Papers

Meta-learning offers a principled framework leveraging task-invariant priors from related tasks, with which task-specific models can be fine-tuned on downstream tasks, even with limited data records. Gradient-based meta-learning (GBML) relies on gradient descent (GD) to adapt the prior to a new task. Albeit effective, these methods incur high computational overhead that scales linearly with the number of GD steps. To enhance efficiency and scalability, existing methods approximate the gradient of prior parameters (meta-gradient) via truncated backpropagation, yet suffer large approximation errors. Targeting accurate approximation, this work puts forth binomial GBML (BinomGBML), which relies on a truncated binomial expansion for meta-gradient estimation. This novel expansion endows more information in the meta-gradient estimation via efficient parallel computation. As a running paradigm applied to model-agnostic meta-learning (MAML), the resultant BinomMAML provably enjoys error bounds that not only improve upon existing approaches, but also decay super-exponentially under mild conditions. Numerical tests corroborate the theoretical analysis and showcase boosted performance with slightly increased computational overhead.

In medical image segmentation, uncertainty estimates are often reported but rarely used to guide decisions. We study the missing step: how uncertainty maps are converted into actionable policies such as accepting, flagging, or deferring predictions. We formulate segmentation as a two-stage pipeline, estimation followed by decision, and show that optimizing uncertainty alone fails to capture most of the achievable safety gains. Using retinal vessel segmentation benchmarks (DRIVE, STARE, CHASE_DB1), we evaluate two uncertainty sources (Monte Carlo Dropout and Test-Time Augmentation) combined with three deferral strategies, and introduce a simple confidence-aware deferral rule that prioritizes uncertain and low-confidence predictions. Our results show that the best method and policy combination removes up to 80 percent of segmentation errors at only 25 percent pixel deferral, while achieving strong cross-dataset robustness. We further show that calibration improvements do not translate to better decision quality, highlighting a disconnect between standard uncertainty metrics and real-world utility. These findings suggest that uncertainty should be evaluated based on the decisions it enables, rather than in isolation.

Neural models for TCR-pMHC binding prediction are susceptible to shortcut learning: they exploit spurious correlations in training data – such as peptide length bias or V-gene co-occurrence – rather than the physical binding interface. This renders predictions brittle under family-held-out and distance-aware evaluation, where such shortcuts do not transfer. We introduce Counterfactual Invariant Prediction (CIP), a training framework that generates biologically constrained counterfactual peptide edits and enforces invariance to edits at non-anchor positions while amplifying sensitivity at MHC anchor residues. CIP augments the base classifier with two auxiliary objectives: (1) an invariance loss penalizing prediction changes under conservative non-anchor substitutions, and (2) a contrastive loss encouraging large prediction changes under anchor-position disruptions. Evaluated on a curated VDJdb-IEDB benchmark under family-held-out, distance-aware, and random splits, CIP achieves AUROC 0.831 and counterfactual consistency (CFC) 0.724 under the challenging family-held-out protocol – a 39.7% reduction in shortcut index relative to the unconstrained baseline. Ablations confirm that anchor-aware edit generation is the dominant driver of OOD gains, providing a practical recipe for causally-grounded TCR specificity modeling.

Adaptive Conformal Inference (ACI) provides distribution-free prediction intervals with asymptotic coverage guarantees for time series under distribution shift. However, ACI only adapts the quantile threshold – it cannot shift the interval center. When a base forecaster develops persistent bias after a regime change, ACI compensates by widening intervals symmetrically, producing unnecessarily conservative bands. We propose Bias-Corrected ACI (BC-ACI), which augments standard ACI with an online exponentially weighted moving average (EWM) estimate of forecast bias. BC-ACI corrects nonconformity scores before quantile computation and re-centers prediction intervals, addressing the root cause of miscalibration rather than its symptom. An adaptive dead-zone threshold suppresses corrections when estimated bias is indistinguishable from noise, ensuring no degradation on well-calibrated data. In controlled experiments across 688 runs spanning two base models, four synthetic regimes, and three real datasets, BC-ACI reduces Winkler interval scores by 13–17% under mean and compound distribution shifts (Wilcoxon p < 0.001) while maintaining equivalent performance on stationary data (ratio 1.002x). We provide finite-sample analysis showing that coverage guarantees degrade gracefully with bias estimation error.

Anomaly detection aims to identify observations that deviate from expected behavior. Because anomalous events are inherently sparse, most frameworks are trained exclusively on normal data to learn a single reference model of normality. This implicitly assumes that normal behavior can be captured by a single, unconditional reference distribution. In practice, however, anomalies are often context-dependent: A specific observation may be normal under one operating condition, yet anomalous under another. As machine learning systems are deployed in dynamic and heterogeneous environments, these fixed-context assumptions introduce structural ambiguity, i.e., the inability to distinguish contextual variation from genuine abnormality under marginal modeling, leading to unstable performance and unreliable anomaly assessments. While modern sensing systems frequently collect multimodal data capturing complementary aspects of both system behavior and operating conditions, existing methods treat all data streams equally, without distinguishing contextual information from anomaly-relevant signals. As a result, abnormality is often evaluated without explicitly conditioning on operating conditions. We argue that multimodal anomaly detection should be reframed as a cross-modal contextual inference problem, in which modalities play asymmetric roles, separating context from observation, to define abnormality conditionally rather than relative to a single global reference. This perspective has implications for model design, evaluation protocols, and benchmark construction, and outline open research challenges toward robust, context-aware multimodal anomaly detection.

Analog optical computers promise large efficiency gains for machine learning inference, yet no demonstration has moved beyond small-scale image benchmarks. We benchmark the analog optical computer (AOC) digital twin on mortgage approval classification from 5.84 million U.S. HMDA records and separate three sources of accuracy loss. On the original 19 features, the AOC reaches 94.6% balanced accuracy with 5,126 parameters (1,024 optical), compared with 97.9% for XGBoost; the 3.3 percentage-point gap narrows by only 0.5pp when the optical core is widened from 16 to 48 channels, suggesting an architectural rather than hardware limitation. Restricting all models to a shared 127-bit binary encoding drops every model to 89.4–89.6%, with an encoding cost of 8pp for digital models and 5pp for the AOC. Seven calibrated hardware non-idealities impose no measurable penalty. The three resulting layers of limitation (encoding, architecture, hardware fidelity) locate where accuracy is lost and what to improve next.

Mapping the spatial distribution of species is essential for conservation policy and invasive species management. Species distribution models (SDMs) are the primary tools for this task, serving two purposes: achieving robust predictive performance while providing ecological insights into the driving factors of distribution. However, the increasing complexity of deep learning SDMs has made extracting these insights more challenging. To reconcile these objectives, we propose the first implementation of concept-based Explainable AI (XAI) for SDMs. We leverage the Robust TCAV (Testing with Concept Activation Vectors) methodology to quantify the influence of landscape concepts on model predictions. To enable this, we provide a new open-access landscape concept dataset derived from high-resolution multispectral and LiDAR drone imagery. It includes 653 patches across 15 distinct landscape concepts and 1,450 random reference patches, designed to suit a wide range of species. We demonstrate this approach through a case study of two aquatic insects, Plecoptera and Trichoptera, using two Convolutional Neural Networks and one Vision Transformer. Results show that concept-based XAI helps validate SDMs against expert knowledge while uncovering novel associations that generate new ecological hypotheses. Robust TCAV also provides landscape-level information, useful for policy-making and land management. Code and datasets are publicly available.

Exploratory Landscape Analysis (ELA) provides numerical features for characterizing black-box optimization problems. In high-dimensional settings, however, ELA suffers from sparsity effects, high estimator variance, and the prohibitive cost of computing several feature classes. Dimensionality reduction has therefore been proposed as a way to make ELA applicable in such settings, but it remains unclear whether features computed in reduced spaces still reflect intrinsic properties of the original landscape. In this work, we investigate the robustness of ELA features under dimensionality reduction via Random Gaussian Embeddings (RGEs). Starting from the same sampled points and objective values, we compute ELA features in projected spaces and compare them to those obtained in the original search space across multiple sample budgets and embedding dimensions. Our results show that linear random projections often alter the geometric and topological structure relevant to ELA, yielding feature values that are no longer representative of the original problem. While a small subset of features remains comparatively stable, most are highly sensitive to the embedding. Moreover, robustness under projection does not necessarily imply informativeness, as apparently robust features may still reflect projection-induced artifacts rather than intrinsic landscape characteristics.

Large Language Models (LLMs) rely heavily on Key-Value (KV) caching to minimize inference latency. However, standard KV caches are context-dependent: reusing a cached document in a new context requires recomputing KV states to account for shifts in attention distribution. Existing solutions such as CacheBlend, EPIC, and SAM-KV mitigate this issue by selectively recomputing a subset of tokens; however, they still incur non-negligible computational overhead (FLOPs) and increased Time-to-First-Token (TTFT) latency. In this paper, we propose KV Packet, a recomputation-free cache reuse framework that treats cached documents as immutable "packets" wrapped in light-weight trainable soft-token adapters, which are trained via self-supervised distillation to bridge context discontinuities. Experiments on Llama-3.1 and Qwen-3 demonstrate that the proposed KV Packet method achieves near-zero FLOPs and lower TTFT than recomputation-based baselines, while retaining F1 scores comparable to those of the full recomputation baseline. Code and experimental results are available at https://github.com/ChuangtaoChen-TUM/KVPacket.

Causal representation learning (CRL) aims to identify the underlying latent variables from high-dimensional observations, even when variables are dependent with each other. We study this problem for latent variables that follow a potentially degenerate Gaussian mixture distribution and that are only observed through the transformation via a piecewise affine mixing function. We provide a series of progressively stronger identifiability results for this challenging setting in which the probability density functions are ill-defined because of the potential degeneracy. For identifiability up to permutation and scaling, we leverage a sparsity regularization on the learned representation. Based on our theoretical results, we propose a two-stage method to estimate the latent variables by enforcing sparsity and Gaussianity in the learned representations. Experiments on synthetic and image data highlight our method's effectiveness in recovering the ground-truth latent variables.

Rare events such as conformational changes in biomolecules, phase transitions, and chemical reactions are central to the behavior of many physical systems, yet they are extremely difficult to study computationally because unbiased simulations seldom produce them. Transition Path Theory (TPT) provides a rigorous statistical framework for analyzing such events: it characterizes the ensemble of reactive trajectories between two designated metastable states (reactant and product), and its central object–the committor function, which gives the probability that the system will next reach the product rather than the reactant–encodes all essential kinetic and thermodynamic information. We introduce a framework that casts committor estimation as a stochastic optimal control (SOC) problem. In this formulation the committor defines a feedback control–proportional to the gradient of its logarithm–that actively steers trajectories toward the reactive region, thereby enabling efficient sampling of reactive paths. To solve the resulting hitting-time control problem we develop two complementary objectives: a direct backpropagation loss and a principled off-policy Value Matching loss, for which we establish first-order optimality guarantees. We further address metastability, which can trap controlled trajectories in intermediate basins, by introducing an alternative sampling process that preserves the reactive current while lowering effective energy barriers. On benchmark systems, the framework yields markedly more accurate committor estimates, reaction rates, and equilibrium constants than existing methods.

As Large Language Models (LLMs) are increasingly integrated into agentic workflows, their unpredictability stemming from numerical instability has emerged as a critical reliability issue. While recent studies have demonstrated the significant downstream effects of these instabilities, the root causes and underlying mechanisms remain poorly understood. In this paper, we present a rigorous analysis of how unpredictability is rooted in the finite numerical precision of floating-point representations, tracking how rounding errors propagate, amplify, or dissipate through Transformer computation layers. Specifically, we identify a chaotic "avalanche effect" in the early layers, where minor perturbations trigger binary outcomes: either rapid amplification or complete attenuation. Beyond specific error instances, we demonstrate that LLMs exhibit universal, scale-dependent chaotic behaviors characterized by three distinct regimes: 1) a stable regime, where perturbations fall below an input-dependent threshold and vanish, resulting in constant outputs; 2) a chaotic regime, where rounding errors dominate and drive output divergence; and 3) a signal-dominated regime, where true input variations override numerical noise. We validate these findings extensively across multiple datasets and model architectures.

Many materials show anisotropic light scattering patterns due to the shape and local alignment of their underlying micro structures: surfaces with small elements such as fibers, or the ridges of a brushed metal, are very sparse and require a high spatial resolution to be properly represented as a volume. The acquisition of voxel data from such objects is a time and memory-intensive task, and most rendering approaches require an additional Level-of-Detail (LoD) data structure to aggregate the visual appearance, as observed from multiple distances, in order to reduce the number of samples computed per pixel (E.g.: MIP mapping). In this work we introduce first, an efficient parallel voxelization method designed to facilitate fast data aggregation at multiple resolution levels, and second, a novel representation based on hierarchical SGGX clustering that provides better accuracy than baseline methods. We validate our approach with a CUDA-based implementation of the voxelizer, tested both on triangle meshes and volumetric fabrics modeled with explicit fibers. Finally, we show the results generated with a path tracer based on the proposed LoD rendering model.

This paper presents HUANet, a constrained deep neural network architecture that unrolls the Alternating Direction Method of Multipliers (ADMM) into a trainable neural network for accelerating parametric constrained convex optimization. Existing end-to-end learning methods operate as black-box mappings from parameters to solutions, often without explicitly incorporating optimality principles or guaranteeing constraint satisfaction. To address these limitations, HUANet embeds a hard-constrained neural network within each unrolled ADMM iteration, where a differentiable correction stage enforces the affine equalities of the primal subproblem. Furthermore, we incorporate first-order optimality conditions into a self-supervised training loss to promote the convergence of the proposed unrolled algorithm. Extensive numerical experiments for benchmark optimization problems and a control application demonstrate and validate the effectiveness of HUANet in accelerating constrained convex optimization solving.

Large language models can be aligned with human preferences through offline reinforcement learning (RL) on small labeled datasets. While single-objective alignment is well-studied, many real-world applications demand the simultaneous optimization of multiple conflicting rewards, e.g. activity and specificity for proteins, or helpfulness and harmlessness for chatbots. Prior work has largely relied on linear reward scalarization, which provably fails to recover non-convex regions of the Pareto front. In this paper, instead of scalarizing the rewards directly, we frame multi-objective RL itself as an optimization problem to be scalarized via smooth Tchebycheff scalarization, a recent technique that overcomes the drawbacks of linear scalarization. We use this formulation to derive Smooth Tchebycheff Optimization of Multi-Objective Preferences (STOMP), a novel offline RL algorithm that extends direct preference optimization to the multi-objective setting in a principled way by standardizing the individual rewards based on their observed distributions. We empirically validate STOMP using multiple base models on 3 protein engineering and 3 chatbot alignment datasets. Compared to state-of-the-art baselines, STOMP achieves or ties for the highest hypervolumes on 16/18 protein tasks and 5/6 natural language tasks. We thus demonstrate that STOMP is a powerful, robust multi-objective alignment algorithm that can meaningfully improve post-training in multiple domains.

3D Gaussian Splatting (3DGS) has recently enabled highly photorealistic 3D reconstruction from casually captured multi-view images. However, this accessibility raises a privacy concern: publicly available images or videos can be exploited to reconstruct detailed 3D models of scenes or objects without the owner's consent. We present PatchPoison, a lightweight dataset-poisoning method that prevents unauthorized 3D reconstruction. Unlike global perturbations, PatchPoison injects a small high-frequency adversarial patch, a structured checkerboard, into the periphery of each image in a multi-view dataset. The patch is designed to corrupt the feature-matching stage of Structure-from-Motion (SfM) pipelines such as COLMAP by introducing spurious correspondences that systematically misalign estimated camera poses. Consequently, downstream 3DGS optimization diverges from the correct scene geometry. On the NeRF-Synthetic benchmark, inserting a 12 X 12 pixel patch increases reconstruction error by 6.8x in LPIPS, while the poisoned images remain unobtrusive to human viewers. PatchPoison requires no pipeline modifications, offering a practical, "drop-in" preprocessing step for content creators to protect their multi-view data.

This paper studies continuous-time stochastic control problems whose controlled states are fully non-Markovian and depend on unknown model parameters. Such problems arise naturally in path-dependent stochastic differential equations, rough-volatility hedging, and systems driven by fractional Brownian motion. Building on the discrete skeleton approach developed in earlier work, we propose a Monte Carlo learning methodology for the associated embedded backward dynamic programming equation. Our main contribution is twofold. First, we construct explicit dominating training laws and Radon–Nikodym weights for several representative classes of non-Markovian controlled systems. This yields an off-model training architecture in which a fixed synthetic dataset is generated under a reference law, while the dynamic programming operators associated with a target model are recovered by importance sampling. Second, we use this structure to design an adaptive update mechanism under parametric model uncertainty, so that repeated recalibration can be performed by reweighting the same training sample rather than regenerating new trajectories. For fixed parameters, we establish non-asymptotic error bounds for the approximation of the embedded dynamic programming equation via deep neural networks. For adaptive learning, we derive quantitative estimates that separate Monte Carlo approximation error from model-risk error. Numerical experiments illustrate both the off-model training mechanism and the adaptive importance-sampling update in structured linear-quadratic examples.

Thermodynamic cycles are pivotal in determining the efficacy of energy conversion systems. Traditional design methodologies, which rely on expert knowledge or exhaustive enumeration, are inefficient and lack scalability, thereby constraining the discovery of high-performance cycles. In this study, we introduce a graph-based hierarchical reinforcement learning approach for the co-design of structure parameters in thermodynamic cycles. These cycles are encoded as graphs, with components and connections depicted as nodes and edges, adhering to grammatical constraints. A deep learning-based thermophysical surrogate facilitates stable graph decoding and the simultaneous resolution of global parameters. Building on this foundation, we develop a hierarchical reinforcement learning framework wherein a high-level manager explores structural evolution and proposes candidate configurations, whereas a low-level worker optimizes parameters and provides performance rewards to steer the search towards high-performance regions. By integrating graph representation, thermophysical surrogate, and manager-worker learning, this method establishes a fully automated pipeline for encoding, decoding, and co-optimization. Using heat pump and heat engine cycles as case studies, the results demonstrate that the proposed method not only replicates classical cycle configurations but also identifies 18 and 21 novel heat pump and heat engine cycles, respectively. Relative to classical cycles, the novel configurations exhibit performance improvements of 4.6% and 133.3%, respectively, surpassing the traditional designs. This method effectively balances efficiency with broad applicability, providing a practical and scalable intelligent alternative to expert-driven thermodynamic cycle design.

Satellite sea surface temperature (SST) products underpin global coral bleaching monitoring, yet they measure only the ocean skin. Corals inhabit depths from the shallows to beyond 20 metres, where temperatures can be 1-3{\deg}C cooler than the surface; applying satellite SST uniformly to all depths therefore overestimates subsurface thermal stress. We present a physics-informed neural network (PINN) that fuses NOAA Coral Reef Watch SST with sparse in-situ temperature loggers within the one-dimensional vertical heat equation, enforcing SST as a hard surface boundary condition and jointly learning effective thermal diffusivity (ąppa}) and light attenuation (Kd). Validated across four Great Barrier Reef sites (30 holdout experiments), the PINN achieves 0.25-1.38{\deg}C RMSE at unseen depths. Under extreme sparsity (three training depths), the PINN maintains 0.27{\deg}C RMSE at the 5 metre holdout and 0.32{\deg}C at the 9.1 metre holdout, where statistical baselines collapse to >1.8{\deg}C; it outperforms a physics-only finite-difference baseline in 90% of experiments. Depth-resolved Degree Heating Day (DHD) profiles show that thermal stress attenuates with depth: at Davies Reef, DHD drops from 0.29 at the surface to zero by 10.7 metres, consistent with logger observations, while satellite DHD remains constant at 0.31 across all depths. However, the PINN underestimates absolute DHD at shallow depths because its smooth predictions attenuate the short-duration peaks that drive threshold exceedances; PINN DHD values should be interpreted as conservative lower bounds on depth-resolved stress. These results demonstrate that physics-constrained fusion of satellite SST with sparse loggers can extend bleaching assessment to the depth dimension using existing observational infrastructure.

The explosive growth of system logs makes streaming compression essential, yet existing log anomaly detection (LAD) methods incur severe pre-processing overhead by requiring full decompression and parsing. We introduce CLAD, the first deep learning framework to perform LAD directly on compressed byte streams. CLAD bypasses these bottlenecks by exploiting a key insight: normal logs compress into regular byte patterns, while anomalies systematically disrupt them. To extract these multi-scale deviations from opaque bytes, we propose a purpose-built architecture integrating a dilated convolutional byte encoder, a hybrid Transformer–mLSTM, and four-way aggregation pooling. This is coupled with a two-stage training strategy of masked pre-training and focal-contrastive fine-tuning to effectively handle severe class imbalance. Evaluated across five datasets, CLAD achieves a state-of-the-art average F1-score of 0.9909 and outperforms the best baseline by 2.72 percentage points. It delivers superior accuracy while completely eliminating decompression and parsing overheads, offering a robust solution that generalizes to structured streaming compressors.

We present the first analytical study of ECD, focusing on the one-dimensional setting for this first installment. We formalize a stochastic ECD dynamics (sECD) with energy-preserving noise, as well as a quantum analog of the ECD Hamiltonian (qECD), providing the foundation for a quantum algorithm through Hamiltonian simulation in a tractable model where the barrier-crossing mechanism can be computed explicitly. For one-dimensional double-well objectives in the under-guessing regime, we compute the expected dynamical hitting times from a local minimum to the global minimum. We prove that both sECD and qECD exhibit exponential improvements in continuous hitting time relative to their respective gradient-based baselines, stochastic gradient descent (SGD) and quantum tunneling walk (QTW). For objectives with tall barriers, qECD admits a further hitting time improvement over sECD. Mechanistically, ECD sidesteps the exponential cost associated with rare-escape events of SGD from local minima by moving from dissipative to energy-conserving dynamics.

On-policy distillation (OPD) has become a core technique in the post-training of large language models, yet its training dynamics remain poorly understood. This paper provides a systematic investigation of OPD dynamics and mechanisms. We first identify that two conditions govern whether OPD succeeds or fails: (i) the student and teacher should share compatible thinking patterns; and (ii) even with consistent thinking patterns and higher scores, the teacher must offer genuinely new capabilities beyond what the student has seen during training. We validate these findings through weak-to-strong reverse distillation, showing that same-family 1.5B and 7B teachers are distributionally indistinguishable from the student's perspective. Probing into the token-level mechanism, we show that successful OPD is characterized by progressive alignment on high-probability tokens at student-visited states, a small shared token set that concentrates most of the probability mass (97%-99%). We further propose two practical strategies to recover failing OPD: off-policy cold start and teacher-aligned prompt selection. Finally, we show that OPD's apparent free lunch of dense token-level reward comes at a cost, raising the question of whether OPD can scale to long-horizon distillation.

On-policy distillation (OPD) is an effective post-training paradigm for large language models but requires a live teacher server throughout training, resulting in substantial infrastructure overhead. We investigate whether OPD can be performed offline by precomputing teacher log-probabilities once over SFT rollouts and reusing them during training. We find that naively doing so fails to reliably match standard OPD, and trace the root cause to a previously overlooked condition we term teacher consistency, requiring that the same teacher be used for both supervised fine-tuning and OPD. Violating this condition introduces a gradient bias that degrades performance for both offline and online OPD. Building on this insight, we propose Lightning OPD, an offline on-policy distillation framework that enforces teacher consistency and eliminates the need for a live teacher server entirely. We prove that, under teacher consistency, Lightning OPD shares the same optimum as standard OPD, with bounded gradient discrepancy and an implicit regularization effect that helps prevent policy drift. Experiments on math reasoning and code generation show that Lightning OPD achieves comparable performance to standard OPD while delivering 4.0x higher training efficiency. Starting from an SFT-initialized Qwen3-8B-Base model, Lightning OPD reaches 69.9% on AIME 2024 in just 30 GPU hours. Lightning OPD further scales to MoE architectures, training Qwen3-30B-A3B to 71.0% on AIME 2024 on a single 8xH100 node, substantially lowering the barrier for academic research on LLM post-training. Our code is released at https://github.com/jet-ai-projects/Lightning-OPD.

Predicting counterfactual outcomes in longitudinal data, where sequential treatment decisions heavily depend on evolving patient states, is critical yet notoriously challenging due to complex time-dependent confounding and inadequate uncertainty quantification in existing methods. We introduce the Causal Diffusion Model (CDM), the first denoising diffusion probabilistic approach explicitly designed to generate full probabilistic distributions of counterfactual outcomes under sequential interventions. CDM employs a novel residual denoising architecture with relational self-attention, capturing intricate temporal dependencies and multimodal outcome trajectories without requiring explicit adjustments (e.g., inverse-probability weighting or adversarial balancing) for confounding. In rigorous evaluation on a pharmacokinetic-pharmacodynamic tumor-growth simulator widely adopted in prior work, CDM consistently outperforms state-of-the-art longitudinal causal inference methods, achieving a 15-30% relative improvement in distributional accuracy (1-Wasserstein distance) while maintaining competitive or superior point-estimate accuracy (RMSE) under high-confounding regimes. By unifying uncertainty quantification and robust counterfactual prediction in complex, sequentially confounded settings, without tailored deconfounding, CDM offers a flexible, high-impact tool for decision support in medicine, policy evaluation, and other longitudinal domains.

Balancing convergence speed, generalization capability, and computational efficiency remains a core challenge in deep learning optimization. First-order gradient descent methods, epitomized by stochastic gradient descent (SGD) and Adam, serve as the cornerstone of modern training pipelines. However, large-scale model training, stringent differential privacy requirements, and distributed learning paradigms expose critical limitations in these conventional approaches regarding privacy protection and memory efficiency. To mitigate these bottlenecks, researchers explore second-order optimization techniques to surpass first-order performance ceilings, while zeroth-order methods reemerge to alleviate memory constraints inherent to large-scale training. Despite this proliferation of methodologies, the field lacks a cohesive framework that unifies underlying principles and delineates application scenarios for these disparate approaches. In this work, we retrospectively analyze the evolutionary trajectory of deep learning optimization algorithms and present a comprehensive empirical evaluation of mainstream optimizers across diverse model architectures and training scenarios. We distill key emerging trends and fundamental design trade-offs, pinpointing promising directions for future research. By synthesizing theoretical insights with extensive empirical evidence, we provide actionable guidance for designing next-generation highly efficient, robust, and trustworthy optimization methods. The code is available at https://github.com/APRIL-AIGC/Awesome-Optimizer.

The Sauer-Shelah-Perles Lemma is a cornerstone of combinatorics and learning theory, bounding the size of a binary hypothesis class in terms of its Vapnik-Chervonenkis (VC) dimension. For classes of functions over a $k$-ary alphabet, namely the multiclass setting, the Natarajan dimension has long served as an analogue of VC dimension, yet the corresponding Sauer-type bounds are suboptimal for alphabet sizes $k>2$. In this work, we establish a sharp Sauer inequality for multiclass and list prediction. Our bound is expressed in terms of the Daniely–Shalev-Shwartz (DS) dimension, and more generally with its extension, the list-DS dimension – the combinatorial parameters that characterize multiclass and list PAC learnability. Our bound is tight for every alphabet size $k$, list size $\ell$, and dimension value, replacing the exponential dependence on $\ell$ in the Natarajan-based bound by the optimal polynomial dependence, and improving the dependence on $k$ as well. Our proof uses the polynomial method. In contrast to the classical VC case, where several direct combinatorial proofs are known, we are not aware of any purely combinatorial proof in the DS setting. This motivates several directions for future research, which are discussed in the paper. As consequences, we obtain improved sample complexity upper bounds for list PAC learning and for uniform convergence of list predictors, sharpening the recent results of Charikar et al.~(STOC~2023), Hanneke et al.~(COLT~2024), and Brukhim et al.~(NeurIPS~2024).

The most cited calibration result in deep learning – post-temperature-scaling ECE of 0.012 on CIFAR-100 (Guo et al., 2017) – is below the statistical noise floor. We prove this is not a failure of the experiment but a law: the minimax rate for estimating calibration error with model error rate epsilon is Theta((Lepsilon/m)^{1/3}), and no estimator can beat it. This "verification tax" implies that as AI models improve, verifying their calibration becomes fundamentally harder – with the same exponent in opposite directions. We establish four results that contradict standard evaluation practice: (1) self-evaluation without labels provides exactly zero information about calibration, bounded by a constant independent of compute; (2) a sharp phase transition at mepsilon approx 1 below which miscalibration is undetectable; (3) active querying eliminates the Lipschitz constant, collapsing estimation to detection; (4) verification cost grows exponentially with pipeline depth at rate L^K. We validate across five benchmarks (MMLU, TruthfulQA, ARC-Challenge, HellaSwag, WinoGrande; ~27,000 items) with 6 LLMs from 5 families (8B-405B parameters, 27 benchmark-model pairs with logprob-based confidence), 95% bootstrap CIs, and permutation tests. Self-evaluation non-significance holds in 80% of pairs. Across frontier models, 23% of pairwise comparisons are indistinguishable from noise, implying that credible calibration claims must report verification floors and prioritize active querying once gains approach benchmark resolution.

Traditional fixed-depth architectures scale quality by increasing training FLOPs, typically through increased parameterization, at the expense of a higher memory footprint, or data. A potential alternative is looped architectures, which instead increase FLOPs by sending activations through a block of layers in a loop. While promising, existing recipes for training looped architectures can be unstable, suffering from residual explosion and loss spikes. We address these challenges by recasting looping as a nonlinear time-variant dynamical system over the residual stream. Via a linear approximation to this system, we find that instability occurs in existing looped architectures as a result of large spectral norms in their injection parameters. To address these instability issues, we propose Parcae, a novel stable, looped architecture that constrains the spectral norm of the injection parameters via discretization of a negative diagonal parameterization. As a result, Parcae achieves up to 6.3% lower validation perplexity over prior large-scale looped models. Using our stable looped architecture, we investigate the scaling properties of looping as a medium to improve quality by increasing FLOPs in training and test-time. For training, we derive predictable power laws to scale FLOPs while keeping parameter count fixed. Our initial scaling laws suggest that looping and data should be increased in tandem, given a fixed FLOP budget. At test-time, we find that Parcae can use looping to scale compute, following a predictable, saturating exponential decay. When scaled up to 1.3B parameters, we find that Parcae improves CORE and Core-Extended quality by 2.99 and 1.18 points when compared to strong Transformer baselines under a fixed parameter and data budget, achieving a relative quality of up to 87.5% a Transformer twice the size.

Deep neural networks are typically trained by uniformly sampling large datasets across epochs, despite evidence that not all samples contribute equally throughout learning. Recent work shows that progressively reducing the amount of training data can improve efficiency and generalization, but existing methods rely on fixed schedules that do not adapt during training. In this work, we propose Adaptive Data Dropout, a simple framework that dynamically adjusts the subset of training data based on performance feedback. Inspired by self-regulated learning, our approach treats data selection as an adaptive process, increasing or decreasing data exposure in response to changes in training accuracy. We introduce a lightweight stochastic update mechanism that modulates the dropout schedule online, allowing the model to balance exploration and consolidation over time. Experiments on standard image classification benchmarks show that our method reduces effective training steps while maintaining competitive accuracy compared to static data dropout strategies. These results highlight adaptive data selection as a promising direction for efficient and robust training. Code will be released.

In generative semantic communication, semantic tokens guide receiver-side generative models to synthesize high-dimensional content. In challenging network environments, however, frequent token erasure distorts the conveyed semantics beyond what receiver-side recovery can restore. In this paper, we propose a token encoding framework (TokCode) for robust semantic recovery, achieving erasure resilience by restructuring redundancy in the semantic domain. TokCode uses a lightweight adapter to recast a general-purpose large language model (LLM) at the transmitter into a token encoder, exploiting the LLM's pretrained semantic prior to avoid introducing a dedicated deep model. To optimize the adapter efficiently and make it applicable across diverse channels, we develop a channel-quality-aware distillation approach for token encoder training~(CADET). Using a differentiable sentence-level semantic surrogate, CADET tunes T5 foundation models into experts for distinct erasure rates and distills them into a single reconfigurable low-rank adapter, enabling subsequent reinforcement learning (RL) to start above the plateau where direct RL stalls. Simulation results on token-based generative image transmission show that TokCode improves the image-level similarity over the best-performing receiver-side recovery benchmark by 14.1%–22.4%, closing 71.9%–76.5% of its gap to the erasure-aware oracle encoding, when only 20% to 50% of the tokens survive.