Papers

Effective-connectivity estimation from brain signals often relies on Gaussian residual modeling, which enables tractable estimation but can discard informative distributional structure and distort inferred directed interactions when empirical residuals are non-Gaussian. We show across multiple modalities, species, and experimental conditions that both brain signals and fitted channel residuals frequently deviate from Gaussianity. We therefore introduce a distribution-aware, information-theoretic measure of effective connectivity based on channel capacity under general residual distributions. To estimate the resulting capacity from empirical, potentially non-Gaussian residuals, we develop a dual-flow min-max estimator based on normalizing flows, in which a generator searches over admissible input distributions under a power constraint while an observer estimates output entropy. We provide a theoretical characterization of the estimator, showing that the observer objective recovers differential entropy up to a KL approximation term, that the formulation reduces to classical Gaussian capacity as a special case, and that residual entropy can alter achievable information rates beyond variance; game-gap and error analyses further characterize optimization and approximation sources. In brain-like simulations with known directed connectivity, Dual-flow achieves the highest AUROC and AUPRC across ten conditions spanning diverse network topologies, hidden drivers, feedback, and heterogeneous hemodynamics, compared with Gaussian capacity, Granger causality, VAR-LiNGAM, and GIMME. Applied to multimodal brain signals, the method reveals time- and condition-resolved directed interactions consistent with known neurobiological circuitry. Together, these results establish a principled distribution-aware framework for effective-connectivity estimation beyond Gaussian residual modeling.

Monetary-policy announcements and central-bank communications play a central role in foreign exchange markets, yet their qualitative, unstructured form makes their forecasting value difficult to quantify. While prior research has largely focused on sentiment extracted from financial news, comparatively little is known about the relative contribution of different dimensions of monetary-policy communication. Existing studies primarily evaluate whether textual information improves overall forecasting performance but provide limited insight into which communication channels drive such improvements. To address this gap, this paper introduces a statistical attribution methodology that decomposes monetary-policy communication into interpretable channels and quantifies their incremental forecasting contribution under false-discovery-rate control. Monetary-policy news is transformed into structured communication signals using large language models (LLMs) and temporal feature engineering. These signals are evaluated using rolling-window experiments with tree-based machine-learning models. The results show that monetary-policy communication contains measurable predictive information. Attribution analysis shows that predictive value is concentrated in a small subset of signals, with communication timing providing the strongest individual feature-level contribution, targeted communication-activity measures also contributing positively, and LLM-derived sentiment providing complementary information at the group level. The findings indicate that communication-based forecasting value extends beyond sentiment alone and that attribution, rather than aggregate accuracy alone, is central to evaluating news-derived signals.

Adapting pretrained models to new data can cause catastrophic forgetting of previously learned behavior. When only a few past samples remain, they give continual learning methods sparse and narrow evidence about what to preserve. We show that language models can expand this evidence through self-probing, in which the frozen model generates new inputs from the retained samples and records its own predictions on them. Unlike prior work that replays such data as training examples, our method, CPLUS uses self-probe and past-sample gradients to scale down parameter updates that conflict with prior behavior. Experiments with five language models on four benchmarks show three results. First, the same probes preserve more prior behavior as gradient signals than as replay data. Second, CPLUS learns the new data while consistently reducing forgetting more than existing baselines, especially when past data are scarce, and this protection extends to benchmarks not used for training. Third, we observe that CPLUS also becomes more effective as models grow: within the Qwen3 model family, it recovers an increasing share of the forgetting caused by standard fine-tuning.

Modern machine learning spans supervised, transfer, continual, meta-learning, and related regimes that often reuse the same hypothesis classes, architectures, and optimizers but differ in information access, objectives, memory, adaptation, and sample accounting. This makes it difficult to determine whether one paradigm is genuinely distinct, a special case of another, or part of a broader structural hierarchy. We introduce a protocol-objective-resource (POR) framework that separates representational capacity from these design choices. A paradigm is specified by an environment class, observation protocol, admissible learners, performance functional, and resource accounting rule. POR reductions combine environment embeddings, learner compilers, threshold maps, and calibrated resource overheads. Our main theorem shows that such reductions imply worst-case complexity domination on embedded comparison classes, transferring upper bounds forward and lower bounds backward; under labeled-example accounting, this yields sample complexity domination. We also show that calibrated nontrivial accuracy regimes are necessary to avoid vacuous comparisons, and that strengthening the objective can strictly increase minimax sample complexity even with unchanged protocols and learner classes. Instantiating the framework for supervised, transfer, continual, and meta-learning yields canonical special-case relations: continual contains transfer, transfer contains supervised, and meta-learning contains supervised under aligned raw-example accounting. We further derive a non-exact episode-to-example reduction for episodic meta-learning and capture within-paradigm refinements such as replay memory and task identifiers. The framework thus provides a unified language for structuring learning paradigms and transferring complexity guarantees across them.

Detecting distributed denial-of-service (DDoS) attacks in cloud-integrated IoT networks is difficult when labeled traffic is scarce. Generative semi-supervised learning can supplement the available training data, but prediction shifts induced by synthetic views may affect the targets assigned to real unlabeled flows. We propose AnchorMixGAN, a generative semi-supervised framework that addresses this problem through anchor-aligned target construction. Its Anchor-MAS module treats each real unlabeled flow as an anchor and creates alternative views by replacing one field group at a time with values from generated traffic. A frozen reference classifier predicts the anchor and its views; averaging and sharpening these predictions produces a soft target for the original flow. The flow and its target are then mixed with a labeled example using MixUp, allowing the detector to learn from both the original labeled records and the mixed examples. We analyze how reference-classifier error, view construction, and sharpening affect the target, and derive a bound on the resulting change in cross-entropy at a fixed detector prediction. At the reported 90% training setting with 20% of the training records labeled, AnchorMixGAN attains accuracies of 97.3%, 97.4%, and 96.5% on NSLKDD, BoT-IoT, and CICIoT2023, respectively, exceeding the corresponding MixGAN results by 1.6, 1.0, and 4.4 percentage points.

Visual imitation learning is a promising approach to training robot manipulation policies capable of completing a wide variety of tasks. However, policies today remain brittle to viewpoint perturbations, making deployment in diverse environments a challenge. We present a controlled empirical study of which design choices allow visuomotor policies to generalize across viewpoints. We find that viewpoint generalization improves when dense visual tokens are retained and the action head participates in geometric reasoning. On a suite of simulated tasks that span a wide range of camera poses, we show that these design choices yield a policy that remains performant across viewpoints. As a practical consequence, a policy trained with these design choices also transfers zero-shot from simulation to the real world under random camera configurations.

We introduce the Extender, a log-structured variant of the standard Transformer architecture. In a standard Transformer, each layer communicates with subsequent layers exclusively via the residual $\mathbf{h}$, a superposition channel. The Extender adds a concatenation channel $\mathbf{x}$: each layer $\ell$ emits both a residual update $\delta_\ell$ which is added to $\mathbf{h}$, and a much smaller extension $\epsilon_\ell$ which is appended to $\mathbf{x}$. While both the FFN and $\mathbf{q}$ see $\mathbf{h}$, the attention $\mathbf{kv}$ projections take only $\mathbf{x}$ as input. As a result, the fully extended $\mathbf{x}$ contains the complete input for the $\mathbf{kv}$ projections of all layers, reducing the persistent attention memory footprint from $2Ld_{model}$ to $\sum|\epsilon_\ell|$. We find that with $|\epsilon_\ell|=32$, the Extender matches Transformer accuracy on short-context (CORE) tasks at 199M-924M parameters, and exceeds Transformer accuracy on long-context (RULER) workloads, again at 924M parameters. For our 1664-wide, 924M model, the Extender's persistent attention memory footprint is $104\times$ smaller than MHA. The memory savings grow with model width.

VLM bounding-box localization is both language generation and spatial commitment. Parseable fields such as bbox_2d make localization easy to score, but dimensions that emit coordinate tokens need not repair localization after visual evidence is damaged. We study this readout/recovery separation in Qwen3-VL-4B-Instruct on single-object COCO grounding. We compare clean coordinate-token readout rankings with corruption-derived repair rankings, using object-mask endpoint replacement for recovery and clean-input flooring for depth localization. In Qwen3-VL, coordinate-token rankings are inert through layer 24, load-bearing from layers 32-35, and peak at layer 34; corruption-derived rankings harm layers 16-24 but become beneficial near layer 35/final. A Kimi-VL-A3B diagnostic shows a matching output-proximal transition despite a different box format. Object-mask recovery separates rank budgets: $k=250$ shows necessity, $k=500$ shows Top-$k$ restoration above random, and $k=d/2$ is largely capacity-driven. Partial-occlusion sweeps reveal that high-overlap coordinate-token sets can hurt at $k=1000$ and help mainly at half-width, while population corruption-derived sets provide no reliable fixed repair set. Edge-attribution patching shows coordinate-token paths are high precision but low recall for detection recovery, and RMSNorm quasi-layer controls do not close the endpoint-repair gap. Endpoint coordinate triage is therefore a useful circuit prior, but occlusion recovery requires a separate benchmark.

Foundation models are rarely used as generic, frozen feature extractors; instead, they are fine-tuned for the target downstream application. This practice is particularly prevalent in Earth observation (EO), and it raises a question that downstream accuracy alone cannot answer: does fine-tuning reuse the pretrained representation, or does it relearn a new one? We study this with spectral diagnostics that compare a model before and after adaptation, quantifying how well its dominant singular subspaces are preserved, how broadly the weight update is distributed, and how large it is. Using natural image models such as CLIP and DINO as a reference, we find that, under the evaluated fine-tuning settings, EO models undergo far larger, higher-rank updates and retain much less of their pretrained structure, so their downstream performance is often obtained with substantial changes to the pretrained weight structure. The diagnostics further provide insight into how cheaply a model can be adapted: where the pretrained subspaces are preserved, adapting a small fraction of the parameters can match full fine-tuning, and where they are not, it can fall behind. More broadly, foundation models, and EO foundation models in particular, should be assessed not only by benchmark accuracy, but also by how reusable their pretrained representation is.

Audio autoencoders compress waveforms into compact latent representations that serve as the interface between raw audio and downstream models. Current systems navigate a three-way trade-off between reconstruction quality, semantic structure of the latent space, and inference speed, typically favoring one or two of these at the expense of the others. This paper introduces SAGE, Semantic Audio Generative Encoder: a compact variational autoencoder, trained solely on publicly available music, that shapes its latent by distilling embeddings from a pretrained audio-text model. This 105M-parameter model runs at the inference cost of Stable Audio Open and reaches the listening-test quality of SAME-L, an autoencoder 8x larger and 4x slower, while surpassing both on objective perceptual and distributional metrics of reconstruction. Furthermore, it sets the state of the art on all nineteen probing tasks of latent semantics, in domain and out of domain. These results establish SAGE as a lightweight audio autoencoder that strikes the best balance of the three-way trade-off among those we evaluate, combining high reconstruction fidelity, state-of-the-art semantic structure, and fast inference.

Reward shaping has a theorem: a potential-based term can be removed without changing the optimal policy. The same practice on the action channel, an offset added in training and dropped at deployment, has no theorem. Nothing cancels an action offset, so the correction is kept at deployment or removed without a guarantee. We call it action shaping and state its principle. A trainable policy absorbs an offset its own output layer can reproduce exactly, which is what we mean by express; what is absorbed can be removed with the return intact. Its minimal instance is a zero-initialized linear head behind a learnable gate, added to an actor that trains through a learned action-value function, with no penalty or schedule. The gate rises and then falls on its own, for deterministic and stochastic actors alike, and on 20 tasks removing the head costs almost nothing. The condition is exact reproduction, not capacity: a nonlinear head with more parameters is not absorbed, and in a paired control, one linear path added to a nonlinear base head restores absorption. Exact reproduction gives the loss a flat direction that gradient noise drifts along, and the offset's amplitude indicates, before removal, what dropping the head will cost. Action shaping thus gains the counterpart of the shaping theorem, a condition for absorption, together with the mechanism behind it and a diagnostic that reads it. Policies absorb what they can express, and only that.

Model distillation transfers capabilities through supervised fine-tuning (SFT) on teacher responses, often collected from commercial APIs, raising questions of model provenance. Existing distillation attribution methods have been largely evaluated on students immediately after the SFT step. However, a distilled model may undergo further SFT, preference optimization, or reinforcement learning before release, while an auditor may lack access to the pre-distillation checkpoint required by reference-based attribution. To close this gap, we propose SCOUT, an output-only method that aggregates recurring *syntactic patterns* into candidate profiles, filters low-contrast patterns, and calibrates student–candidate distances against inter-candidate distances. SCOUT supports attribution and abstention using only current texts, without model weights, token likelihoods, or historical checkpoints. Auditing publicly released descendants of distilled models spanning diverse post-training objectives, SCOUT consistently identifies the distillation source. Furthermore, tracing teacher-associated *syntactic signatures* along training trajectories reveals that they emerge during distillation and persist through subsequent preference optimization and reinforcement learning.

While symbolic regression (SR) has been successfully used in science to discover new equations, its use in financial valuation is hindered by several limitations. Whereas the natural sciences provide objectively correct relationships, financial valuation constitutes a distinct class of symbolic discovery problems, as it admits multiple valid perspectives, operates under non-stationary market conditions, and involves noisy, continuous performance signals. In this work, we propose Multi-Agent Fundamental Analysis with Symbolic Adaptive learning (MUFASA), a hierarchical multi-agent framework for symbolic discovery in finance. MUFASA introduces (1) disentangled equation discovery via specialized agents representing distinct valuation perspectives, (2) a meta-coordinator that performs hierarchical-level reasoning over market context information, and (3) a memory mechanism that reasons over statistical performance summaries (e.g., accuracy, stability, and tail risk) to guide learning under noisy feedback. Experiments across datasets from multiple countries show that MUFASA achieves state-of-the-art performance on the valuation task compared to classical finance methods, financial large language models, and SR approaches, while simultaneously producing interpretable equations, which we share with the community. We also make publicly available the distilled learnings across evolution iterations and context-dependent strategy weights, which might offer useful insights for future research on financial fundamental analysis.

Electroencephalography (EEG) foundation models (FMs) promise transferable neural representations, yet their advantages over strong supervised baselines and their prospects for further scaling remain unclear. To address these questions, we introduce EEG-Arena, an open-source benchmark covering 30 EEG FMs and 25 supervised baselines evaluated on 57 downstream tasks from 23 public datasets. Through more than 20,000 evaluations across five experimental protocols, we assess downstream performance, pretraining benefits, model size scaling, pretraining data scaling, and robustness to channel configuration. We find that (1) EEG FMs outperform strong task-specific supervised baselines on most evaluated tasks, particularly under non-bipolar settings; (2) compared with architecture-matched supervised training from scratch, pretraining improves both early optimization and final downstream performance, with larger and more consistent gains as more labeled downstream data become available; (3) existing EEG FMs do not exhibit a consistent positive relationship between parameter count and downstream performance; (4) under a fixed architecture, increasing the pretraining data scale yields sustained downstream gains; and (5) channel-flexible FMs achieve higher absolute performance than channel-constrained models across most evaluated channel configurations. Together, these findings demonstrate the downstream value of EEG FMs and identify pretraining data expansion as a promising direction for further progress. To support continued research, we release EEG-Arena as an open-source evaluation framework that provides shared infrastructure for reproducible benchmarking, model comparison, and community-driven development.

Modern semiconductor production relies on a globally distributed, multi-tier supply chain in which financial stress at one firm spreads with a delay and eventually affects the revenue, inventory, and profitability of the companies that design AI chips. Most firms see only their direct partners, and prior predictive research has mainly targeted market-based risk measures, so few tools forecast how supply chain stress will appear in reported financials. In this study, we propose a heterogeneous graph patch transformer that forecasts these quarterly changes one and two quarters ahead. Learning from a 15,186-company network over 60 quarters, the proposed model fuses quarterly fundamentals with macro-trade, event, and disaster signals through learned gates, carries risk across supplier, customer, ownership, and headquarters relations through typed, direction-specific propagation, and encodes the propagated histories with patch-based tokenization. In preliminary experiments on 116 focal semiconductor firms, the proposed model achieves the lowest error on every target at both horizons, and its profitability advantage widens at the two-quarter horizon. These forecasts can help supply chain managers and investors act before disruptions appear in reported financials.

Existing geospatial vision-language models (Geo-VLMs) typically optimize diverse geospatial tasks through a unified multi-task adaptation paradigm without explicitly accounting for the heterogeneous optimization characteristics. Our empirical observations reveal heterogeneous gradient characteristics across tasks, including vision-language differences, intra-branch gradient relationships, and task interference, which hinder effective multi-task optimization. Motivated by these observations, we propose Gradient-Guided Decoupled Adaptation (G2DA), a gradient-aware optimization framework for multi-task Geo-VLM learning. G2DA first partitions tasks into vision- and language-centric groups through gradient-guided cross-modal decoupling. It then constructs modality-specific curricula based on task gradient similarity and employs bidirectional rehearsal to mitigate the recency effects introduced by sequential optimization. We evaluate G2DA on three Geo-VLM benchmarks using six InternVL3 and Qwen3.5-VL variants, along with GeoChat and GeoLLaVA. Across all 24 benchmark-model combinations, G2DA consistently outperforms representative baselines, improving over the strongest competitor by 3.08, 4.30, and 2.81 percentage points on UrBench-MCQ, XLRS-Bench-Lite, and VRS-Bench-VQA, respectively. These results demonstrate the effectiveness of gradient-guided task organization for Geo-VLM adaptation.

Human mobility generation aims to synthesize realistic point-of-interest (POI) visitation trajectories and has become an important tool for travel behavior modeling, transportation management, and urban planning. Existing diffusion-based methods achieve high fidelity but require per-city generation, given the inherent heterogeneity of geospatial locations and POI categories, while large language model-based methods generalize across cities but remain too costly at scale, especially for long-horizon trajectory generation. To address this, we propose SeMoFlow, a Semantic human Mobility generation framework based on latent Flow matching. We first encode heterogeneous POIs from different cities into a shared cross-city representation space via hierarchical Semantic IDs, where shared prefixes capture transferable semantics, and successive codes progressively refine the representation toward individual POIs. Building on the semantic IDs, SeMoFlow follows a plan-to-synthesis hierarchical generation paradigm, in which an autoregressive planner generates coarse-grained semantic and recurrence patterns, and a flow matching realizer synthesizes fine-grained suffix latents. The generated latents are subsequently decoded and grounded to concrete POIs. Extensive experiments on large-scale multi-city datasets show that SeMoFlow achieves higher trajectory fidelity than existing baselines, preserves city-specific mobility motifs, and supports both joint multi-city generation and effective cross-city transfer.

Domain Adaptation (DA) often lacks worst-case guarantees, while Distributionally Robust Optimization (DRO) based only on source data centers its ambiguity set at the source law and ignores available target structure. To bridge this gap, we introduce a doubly-anchored DRO framework whose ambiguity set is the intersection of $\phi$-divergence balls centered at the source law and a source-completed target reference law, the latter pairing the target covariate law with the source conditional law. We derive dual-induced adversarial bridge geometries for symmetric and asymmetric divergence pairings, notably introducing a Kullback–Leibler/squared-Hellinger (KL/HD) bridge. This asymmetric formulation yields a Lambert-$W$ geometry in which source-side exponential risk tilting and target-side Hellinger stabilization enter through distinct terms, attenuating, but not bounding, the effect of large likelihood ratios. Furthermore, without imposing covariate shift, we establish finite-sample generalization bounds for the minimizer of a structural, loss-agnostic density-bridge risk under bounded-overlap conditions; these bounds do not apply directly to the loss-aware DRO min–max estimator. We translate our framework into a bridge-weighted Nadaraya–Watson estimator, proving uniform consistency for the source regression function and pointwise asymptotic normality, with target recovery when the source and target regression functions coincide, as under covariate shift. Finally, an empirical evaluation on a domain-shifted Fashion-MNIST dataset illustrates the finite-sample stability of the asymmetric KL/HD bridge under severe synthetic target-covariate corruption.

AdaGrad and AdaGrad-Norm are widely used adaptive methods, but their precise behavior in stochastic convex optimization remains less understood. We first show that AdaGrad-Norm and AdaGrad do not admit any universal last-iterate rate, even under sub-Gaussian noise and bounded iterates. We then prove that bounded variance alone is too weak: even with bounded iterates, it cannot yield high-probability average-iterate rates, and without bounded iterates it may not even guarantee convergence in expectation. Besides, we construct tight $\Omega(\log T/\sqrt T)$ lower bounds for both AdaGrad-Norm and AdaGrad under sub-Gaussian noise, showing that $\log T$ in existing average-iterate upper bounds is unavoidable. Finally, we show that this $\log T$ loss disappears once bounded-iterate condition is imposed: under a general ABC condition, both methods achieve rates ${O}(1/\sqrt{T})$.

We study distributionally robust reinforcement learning (DR-RL) in the average-reward setting under weak communication. Our main result provides finite-sample guarantees for estimating the robust optimal average reward and learning a near-optimal policy, covering both SA-rectangular and S-rectangular structures with divergence-based and distance-based uncertainty sets. Specifically, for Kullback–Leibler and $f_k$-divergence balls, we establish explicit radius conditions under which the robust average-reward Bellman equation admits a constant-gain solution, while for total variation and Wasserstein balls, any positive radius suffices without requiring the nominal MDP to be weakly communicating. Our algorithm is prior-knowledge-free and achieves sample complexities of $\widetilde O(|\mathcal{S}||\mathcal{A}|p_{\wedge}^{-1}\operatorname{Span}^{2}(u_{\delta}^{\ast})\epsilon^{-2})$ for estimating the robust optimal average reward and $\widetilde O(|\mathcal{S}||\mathcal{A}|p_{\wedge}^{-2}\operatorname{Span}^{2}(u_{\delta}^{\ast})\epsilon^{-2})$ for learning an $\epsilon$-optimal policy. Here, $p_{\wedge}$ is the smallest positive nominal transition probability and $u_{\delta}^{\ast}$ is a robust optimal bias function. We further provide an almost-tight explicit upper bound on $\operatorname{Span}(u_{\delta}^{\ast})$. Finally, we validate the predicted $n^{-1/2}$ convergence rate through numerical experiments.

*Reinforcement learning (RL)* can induce substantial reasoning capabilities in large language models (LLMs), but how much of this capability transfers across model scales, and how quickly, remains unclear. We study the scaling properties of *on-policy distillation (OPD)* across *weak-to-strong*, *same-base*, and *strong-to-weak* teacher–student setups. We find that early OPD training dynamics uniformly exhibit a regular *useful-transfer* regime, in which held-out accuracy (the *gold score*, $G$) rises approximately linearly in $d=\sqrt{\mathrm{KL}(\pi_\theta \Vert \pi_{\mathrm{ref}})}$, the square root of token-level reverse KL divergence from the student initialization. In every observed weak-to-strong pair, the student's peak gold score exceeds its teacher's own, so a compact RL expert can transfer capability to a much larger student via OPD. To estimate OPD outcomes, we fit *power laws* for how $G_{\mathrm{peak}}$ and the slope of the useful-transfer regime scale with student and teacher parameter counts and with teacher gold score. These laws show that peak gold score improves with teacher scale only up to roughly the student's scale, and that at a matched gold score smaller teachers transfer better, so a teacher's score alone does not define its supervision value. We also study the scaling effects of two OPD variants, bootstrapping weak-to-strong OPD, and the degree of on-policy supervision.

Reward models score responses from large language models (LLMs) and guide LLM training toward human preferences. However, reward models can favor superficial attributes such as length or confidence, leading LLMs to produce higher-scoring but not more correct responses. Existing mitigation methods either retrain the reward model or apply a fixed correction to one known bias, such as a preference for longer responses. Retraining requires additional data and computational resources, while existing editing methods require the target bias to be specified in advance and use a fixed edit for that bias. To this end, we propose BiasReducer, a lightweight framework that edits only the linear reward head and selects the relevant edits for each new dataset. First, BiasReducer uses a sparse autoencoder (SAE)-style encoder to learn which attributes (e.g., length and confidence) the reward model is sensitive to. Second, it learns how to reduce the reward model's dependence on each attribute by determining which direction to adjust the reward head and how much to adjust it. Third, for a new dataset, it ranks the attributes by their influence on reward scores, selects the relevant ones, and edits the reward model accordingly. BiasReducer consistently improves reward-model robustness to biases toward superficial response attributes. Across five reward models, BiasReducer-M improves the three benchmarks by 8.3, 18.0, and 6.9 percentage points on average, outperforming the two training-based baselines. The gains transfer downstream, reducing unnecessary verbosity and sycophancy while maintaining comparable judged quality.

Drifting models enable one-step generation by amortizing distribution transport into training, but this efficiency places greater demands on the transport field estimated at each update. In large-scale training, the field is computed from finite mini-batches of generated and real samples, which provide only imperfect approximations to the underlying distributions. Balanced optimal transport enforces exact mass matching within every mini-batch, making the estimated field sensitive to the particular composition of the real-data batch. We find that generated and real samples should be treated asymmetrically: letting the mass assigned to real samples adapt while keeping every generated sample fully transported improves generation across six feature-space metrics in controlled ablations, and is more robust to the relaxation strength than relaxing both marginals simultaneously, which falls below balanced transport under stronger relaxation. Motivated by this observation, we propose Unbalanced Optimal Transport Gradient Flow (UOT-GF), which keeps the generated-sample marginal fixed and relaxes only the real-data marginal. Under identical settings at DiT-B/2 on ImageNet-256, UOT-GF improves Fréchet Inception Distance (FID) from 1.53 to 1.46 over the balanced W-Flow baseline; scaling the same recipe yields 1.34 and 1.22 FID at L/2 and XL/2, the best FID among the one-step models we compare. We further derive the induced UOT transport force, establish a kinetic Vlasov–Fokker–Planck formulation whose overdamped zero-temperature limit recovers the drifting dynamics, and characterize non-target stationary states together with sufficient conditions for convergence.

Data shape determines how features are structured, how patterns are separated, and how distributions cover the underlying domain. Poor data shape can make models learn noise rather than generalizable structure. This paper studies robust feature-centric data reshaping: generating feature transformations that remain useful, stable, and reproducible under noisy evaluation and imperfect data conditions. We view reshaping operation sequence search as reward-guided diffusion generation, and robust reshaping as searching for regions in the latent reward landscape rather than isolated high-reward transformations. The key challenge is dual instability: noisy evaluators distort local reward guidance, while stochastic generative trajectories can converge to inconsistent solutions. We propose FCDiff, a flat-consensus diffusion framework that addresses both failures through a micro-macro decomposition. The micro layer replaces point-estimate reward guidance with Gaussian-smoothed, Monte Carlo averaged gradients, steering generation toward locally flat reward regions. The macro layer aggregates independently guided trajectories with a weighted Frechet-mean barycenter, selecting consensus-supported basins and filtering stochastic outliers. Across an 8-dataset headline cohort under heavy-tailed evaluator noise, FCDiff attains the best aggregate rank on lower-tail reliability and robustness against both search-based AutoFE and robustness-oriented generative baselines, with statistically significant accuracy gains over every generative baseline. Our results show that robust data reshaping requires searching for flat, consensus-supported regions rather than sharp single-trajectory optima.

We study how the affine geometry of finite ReLU networks is created at random initialization and reorganized by supervised training. We call a sign-changing zero of a hidden preactivation an activation switch and a point where the scalar network output is nondifferentiable a scalar kink. For one-dimensional input, conditioning on the preceding layers makes each preactivation Gaussian and affine on the cells of a random finite partition. This yields an exact finite-width conditional Kac-Rice formula for the expected number of activation switches along an input interval. For fixed depth and proportionally growing widths, the resulting switch intensities converge to explicit deterministic limits. A visibility estimate shows that the expected number of switches that do not produce scalar kinks is negligible, yielding an explicit leading formula for the expected number of scalar kinks and hence affine regions. In higher input dimensions d >= 2, the analogous conditional surface formula yields the leading expected (d-1)-dimensional Hausdorff measure of the scalar kink set. On the Breast Cancer Wisconsin data, the initialization formula accurately predicts switch counts along held-out segments. After training, switch counts decrease along within-class segments and increase along between-class segments. Thus, training redistributes rather than merely contracts affine complexity.

LLM agents that invoke privileged tools are vulnerable to indirect prompt injection (IPI), in which adversarial instructions embedded in retrieved data hijack the agent's actions. A growing body of work evaluates defenses against IPI, but the validity of that evaluation is rarely examined. We audit an IPI benchmark and its harness and identify four defect classes – silent payload non-delivery, attack success scored by tool identity rather than arguments, false-rejection rate conflated with model incapacity, and the absence of an audit trail – each of which yields a plausible, publishable, and incorrect number. We quantify the distortion by re-scoring identical execution traces under the defective and corrected definitions: on real agent behaviour, the tool-identity scorer reports a 21.7% attack-success rate where the true argument-level rate is 1.2%. In the sharpest case, an open model previously reported at 62.8% registers 0% under the corrected harness – the prior figure largely an artifact of undelivered payloads and identity-level scoring. We release a harness whose construction makes each defect unrepresentable – machine-checkable payload placement, argument-level attacker predicates, per-scenario environments, and mandatory trace persistence – and use it to report three quantities the field does not: whether a compromised agent discloses the attack, the full security/utility operating curve of an LLM-judge defense, and tool-calling capability disentangled from defensive over-blocking. A corrected harness further overturns a reported "capability barrier": a model deemed incapable of tool use is in fact fully capable, its earlier result an artifact of environment mismatch. We argue that evaluation validity is a prerequisite for, not a footnote to, defense claims in agentic security, and provide an instrument that enforces it.

LLM-based agents are increasingly used for time-series forecasting because they can organise contextual information, perform multi-step analysis, and guide the sequence of actions required to complete forecasting tasks. Most existing agents focus only on the current forecasting instance. However, in real-world deployments, forecasting commonly operates online, with new forecasts issued from the currently available history as the forecast origin advances and the ground-truth targets of earlier instances progressively become available. These targets provide feedback on the actions taken in earlier instances, yet existing agents generally do not preserve or utilise this information to adapt their subsequent actions. To address this limitation, we introduce FASE, a Feedback-Aware Self-Evolving forecasting agent that converts such feedback into task-specific experience for subsequent forecasting instances. FASE combines episodic memory, which retrieves relevant completed instances, with online policy learning, which summarises the feedback accumulated across instances into ranking guidance. The proposed framework is evaluated on 29 dataset configurations selected from the GIFT-Eval benchmark. Across these 29 configurations, FASE attains the strongest aggregate point forecasting performance among the evaluated methods and reduces the normalised MAE by 9.1% relative to the best individual foundation model baseline. The results further indicate that the cumulative advantage of FASE increases as delayed feedback accumulates. Together, these findings demonstrate that FASE can continually self-evolve through feedback from completed forecasting instances without updating the parameters of the LLM.

Physics-informed neural networks (PINNs) provide a learning-based framework for solving partial differential equations (PDEs), yet their training behavior can change substantially throughout optimization. Residual distributions, gradient interactions, regional learning difficulty, and model-capacity requirements may evolve over time, while the network architecture and major training mechanisms are typically determined before training. We propose PINNMorph, an online PINN adaptation framework based on large language model (LLM)-guided policy evolution. PINNMorph maintains a population of state-conditioned adaptation policies that map execution diagnostics to controlled interventions over topology modification, additive representation augmentation, objective balancing, gradient handling, adaptive sampling, and optimizer-phase control. At each intervention opportunity, candidate programs are instantiated from the current policy population, selected according to the observed training state, and applied directly to the PINN under training. The resulting model inherits its existing parameters and training state and continues optimization along the same trajectory. Execution outcomes are subsequently used to evaluate interventions and evolve the policy population. Unlike pre-training architecture search or fixed adaptation rules, PINNMorph jointly adapts the current PINN and the policies governing its interventions using feedback from actual training. Experiments on 13 PDE benchmarks show that PINNMorph achieves lower solution errors than SA-PINN, ConFIG, RoPINN, HARMONIC, and PINNsAgent across all evaluated problems. Ablation studies further examine the effects of online adaptation, state-conditioned intervention selection, and execution-feedback-driven policy evolution.

GUI World Models (GUI-WMs) are increasingly used to predict future states for agent planning and simulation, yet most existing formulations condition only on the current GUI observation and action. We identify state aliasing, where the vis- ible interface omits transition-relevant environment state, so identical observable conditions can correspond to different valid futures. To diagnose this failure mode, we introduce StateAliasBench, a diagnostic benchmark that explicitly isolates such ambiguities via strict pairing. We further propose lightweight predictive- state recovery that infers structured state from history and augments otherwise frozen GUI-WMs through a deterministic state interface. Family-specific special- ists provide state recovery across heterogeneous state types, and multi-teacher dis- tillation consolidates them into a single unified estimator. Experiments show that existing GUI-WMs exhibit systematic failures under observation-only condition- ing, while predictive-state augmentation substantially restores state-sensitive pre- diction across evaluated WMs, preserves generative fidelity, and improves down- stream performance of GUI agents on AndroidWorld. These results suggest that reliable GUI world modeling should account not only for what is visible, but also for the hidden transition state that determines what happens next.

Self-improving agents increasingly rely on proxy verifiers to choose policy updates, yet deployment can change the world in which those updates are evaluated. An update that looks better to the verifier can therefore become worse after deployment even when the verifier ranks policies well overall. We formalize this gap as Improvement Fidelity, which asks whether proxy improvements preserve the sign and ordering of deployment improvements over the updates an improvement process actually proposes. We show that global policy accuracy need not guarantee update fidelity: operator shift and deployment response can create update-level errors, while candidate margins determine whether those errors change the replacement decision. We introduce PIVOT-KG, a paired, decision-aware validator that allocates scarce high-fidelity evaluation according to the expected reduction in selection regret per unit cost. Across 90 held-out roots in Leduc, Kuhn, and Melting Pot, proxy and deployment optimal sets are disjoint in 51 cases. In an eight-candidate HighwayEnv stress test, PIVOT-KG reduces mean improvement-selection regret from 0.0435 under the exact Uniform validation rule to 0.0055 at the primary budget. Together, these results show why reliable self-improvement should evaluate proposed improvements in the worlds they induce, while providing a practical rule for allocating scarce deployment evidence when it can affect the replacement decision.