m$ stage-wise residuals depends on at most $\log_2(B)$ bits, a sample-splitting variant of our algorithm trained on $T$ observations returns a six-layer $\operatorname{ReLU}$-MLP (counting the input layer) of width $\mathcal{O}(mB)$ with truth-table error $\mathcal{O}\bigl(\sqrt{m(B+\log(m/\delta))/T}\bigr)$. On synthetic random-junta tasks, our networks outperform depth- and hidden-width-matched Adam-trained MLPs in several data-sparse or projection-aligned regimes, while the trained ReLU-MLPs are stronger in others. Moreover, in our explicit PyEDA truth-table implementation, the iterative procedure completes in regimes where flat ambient-dimensional ESPRESSO exceeds the three-hour computational budget.

Training diffusion models with multiple rewards requires distinguishing user preference from reward informativeness. User preference determines how much each reward should contribute to the overall objective; reward informativeness determines when its feedback is useful during denoising. Some rewards can meaningfully evaluate a sample as soon as global structure emerges, but others become informative only when the sample is nearly clean. To address both questions jointly, we propose ReCAST (Reward Credit ASsignment across T}imesteps), the first method, to our knowledge, for per-reward, timestep-dependent credit assignment in diffusion reward fine-tuning. ReCAST separates user preferences from temporal allocation through a reward-by-timestep weight matrix $W$, whose row sums match the user-specified reward budgets $\lambda$, while its column sums are equal, assigning the same total weight to each denoising step. Under these marginal constraints, ReCAST allocates weight according to each reward's informativeness, quantified by its Rényi discriminability gain at each step. These gains telescope to the total discriminability between the reward-induced positive policy and the current policy, providing a basis for temporal credit assignment. We evaluate ReCAST by training SD3.5-Medium under two distinct four-reward settings, each across five reward budgets $\lambda$. ReCAST improves the training rewards in one setting and matches them in the other, improves every held-out judge in both, and is preferred by an independent LLM-as-a-Judge. Together, these results show that ReCAST yields improvements that generalize beyond the training rewards and support its core principle: assigning each reward greater weight at the denoising timesteps where its feedback is most informative.

LLM judges are increasingly used to evaluate and improve AI-generated outputs, yet their reliability for complex professional work remains unclear. We study this problem through Vibe Patenting, an end-to-end patent-drafting testbed for AI-agent evaluation. A separately-invoked LLM judge evaluates generated patent drafts and provides structured feedback for iterative revision. Across multiple inventions and drafting-agent configurations, judge-guided revision consistently improves judge-assessed quality, while unguided revision tends to saturate. Notably, iterative judge feedback enables a low-reasoning agent to approach the performance of a substantially more expensive high-reasoning agent. Stronger models and increased reasoning generally improve judge-assessed drafting quality, while domain-specific agentic workflows provide further gains. We validate the judge against independent evaluation by a professional patent attorney and find meaningful but strongly metric-dependent agreement and systematic calibration differences. These results highlight both the utility and limitations of LLM judges as evaluators and optimization signals for complex professional workflows.

Specifications face a basic tradeoff: leave details out, and important questions go unanswered; record every detail separately, and the specification becomes large and prolix. We investigate whether a language model can serve as a compact, living specification oracle by learning facts about a target and answering questions about it directly. We compare two ways of storing the learned facts: external text notes and changes to the model's weights. Across four families of 596-fact worlds and two Qwen2.5 model sizes, weight-only oracles benefited substantially more from structure: with the 7B model, their accuracy integrated across storage capacities was 18.5 percentage points higher on structured than unstructured worlds, compared with 1.1 points for note-sheet oracles. This advantage came at a substantial storage cost, with the smallest adapter requiring approximately 175 KiB compared with a maximum note budget of 16 KiB. Adapted weights therefore exploited latent structure more successfully, while external notes required substantially less object-specific storage.

When we speak of recursive self-improvement (RSI), are we speaking of a phenomenon, a mechanism, or a prospect? Towards autonomous and evolving intelligence, RSI is being claimed at many scales, while no single framework that formally describes these emerging instances exists. Its counterpart in the classical realm, iterative policy improvement, is characterized by generalized policy iteration (GPI), a framework of broad applicability with well-understood theoretical properties, but only where the update principle and the evaluation base lie outside the agent. In this paper, we propose Generalized Agent Iteration (GAI), a formal framework that describes iterative policy improvement and RSI as two cases of a single learning paradigm. GAI defines the agent as a configuration of modifiable components within a system and models the learning process as a cycle of agent evaluation and agent improvement. Two pivotal dials then distinguish the instances: whether the improving mechanism is part of the agent and whether the standard it is measured against is grounded outside it. The former dial delineates the boundary between GPI and RSI, and the latter determines a system's polarity as anchored, goal drift, or fully self-referential. Moreover, we use these coordinates to place existing systems on the same two axes and make the defects of recursive self-improvement statable one condition at a time. We see this paper as a first step toward exploring a formal characterization of RSI that rests on the classical account, makes existing systems comparable, and provides a principled basis for analyzing and designing new ones.

Mixture-of-Experts (MoE) has become a widely adopted architecture for Large Language Models (LLMs), as it improves model capacity while limiting computational overhead through sparse expert activation. This property makes MoE-based LLMs particularly attractive for resource-constrained distributed environments. However, federated fine-tuning of MoE-based LLMs remains challenging under heterogeneous client data. Since clients often correspond to different task preferences, directly aggregating their local updates may weaken expert specialization and introduce conflicting update directions on shared experts. To address these challenges, we propose FedTAR, a task-aware federated fine-tuning method for MoE-based LLMs. FedTAR establishes the association between local updates and task preference via routing outputs. Specifically, we apply Singular Value Decomposition (SVD) to both routing features and local updates to extract low-dimensional task coordinates and update directions. Based on the task coordinates, FedTAR performs intra-cluster aggregation among clients with similar task preferences and inter-cluster aggregation across different task groups. The aggregated update is then reconstructed through the learned task-to-update mapping, ensuring that the final update remains aligned with task-specific optimization directions. In this way, FedTAR preserves expert specialization and mitigates destructive interference among heterogeneous clients. We evaluate FedTAR on four benchmark tasks under different non-IID settings. Experimental results demonstrate that FedTAR consistently outperforms strong federated fine-tuning baselines and achieves state-of-the-art performance.

Deep joint source-channel coding (DeepJSCC) transmits learned semantic features efficiently but can leak sensitive attributes such as gender, race, or speaker identity. We propose LEAPSC (LEACE-in-the-loop privacy for semantic communication), whose core contribution is the integration of in-loop least-squares concept erasure (LEACE) within a variational information bottleneck (VIB) encoder. By periodically refitting the projection operator during training, LEAPSC couples the encoder dynamics to the erasure mechanism, driving attribute-conditional mean differences toward zero within each task-label group on the fitting sample. Additional components, namely conditional value-at-risk (CVaR) tail-sensitive privacy, feature-wise linear modulation (FiLM) signal-to-noise ratio conditioning, and Lagrangian dual ascent, improve robustness across channel conditions and over the high-leakage tail of samples. On CelebA, FairFace, and Google Speech Commands, LEAPSC reaches task accuracy of 0.862, 0.755, and 0.925 respectively, with attacker accuracy at or below the label-only floor on CelebA (0.548 vs. floor 0.580) and within 2 percentage points (pp) of chance elsewhere, improving over an information-bottleneck adversarial baseline (IBAL) at a matched 52-epoch budget by +3.6, +2.5, and +1.3 pp (Welch's t-test, p=0.019 on CelebA).

We present three robotic art installations which explore the aesthetics of adaptive behavior. Through embodied machine leaning and digital evolution, these works draw viewers into an artificial ecosystem in which open-ended novelty, trial-and-error learning, competition, and cooperation emerge in real time. Research-creation practices are examined in relation to these works, focusing on how they redefine the role of artists within a human-machine collective while examining points of convergence and divergence between artistic and engineering approaches to adaptive robotics. The systems in question use learning and evolutionary processes not as a means to optimize a specific solution, but as an aesthetic experience on its own, suggesting new modes of interdisciplinary art-science research. Finally, we discuss strategies and practices to elevate the aesthetic experience for audiences, including contexts of presentation as well as temporal and material considerations for artworks based on embodied adaptive systems.

Probabilistic forecasting is central to decision-making under uncertainty, yet its methodological landscape has become increasingly fragmented across temporal and spatiotemporal forecasting, statistical modeling, machine learning, and deep generative modeling. This survey develops a unified perspective by organizing probabilistic forecasting methods according to where and how uncertainty is introduced into the forecasting pipeline. Our taxonomy connects model-agnostic approaches including ensembles and distribution-free calibration, with model-intrinsic approaches spanning Bayesian modeling, parametric predictive distributions, distributional regression, and modern generative models, and further examines the emerging role of time series foundation models. Beyond methodological synthesis, we identify the assumptions, computational demands, and forms of uncertainty represented by different paradigms, and translate these distinctions into data-driven and domain-specific guidance for method selection. We complement the survey with a cross-paradigm empirical study on univariate, multivariate, and spatiotemporal forecasting tasks. The results reveal that no single uncertainty-quantification paradigm dominates across settings. Calibration, sharpness, predictive accuracy, and computational efficiency can lead to substantially different model preferences, while expressive generative models and zero-shot foundation models exhibit markedly different accuracy-efficiency trade-offs. Lastly, we identify unresolved challenges surrounding uncertainty in evolving dependency structures, physics-informed predictive distributions, forecasting extreme events, handling count-valued, directional, and continuous-time series, and the development of unified software resources. Our survey provides both a conceptual framework and a practical roadmap for probabilistic forecasting research.

Stochastic gradient descent (SGD) admits diffusion approximations that replace the complicated randomness of stochastic gradients by Gaussian noise, providing a powerful tool for understanding its dynamics and long-time behavior. We investigate whether an analogous approximation principle holds for optimization over probability measures, where the objective is a functional defined on the Wasserstein space P2. The nonlinear geometry and infinite-dimensional nature of P2 prevent a direct extension of the classical Euclidean theory. Using Lions differentiability, we lift the problem to a linear Hilbert space, where higher-order differential calculus becomes available. We then construct a Gaussian random-field approximation whose velocity field matches the mean and covariance of the original stochastic gradient. By exploiting this moment matching through higher-order Taylor expansions, we show that the Gaussian approximation captures the SGD dynamics with second-order weak accuracy. Our result provides a rigorous foundation for replacing sample-driven randomness by analytically tractable Gaussian fluctuations in stochastic optimization over probability measures.