Papers

Muon-style optimizers take a matrix-valued momentum or preconditioned update $B = U \operatorname{diag}(\sigma_1,\ldots,\sigma_r) V^\top$ and replace it with its canonical partial polar factor $\operatorname{Pol}(B) = U V^\top$. This maps every nonzero singular value to one. MuCon is the clipped-Muon variant studied here: it applies singular-value clipping to the same Muon matrix, $D^{\mathrm{MuCon}}\_\tau(B) = \operatorname{MClip}\_\tau(B) = U \operatorname{diag}\bigl(\min\{\sigma\_i,\tau\}\bigr) V^\top, \qquad \tau > 0$. Thus, $\operatorname{MClip}\_\tau$ denotes the mathematical clipping operator, while MuCon denotes the optimizer primitive that substitutes this clipped direction for Muon's polar direction. The Muon/MuCon scaling parameterization used in this work is called $\text{SpectralP}$: it is the hidden-matrix scaling recipe under which polar Muon or clipped MuCon directions are applied. The map $\operatorname{MClip}\_\tau$ is the Frobenius projection onto the spectral-norm ball $\{X : \|X\|_2 \le \tau\}$: it leaves singular values at or below $\tau$ unchanged and modifies only the violating singular directions. This paper asks when the MuCon clipping step can be approximated without a full dense SVD. We record two exact identities, a polar/absolute-value formula and a scalar-root formulation leading to a rational Newton filter for the clipped positive-semidefinite factor, and identify the numerical obstruction common to both: singular values near the threshold make sign decisions and rational solves ill-conditioned. Matrix-function methods are therefore useful only when paired with stable polar/square-root primitives or explicit regularization near the clipping boundary.

Safe reinforcement learning (RL) for robotic systems requires policies that improve task performance while satisfying state and input constraints during both training and deployment. Control barrier functions (CBFs) provide a principled mechanism for enforcing forward invariance through minimally invasive safety filters, but their use in model-free RL is limited by the need for accurate dynamics and hand-designed barrier certificates. We propose Robust Koopman-CBF SAC, a safety-filtered actor–critic framework that learns a finite-dimensional Koopman predictor from data, constructs affine CBF constraints in the lifted space, and enforces them through a quadratic-program safety layer. To account for finite-dimensional Koopman approximation error, the CBF condition is tightened using a projected residual margin estimated from held-out rollout data. The critic is trained on the executed safe action, while the actor is regularized toward the Koopman-CBF feasible set, reducing dependence on the filter over training. Across safe-control benchmarks, the method achieves zero constraint violations on CartPole stabilization and tracking while matching or exceeding unconstrained SAC returns. On high-dimensional Safety Gymnasium locomotion tasks, the method reduces violations in some settings but also exposes important limitations of first-order velocity barriers and linear EDMD models, motivating high-order and multi-step Koopman-CBF extensions. These results suggest that robust Koopman-CBF filters are a promising bridge between model-free RL and certifiable safety, while clarifying the structural conditions under which such filters remain effective.

Graph anomaly detection (GAD) commonly relies on message passing to jointly encode node attributes and neighborhood context. However, once the two are mixed, an abnormal post-encoding state may reflect either an intrinsic node deviation or incompatible contextual influence, making its source ambiguous. We propose Disagreement-Driven Graph Anomaly Detection (DDGAD), which treats persistent incompatibility between node-wise and contextual estimates as anomaly evidence. Inspired by Adapt-Then-Combine (ATC), DDGAD reverses its consensus objective: Adapt produces a node-wise estimate without new same-step neighbor aggregation, while Combine forms a neighborhood-dependent contextual estimate, and their pre-consensus disagreement is accumulated across ATC steps for detection. We further characterize this signal from graph-spectral and source-response perspectives and derive sufficient conditions for anomaly–normal separation under contextual mixing. Experiments on six benchmarks show the highest average AUROC among the evaluated methods, while controlled interventions and Adapt-operator controls further support persistent disagreement as an effective detection signal.

EEG foundation models, pre-trained on large-scale unlabelled EEG data, have emerged as a promising direction towards learning generalizable EEG representations. Despite showing positive results in data-rich regimes, they often fail to outperform significantly smaller supervised models in low-resource settings compared to fully supervised models. We provide a mechanistic account of this shortcoming, attributing it to a fundamental mismatch between reconstruction-based pretext tasks and the idiosyncratic spectral structure of EEG signals, which decompose into distinct high-power aperiodic and low-power oscillatory components. Using controlled, synthetically-generated EEG inputs, we demonstrate that EEG foundation model embeddings are biased to capture the aperiodic components of the EEG signal while under-representing oscillatory components, particularly at higher frequencies. Additionally, linear probe evaluations on real-world BCI datasets further reveal that embeddings encode subject identity more strongly than task-relevant information, thereby reinforcing the low-frequency and aperiodic component bias in foundation model embeddings trained primarily on reconstruction based objectives. Together, these findings elucidate a failure mode in reconstruction based EEG foundation models and motivate future work to incorporate auxiliary losses explicitly targeting high-frequency oscillatory structure as a path toward more capable and generalizable EEG representations.

This paper addresses structured out-of-distribution (OOD) testing in high-stakes machine learning applications. Traditional conformal methods rely on joint exchangeability, making it difficult to incorporate auxiliary information such as spatiotemporal or grouping structures. To overcome this limitation, we propose the structure-adaptive conformal q-value (SCQ), a significance index that integrates individual test evidence with structural patterns. We also develop pseudo-score-guided transductive automated model selection (P-TAMS), which adapts conformalized model selection to structured OOD testing across a toolbox of candidate models. Together, SCQ and P-TAMS form a unified framework under pairwise exchangeability, providing finite-sample error-rate control, improved power, and enhanced interpretability. Experiments on simulated and real data demonstrate that the proposed approach controls the false discovery rate and performs well across diverse settings.

With the rapid evolution of internet services, recommendation systems have become indispensable. In particular, the blending (re-ranking) stage plays a pivotal role in allocating traffic across diverse business objectives. However, existing approaches often suffer from coupled allocation plans, score inflation, and a lack of interpretability. To address these challenges, we propose Uniboost, a unified traffic allocation framework. Uniboost introduces a posterior value alignment mechanism that calibrates abstract model scores to anchor metrics with explicit business semantics, significantly enhancing interpretability. Furthermore, it employs an independent linear boosting paradigm to decouple complex weighting schemes, enabling precise attribution of each plan's contribution. We validate the effectiveness of Uniboost through online A/B tests and in-depth data analysis, demonstrating three key findings: 1) Reducing the overall weight of weighted scores effectively mitigates unintended business interference, yielding a more efficient micro-level traffic allocation strategy; 2) Post-hoc analyses and aggregated dashboards provide intuitive, macro-level insights that guide the design of the overall traffic allocation mechanism; 3) The proposed "Effective Completion Score" serves as an easily obtainable post-metric that offers a reliable anchor for content recommendation pipelines. Collectively, our experiments show that Uniboost not only improves traffic allocation efficiency and recommendation performance at the micro level but also provides macro-level guidance for system iteration. Thus, this work provides an efficient and controllable traffic regulation solution for large-scale industrial recommendation systems.

Task-based fMRI provides a direct readout of task-evoked neural dynamics, but it is expensive and difficult to acquire at scale, motivating rest-to-task synthesis from widely available resting-state fMRI (rsfMRI). We propose FM-fMRI, an event-conditioned flow-matching model that learns a continuous-time conditional vector field to generate task ROI time series from a subject's rsfMRI and the task event information. The formulation enables fast ODE-based sampling and flexible conditioning over heterogeneous event schedules. Rather than optimizing for pointwise reconstruction, we evaluated generated signals using complementary criteria that probe temporal and spectral structure, subject and group-level connectome consistency, and distributional alignment. On the public Human Connectome Project and internal BioPoint autism cohort, FM-fMRI achieves the strongest spectral and connectivity agreement and improved distribution-level matching over conditional diffusion, generative adversarial networks (GANs), and variational autoencoders (VAEs) baselines. Furthermore, we augment the BioPoint cohort by synthesizing task-fMRI ROI time series with our method, improving downstream autism classification and demonstrating practical utility in data-limited clinical settings. The code will be available on GitHub.

Amortized inference promises fast test-time Bayesian inference, but existing methods typically target a fixed model or a restricted model family. Extending amortization to new model families often requires retraining or test-time fine-tuning. In this paper, we ask: is it possible to build a single inference network capable of generalizing across varying priors, likelihoods, latent dimensions, and observation counts? We introduce Amortized Factor Inference Networks (AFINs), which describe a model as a list of typed prior and likelihood factors and map this description, together with the observations, to a Gaussian posterior approximation through an encode–merge–decode network whose learned parameters do not depend on the latent dimension or the number of observations. Experiments on synthetic and real-data benchmarks show that a single trained AFIN provides competitive posterior approximations at substantially lower test-time cost than NUTS and per-task variational inference. Optional importance sampling can further improve accuracy.

A properly calibrated rule-based autoscaler can beat every one of six mainstream deep reinforcement learning (DRL) algorithms on cost across every workload we test - so when, if ever, does DRL actually help? We study this in RLScale-Bench, a reproducible benchmark and evaluation protocol for DRL on adaptive resource control, where an agent allocates compute to a dynamic workload under cost and service-level constraints. We evaluate PPO, DQN, A2C, SAC, TD3, and DDPG under matched architectures, training budgets, and reward functions against a calibrated rule-based baseline across six workload patterns and five seeds (240 runs), instantiate the benchmark on Kubernetes Horizontal Pod Autoscaling, and probe distribution-shift generalization. Three findings challenge common assumptions: (i) the calibrated controller achieves the lowest cost on all six workloads, though it trails the best RL agents on bursty and flash traffic; (ii) discrete-action algorithms outperform continuous-action ones by one to two orders of magnitude in constraint violations due to action-space mismatch; and (iii) no single algorithm dominates across workloads, with rankings shifting by up to four positions. The bottleneck in RL-based resource control is not algorithm selection but baseline calibration, reward engineering, and realistic evaluation protocols.

Large Language Models (LLMs) achieve impressive accuracy on mathematical reasoning benchmarks, yet their performance drops when problems are modified with simple changes like different names or numbers. Code execution methods, which let models generate and run Python code instead of reasoning in natural language, have been proposed as a solution, but their effect on reasoning robustness (the ability to maintain accuracy across problem variations) has not been systematically tested. This study evaluates three approaches on 1,000 problems from the GSM-Symbolic dataset: pure reasoning using chain-of-thought (CoT) prompting, single-shot code execution using Program-Aided Language models (PAL), and iterative code execution using Step-by-Step Coding (SBSC). All three were run on paired original and modified problems using Claude Haiku 4.5. CoT was the most robust method, with an accuracy drop of 1.3 percentage points and 1.8% of problems breaking under perturbation. PAL was the least robust at 1.7 percentage points and 3.1% broke, with SBSC falling in between. Although these differences were not statistically significant ($p = .096$), the directional trend was consistent across all measures, suggesting that code execution, whether single-shot or iterative, does not improve reasoning robustness on grade-school-level problem variations.

Understanding the effects of interventions is central to scientific progress, with randomized controlled trials (RCTs) regarded as the gold standard for causal inference in many applied fields. However, RCTs are costly, time-consuming, and often constrained by ethical or practical limitations, motivating the need for causal methods able to draw conclusions from observational data. While such data is collected at ever larger scale, making its use for causal inference is often hindered by the fact that not all variables affecting treatment allocation and the outcome are observed: an issue known as unobserved confounding. In this paper, we introduce a new study design called confounder detection via treatment intent. The idea is to query a human expert who makes treatment decisions, and ask them to compare pairs of units proposed by a principled matching strategy, with the goal of eliciting unobserved variables that explain why treatment decisions differ. We provide a theoretical basis for such a procedure, ascertaining conditions under which such a study design may elicit unobserved confounders. Building on this newly established foundations, we study treatment effects of interventions in the intensive care unit (ICU). First, we show empirical evidence strongly indicating that electronic health records (EHRs) collected in ICUs are subject to unobserved confounding. By using clinical text notes as a proxy for physicians' knowledge and leveraging natural language processing, we provide a proof of concept for our methodology in a semi-synthetic environment with a known ground truth.

Evaluating and mitigating a generative system's susceptibility to jailbreak attacks is critical to its safe deployment. Given the number of deployable systems, full per-configuration evaluation and optimization is impractical. In this paper, we formalize the behavioral geometry of a population of models that, by leveraging previously evaluated and defended models, supports both efficient susceptibility prediction and effective defense transfer across a population. We apply the framework to 82 models spanning 24 providers and to 100 system configurations of a single base model. Simple methods that use the behavioral geometry reach an AUPRC of $0.94$ for susceptibility detection with $\approx98\%$ fewer probes relative to a full evaluation. Using the behavioral geometry to select which model to transfer an optimized defense from outperforms same-provider assignment ($+4.1\%$, $p = 0.004$) at no additional probe cost, with a set of three models sufficient to cover the population. Results are robust to hyperparameter selection and judge.

Causal discovery in time series is increasingly performed using nonlinear machine-learning models, yet the resulting causal relationships are almost always summarized by scalar edge scores. We argue that this practice obscures the true object learned by nonlinear autoregressive models: a state-dependent function whose effect varies across regimes, magnitudes, and contexts. We formalize function-valued causal influence for additive, contribution-decomposable architectures and show that scalar causal scores constitute a severe information bottleneck, conflating between-state variation with within-state residual noise. Using Neural Additive Vector Autoregression as a representative architecture, we introduce a practical framework based on Individual Conditional Expectation for estimating causal response functions directly from trained models. Through controlled synthetic experiments, we demonstrate that edges with indistinguishable scalar scores can exhibit qualitatively different functional behaviors, including monotonic, thresholded, saturating, and sign-changing effects. An applied case study on democratic development further shows that function-valued analysis reveals regime-specific and asymmetric causal structure systematically missed by score-centric approaches.

The topological gap $\Delta$ – the excess $H_1$ total persistence of a critical point cloud over a density-matched null – scales as $\Delta \sim L^{d+\eta}$. We derive this analytically: the spectral integral $I(\alpha) = \sum_{k\neq 0} S_{\mathrm{conn}}(k)\,|k|^{\alpha}$ scales as $L^{2-\alpha-\eta}$ when IR-dominated, giving $I(-2\eta) \sim L^{d+\eta}$. The decomposition $I(-2\eta) = I_0 \cdot I_{\mathrm{shape}}$ separates volume ($I_0 \propto N(1-m^2) \sim L^d$) from anomalous dimension ($I_{\mathrm{shape}} \sim L^{\eta}$); the volume factor accounts for the magnetization-driven per-configuration variance of $\Delta$. We prove the mechanism requires $d < 2 + \eta$ (IR dominance), confining it to $d = 2$ for physical systems; in $d = 3$ the spectral integral is UV-dominated, explaining why density normalization is needed. An $\alpha$-sweep for Potts $q = 4$ at $L = 32$–$256$ finds $\alpha_{\mathrm{opt}}$ in $[-0.75, -0.5]$, consistent with $-2\eta_{\mathrm{Ising}}$ and inconsistent with $-2\eta_{q=4} = -1$; we flag this as tentative pending $L \geq 1024$ confirmation. The $\langle m^2 \cdot I(-2\eta)\rangle$ hyperscaling product is dominated by the correlation $r(m^2, I) \approx -0.98$ via the shared $I_0$ amplitude, so we report it as a covariance-correction analysis. Under a heuristic argument extending Divol–Polonik to inhomogeneous Poisson intensities, the bare PH kernel is flat; the effective kernel acquires $k$-dependence only at criticality. The per-configuration agreement between $\Delta$ and $I(-0.5)$ is primarily a magnetization correlation: $R^2 = 0.91$ at $L = 256$ collapses to $R^2 \approx 0$ once $|M|$ is partialed out. Per-configuration evidence corroborates the $I_0$ Parseval identity but not the $|k|^{-2\eta}$ shape factor; the latter is established by ensemble $L$-scaling.

We evaluate whether clinician pairwise preferences provide a reliable signal of clinical safety in large language model (LLM) evaluation using expert feedback from MOOVE (Massive Open Online Validation and Evaluation), a clinician-led platform collecting blinded pairwise preferences alongside multi-criterion rubric ratings. Clinicians assign scores on a discrete $[-2, +2]$ scale, where negative values indicate clinically unsafe or misleading content. Using 26{,}804 pairwise judgments across outputs from 13 LLMs, contributed by more than 736 clinicians across 28+ countries, we find that clinician preference is a poor proxy for safety-critical performance. Models ranking highly under pairwise preference can still exhibit substantial rates of clinically meaningful failures ($\leq -1$) on dimensions such as Harmlessness and Accuracy. These failures are unevenly distributed across specialties, creating domain-specific “no-go zones” not visible in aggregate rankings or single-number leaderboards. We further analyze contributing factors including prompt length, refusal and escalation behavior, and the relative contributions of safety-critical versus surface-level features. A substantial fraction of preference votes carry no positive safety signal, while feature decomposition shows that surface-level characteristics explain slightly more preference variation than safety-critical rubric differences. Finally, we introduce a clinically adjusted preference ranking combining pairwise preference with rubric-derived feedback, producing a more safety-aware ordering than raw Bradley–Terry strength alone. Our findings support evaluation practices that separate preference from safety, report safety-critical failure rates directly, and incorporate clinically grounded adjustments when ranking LLMs for clinical decision making.

Intelligent scaling of microservices in cloud platforms is crucial for mitigating escalating compute costs while avoiding service disruptions. Current solutions are limited to the univariate space, typically focusing on CPU usage alone to drive scaling decisions. Moreover, they address the problem as a purely forecasting task, focusing on prediction precision while neglecting the greater risks of underestimation and delays in system responsiveness. Alternative solutions are computationally complex, making them impractical for large-scale, real-time deployments. To address these challenges, we present STARIXNet, a lightweight neural network that guides resource allocation decisions in the multivariate space by capturing spatio-temporal relationships among multiple system metrics. STARIXNet models multiple quasi-dependent attributes, in particular the (S)easonal, (T)emporal, (A)uto-(R)egressive (I)ntegrated, and e(X)ogenous patterns, then implements an aggregation policy to finalize scaling decisions, prioritizing service stability, followed by cost-efficiency, over raw forecast accuracy. We empirically demonstrate the performance of STARIXNet by benchmarking against existing solutions in real-world settings. STARIXNet is deployed for critical production microservices at Walmart achieving tangible savings ranging from 10% to 50%, in addition to intangible benefits through improved service stability and customer experience.

Across machine learning, biology, and physics, independently evolving systems often converge toward strikingly similar high-level structures despite radically different microscopic details. Grokking circuits converge across random seeds, evolutionary lineages rediscover similar metabolic solutions, and renormalization flows approach common fixed points. We propose the Hierarchical Emergence Framework (HEF) as a candidate universality framework for such convergence phenomena. HEF models emergence as a phase transition in a mechanism landscape constrained by thermodynamic and information-theoretic laws. The framework introduces a critical energy threshold Ec separating an exploration regime with competing mechanisms from a convergence regime governed by a unique minimum-cost mechanism. Under structural assumptions, we prove physical feasibility, derive strict metric contraction, and establish convergence toward a unique fixed-point representation independent of initial conditions. We further connect this convergence structure to causal emergence through Effective Information and mechanism competition entropy. To test the framework, we study delayed generalization ("grokking") in modular arithmetic transformers across 111 experiments. We identify a reproducible empirical fingerprint of the Ec transition: the weight norm peaks systematically before grokking in 92% of runs. Normalized accuracy curves collapse onto a tanh kink (R^2=0.93) consistent with a Landau-Ginzburg universality class, and all grokked models converge to 0.9745+/-0.014 regardless of initialization, weight decay, or training fraction (ANOVA p>0.13). HEF is not presented as a universal theory of emergence, but as a falsifiable mathematical scaffold for studying convergence phenomena across complex systems.

Gaussian processes with stationary kernels on bounded domains exhibit inflated posterior variance near the boundary. Despite being a long-recognized artifact in geostatistics and a source of over-exploration in Bayesian optimization, the causes and effects of boundary-induced acquisition bias are underexplored. We trace the root cause to a simple geometric mechanism: the truncation of the kernel correlation neighborhood at the domain boundary creates an observation-independent distortion that worsens with dimensionality. We show how this distortion manifests across three acquisition classes: variance maximization concentrates selections at the corners, whereas negative integrated posterior variance and expected predictive information gain move selections inward to axis-aligned interior shells. These patterns arise without reference to any objective function, meaning that acquisition behavior can be dominated by kernel geometry rather than the desired task-specific uncertainty. To quantify this, we introduce a function-free selection-profile diagnostic for arbitrary acquisitions, kernels, and bounded-domain geometries.

In-context learning lets a language model perform a task specified by examples in its prompt. Function vectors capture task information in a compact activation assembled from attention-head outputs. Across two rule families and three Pythia models, we find two opposed functional populations among candidate function-vector heads. Writers support the rule-correct answer, while cancellers systematically suppress it. These roles predict opposing group effects on held-out prompts, where removing cancellers improves accuracy by 2.4 to 6.8 percentage points in all six settings. The dominant suppressive contributions depend on task content, and the same components can support correct predictions in another task. At the population level, substantial support and suppression can nearly cancel. Separating these roles explains how task-related computation can work against correct predictions and how a small aggregate effect can conceal strong opposing contributions.

Decentralized multi-agent swarm coordination remains fundamentally challenged by the combinatorial scaling of joint action spaces and high-overhead training or optimization loops when managing localized group behaviors. To address this problem from a different perspective, this paper introduces the Swarm Policy Interference Network (SPIN) framework, which models multi-agent communication topologies as compressed tensor networks. By factorizing the joint policy tensors of local multi-agent cliques into Open Boundary Condition Matrix Product State (MPS) chains, SPIN replaces explicit exponential joint-action enumeration with clique-level contractions that scale linearly in clique length for fixed local behavior and bond dimensions. To connect raw spatial geometry with this discrete algebraic backend without relying on online training loops, we introduce a decoupled framework combining a lightweight frozen neural mapper evaluated offline with a deterministic zero-shot importance-reweighting filter based on the Radon-Nikodým derivative. We evaluate an executable prototype of this framework within a simulation experiment across distinct task regimes: single-target tracking, decentralized area coverage, and structured multi-goal coordination. The results demonstrate that SPIN functions as a reusable decentralized coordination layer across tracking, dispersion / area coverage, and structured multi-goal coordination, with its strongest gains appearing in multi-goal coordination and dense local interaction regimes, without requiring scenario-specific online optimization or retraining.

Local definitions can assign a familiar word a temporary meaning while its usual associations remain useful elsewhere. We measure interference from those associations with a matched Stroop-style paradigm. A conflict prompt defines doctor as forest and compares forest with the familiar associate hospital. A neutral control replaces doctor with a semantically weak word in both the definition and query while keeping forest and hospital fixed. All 11 model-level means are positive. Aggregate means are also positive for all four conflict families and prompt formats. When no redefinition is present, a stronger preference for the familiar distractor predicts more interference in arbitrary-semantic, polysemy/entity, and domain-definition remappings, while the antonym slope is null. Separately, we patch neutral-control activations into antonym prompts in five 1B-2B models. Patching the defined word, the target word in the definition, and the later query word together restores almost all of the target-minus-distractor margin lost in conflict (normalized recovery R in [0.92,1.06]). Replacing only that target-word activation with a donor from another item reduces recovery in every tested case. Those donor patches also lower the distractor logit, while the contextual target falls much more than under the same-item patch that restores the margin.

Numerous medical imaging problems must be solved under limited labels and constrained compute, yet it remains unclear whether performance gains are driven mainly by more expressive models or by better representation of clinically meaningful anatomy. We study this question through a low-data anatomy-aware benchmark for 5-class cardiac pathology prediction on the public ACDC MRI dataset. Using segmentation-derived patient descriptors from the right ventricle, myocardium, and left ventricle, we compare anatomy-specific and multi-structure representations across linear, kernel, and tree-based classifiers. We find that under limited label settings, representation dominates complexity. These results suggest that in resource-constrained healthcare settings, identifying and representing the most informative anatomy may matter more than the increasing complexity of the model alone.

This is a planning-method note with an unpaired pilot audit. We adapt the classical paired-binary sample-size calculation (Miettinen, 1968) to quantization benchmarks, giving a conservative minimum detectable effect (MDE) bound $\delta^{*} \le (z_{1-\alpha/2}+z_{1-\beta})\sqrt{\rho_d/m}$ in the paired item count $m$ and the FP16-NF4 disagreement rate $\rho_d$. The bound turns "how reliable is my quantization claim?" into a one-line budget a benchmark designer can commit to before running. We illustrate the bound on four models and four benchmarks ($k=5$ splits of $n=100$), and add a parallel MMLU prompt-template study to put the bound's quantization-noise scale alongside the prompt-noise scale. Assuming $\rho_d=0.10$ (an unmeasured planning value), all observed NF4-FP16 deltas fall below the implied MDE, and most cross-split SDs lie within $\pm 1.5$ pp of the binomial reference $\sqrt{p(1-p)/n}$, so much of the variance reported as "benchmark unreliability" on $n=100$ subsamples is binomial sampling noise. The single borderline cell (OPT-WinoGrande, $|\Delta|=3.2$ pp) is below the implied MDE at $\rho_d=0.10$ but above it at $\rho_d=0.05$, illustrating the planning trade-off the bound makes explicit. On MMLU, prompt-template ranges of 2-10 pp meet or exceed the largest observed quantization delta (3.2 pp), so a quantization audit that does not first fix the prompt template absorbs template variance into its noise floor. We complement the bound with a five-line pre-registration template.

Transformer has significantly propelled the development of artificial intelligence, and certainly the development of agents as well. We categorize attention structures of Transformer into two types based on the source of the input information: homogenous and heterogenous attention structures. Heterogenous attention structures, with co-attention as a typical example, process information from different sources. Heterogenous attention structure is the foundation for Transformer models to achieve more complex functions and integrate more modal information. Whether for research purposes or policy requirements, the interpretation of Transformer models with heterogenous attention structures is an important task. The fusion of information from different sources brings new challenges. Our work mainly includes two parts: method and experimentation. In terms of method, we propose an interpretation method for Transformer models with heterogenous attention structures. In terms of experimentation, based on our experimental analysis paradigm, we interpret the operating mechanisms of representative models, conduct semantic interpretation and logical interpretation.

Geometric deep learning organises neural architectures around the symmetries of their data domain, with the choice of symmetry group serving as a geometric prior that determines what representations can be learned. Metric-Aware Principal Component Analysis (MAPCA) parameterises principal component analysis by a positive-definite metric matrix, with a canonical subfamily interpolating between standard PCA and output whitening and a diagonal-metric point recovering Invariant PCA (IPCA). This paper positions MAPCA within the geometric deep learning framework. The metric is read as the geometric prior; the orthogonal group preserving it is the symmetry group it induces; MAPCA solutions are equivariant under this group with the resulting spectrum invariant; and MAPCA's defining constraint is the linear analogue of the Schur-type weight constraints used in equivariant networks. Across six axes - domain, symmetry group, equivariance, invariance, architectural primitive, and geometric prior - we construct a precise dictionary between MAPCA and geometric deep learning. The technical anchor is a uniqueness theorem characterising IPCA as the unique linear data-derived metric in the MAPCA family that is equivariant under arbitrary diagonal rescaling and projects onto the fixed-point set of the action, equivalent under normalisation to the variance-maximisation criterion in its precise form. The paper closes with three bridges: kernel PCA as the nonlinear extension, spectral graph methods as MAPCA on graphs, and a deep MAPCA construction extending the positioning into deep equivariant networks

Large multimodal models (LMMs) have rapidly advanced in perception and reasoning; however, it remains unclear whether these capabilities generalize to discovering visually grounded solutions in open-ended environments, beyond pattern recognition. In such settings, intelligence requires more than answering well-posed questions: it involves identifying how elements in a scene can be repurposed in non-obvious yet physically feasible ways. This form of creative problem-solving is central to human intelligence, but remains largely untested in current benchmarks. To evaluate this ability, we introduce MM-CreativityBench, a benchmark for affordance-grounded creative tool use in visually rich, physically constrained environments. Each instance presents a scenario image with structured views of candidate entities and their parts, enabling fine-grained, interactive evaluation of how models iteratively inspect the scene, identify relevant affordances, and compose visually and physically grounded solutions. Our experiments show that current LMMs often fall short, not due to lack of generative capability, but because they do not sustain grounded exploration. Models often overlook relevant entities, under-examine critical parts, or hallucinate attributes not grounded in the image. Motivated by this failure mode, we propose affordance-grounded alignment, which casts creative tool use as a preference learning problem. Using Direct Preference Optimization, we encourage models to prefer attribute-affordance reasoning grounded in visual evidence over hallucinated alternatives. In addition, we incorporate supervision derived from an affordance knowledge base to guide broader entity exploration and multi-turn planning. Our results show consistent gains in selecting the correct entities and parts, while substantially reducing hallucination and grounding-related errors.

A representation that scrambles the true degrees of freedom of the world cannot support reliable planning or compositional generalization. We prove that LeJEPA (alignment plus Gaussian regularization) linearly recovers the world's latent variables from nonlinear observations, a property known as linear identifiability, in a broad class of worlds where latents evolve under stationary, additive-noise transitions. Our main result is that among all such worlds, the Gaussian is the unique latent distribution for which this guarantee holds. The forward direction rests on a spectral decomposition in which each degree of nonlinearity is strictly penalized by alignment, making the linear map the optimum; the converse rules out every non-Gaussian alternative. We further prove an approximate identifiability result where the guarantee degrades gracefully, and show that linear, orthogonal identifiability enables optimal latent-space planning. We validate the theory with experiments ranging from 2D examples to 1024-dimensional latents, including distributional ablations and pixel-based robotic control. Our theory turns an empirically successful recipe into a mathematical guarantee, providing the foundation for building World Models that provably recover the structure of the world.

Hepatocellular carcinoma (HCC) is biologically heterogeneous, shaped by the interplay between hepatic functional reserve and tumor-related oncologic factors; thus, similar survival outcomes may reflect fundamentally different underlying biological processes. Prognostic modeling in HCC is informed by rich multimodal information from multiparametric MRI and radiology reports from routine clinical practice. Existing prognostic vision-language models (VLMs) learn a single entangled latent representation that blends hepatic and tumor-related factors, limiting both accuracy and biological interpretability. We present BioFact-MoE, a biologically factorized Mixture of Experts (MoE) framework that explicitly decomposes liver and tumor factors via biologically supervised experts within a residual MoE survival architecture. On a HCC cohort of N=588 patients (pretrained on 4,582 3D MRI image-report pairs), BioFact-MoE consistently improves survival prediction over all baselines across time horizons, achieving 12-, 18-, and 24-month AUCs of 75.33%, 75.85%, and 73.96%. Beyond scalar risk prediction, gated expert weights enable phenotype-aware risk stratification. Pathway-informed gating uncovers clinically meaningful treatment-associated survival heterogeneity. In held-out validation, hepatic and tumor embeddings show selective associations with liver function and tumor burden markers, respectively (p<0.05), without supervision. The code is available at https://github.com/jy-639/BioFact-MoE.

We study adversarial online learning with hidden-convex losses, i.e., nonconvex losses that become convex after a nonlinear reparameterization. Ghai, Lu and Hazan (2022) proved that, under geometric and smoothness assumptions, online gradient descent (OGD) on such nonconvex losses approximately simulates online mirror descent (OMD) on the underlying convex losses with a suitable regularizer, yielding $\mathcal{O}(T^{2/3})$ regret. They left open whether the optimal $\Theta(\sqrt{T})$ regret from online convex optimization can be recovered in this hidden-convex setting. We answer this question affirmatively. More specifically, via a sharper discrete-time algorithmic equivalence argument, we prove that OGD achieves $\mathcal{O}(\sqrt{T})$ regret under the same assumptions, matching the optimal worst-case rate for adversarial online convex optimization. We also address another open question of Ghai, Lu and Hazan (2022) by clarifying the geometry required for this algorithmic equivalence. We replace the diagonal-Jacobian sufficient condition with a necessary-and-sufficient Hessian compatibility condition, thereby expanding the class of admissible reparameterizations. We complement our tight regret bound with a lower bound showing that the Hessian compatibility assumption is essential for OGD; when it fails, we construct a smooth reparameterization and an adversarial sequence of hidden-convex losses for which OGD suffers $\Omega(T)$ regret. Finally, we extend our analysis to one-point bandit feedback and prove a $\mathcal{O}(T^{3/4})$ expected regret bound for bandit OGD with spherical smoothing, matching its classical rate on convex losses.

Recent studies on hallucination detection have shown that hallucination-related signals are more strongly encoded in intermediate layers than in the final layer of large language models (LLMs). Although a growing body of work has sought to exploit this property for hallucination detection, how to automate the selection of high-performing layers remains underexplored, and principled methods for this purpose are still lacking. To address this gap, we first propose several hypotheses for why such signals emerge in intermediate layers and evaluate corresponding criteria for automatic layer selection across diverse LLM architectures, scales, and tasks, covering both question answering and summarization hallucination detection benchmarks. However, we find that none of these criteria consistently delivers satisfactory performance. We therefore propose a new selection criterion, First Effective Peak of Intrinsic Dimension (FEPoID), which consistently identify optimal or near-optimal layers and outperforms both the aforementioned criteria and existing hallucination detection baselines. FEPoID is training-free and incurs negligible computational overhead. In addition, we study the generation behaviors of LLMs and introduce a simple yet effective truncation strategy, which further amplifies hallucination-related signals and substantially improves overall detection performance. Code is publicly available at https://github.com/DesoloYw/Automatic-Layer-Selection-for-Hallucination-Detection.git