Papers

We propose a method for choosing the shared memory parameter $\beta_1=\beta_2=\beta$ in Adam from a short pilot training. The selected $\beta$ remains fixed during the subsequent full training. A local model of Adam's normalized direction balances sampling variability against the delay introduced by averaging past gradients. This balance gives a cubic memory rule, whose two coefficients are estimated from gradient probes at a few pilot checkpoints. The estimator uses the numerator and denominator jointly, preserving their covariance. With a 200-update pilot and sixteen probe gradients at each of four checkpoints, a seed-matched retrospective evaluation on eleven vision and language workloads reduces mean relative validation gap by 40.7% and worst-quarter mean gap by 44.3% against the grid representative of shared $\beta=0.95$. The mean gap is also 32.3% lower than that of the best constant $\beta$ chosen across all eleven workloads.

Traditional approaches for automated bug detection in video games, such as manual testing, can be beneficial for the improvement of quality assurance, but they can be expensive and time-consuming. The scarce number of tools available to detect multiple perceptual bugs in the same video frame introduces detection challenges for automated bug detection tools in real-world scenarios. We propose a deep learning model for multi-label perceptual bug detection and compare it against video classification models such as Inflated 3D ConvNet and 3D ResNet. Our proposed model, ResNet-BiLSTM, achieved an F1 score of 85.78% on the benchmark dataset. Our results demonstrated that temporal dependency modelling is beneficial for accurate video-based bug detection. We believe this work with multi-label perceptual bug detection on gameplay videos will help save resources spent on manual testing workloads in video games. Furthermore, we introduce a new dataset with multi-label perceptual bugs in this work. The dataset contains 77,969 video clips across different genres of games with approximately 1.2 million frames, containing combinations from 5 classes of bugs in the same video frame.

CNet is a C++/CUDA framework for deep complex-valued neural networks (CVNNs), trained by gradient descent with explicit Wirtinger (CR-calculus) derivatives. Complex models are underexplored yet natural where data is intrinsically complex – RF/IQ, MRI k-space, radar/SAR, audio spectra – and phase carries information real networks discard. CNet is physics-native: a network is a cascade of complex (often unitary) operations on an amplitude vector, and classification is a Born-rule measurement p_k = |z_k|^2 / ||z||^2 rather than a softmax over real logits. Every layer ships a CPU reference and a finite-difference-verified CUDA kernel; the static, batch-cloned computation graph captures cleanly into CUDA Graphs, outperforming eager and compiled torch.complex64 on an H100. We present three studies. (1) A fully complex causal language model, built on a new causal (triangular-masked) Fourier mixer with an O(N log N) Bluestein evaluation of the forward pass and its Wirtinger adjoint, matches a parameter-matched real causal FNet on character-level tiny-shakespeare in under half the steps. We introduce Born-rule attention, a content-based causal mixer whose weights are the quantum-measurement probabilities |<Q_k, K_j>|^2 – to our knowledge the first classical, complex-valued attention to replace softmax with the squared-magnitude Born score; with rotary embeddings and per-token complex normalization it reaches 1.52 nats/char, matching softmax self-attention, and transfers unchanged to byte-level (1.31 nats/char) and sub-word WikiText-103. (2) On radio-modulation classification (RML2016.10a/RML22) we trace the pooled-power accuracy ceiling to discarded temporal structure and recover it with a learnable cyclostationary conjugate-lag operator. (3) A bottleneck analysis of the Fourier phase problem in coherent-diffraction imaging isolates where complex-valued networks still need new operators.

Transformers provide a state-of-the-art modeling framework, yet poor calibration limits their reliability in safety-critical applications. A promising direction addresses this issue by interpreting attention as a Gaussian process (GP) posterior, which enables principled uncertainty calibration but incurs cubic complexity in sequence length due to the inversion of the kernel; although decoupled GP variants reduced the cost to quadratic, the computation remains prohibitive in practice. In this paper, we propose the plug-and-play random Fourier feature Gaussian process attention (RFF-GPA) module, which represents the attention as a GP with a stationary kernel approximated by random Fourier features. This low-rank approximation results in linear-time complexity for approximating the posterior mean and variance, making it far more scalable compared to previous work. Empirical results on multiple real-world datasets show that our attention module improves calibration while maintaining predictive accuracy, and simultaneously reduces computational complexity to linear in the sequence length.

While unlearning seeks to negate undesired capabilities acquired through learning, little research has examined how the way models learn shapes their subsequent unlearning. In this paper, we investigate this connection from the perspectives of memorization and generalization, the two most representative yet competing strategies that models employ during training. We first classify memorization- and generalization-heavy models using grokking in modular addition and compare their responses to unlearning, showing that the latter suffer greater retain damage, i.e., a larger performance drop on the retain set. Furthermore, we conduct a finer-grained analysis by introducing bucketed modular addition, in which the respective contributions of the two strategies can be explicitly controlled across the memorization-generalization spectrum. In this setup, we reaffirm that the same trend persists and is nearly monotonic. We further demonstrate that this relationship also holds in LLM unlearning across verbatim and factual recall settings. Finally, we provide two practical insights for developing better unlearning methods, highlighting the importance of accounting for learning dynamics in unlearning.

Skin cancer is a major global health concern, and early detection and accurate lesion delineation are important for effective diagnosis and treatment planning. Automated skin lesion analysis can assist dermatologists, with lesion segmentation serving as a fundamental step in computer-aided diagnostic systems. Conventional deep learning-based segmentation models typically rely on centralized training, where images and their corresponding segmentation masks are collected on a central server. Such data aggregation raises privacy concerns in medical applications and requires substantial centralized computational resources. To address these limitations, we investigate the feasibility of federated learning for privacy-preserving skin lesion segmentation. The training and validation sets of the ISIC 2018 Skin Lesion Segmentation Challenge dataset are used to simulate a distributed learning environment and develop a federated segmentation model. The resulting model is evaluated on the ISIC 2018 test set and the PH2 dataset to assess its performance and generalizability. Experimental results demonstrate that the federated model achieves performance comparable to centralized training while consistently improving upon the locally trained models. These findings demonstrate the potential of federated learning for collaborative skin lesion segmentation without requiring centralized aggregation of medical images.

Retrieval-Augmented Generation (RAG) systems are vulnerable to prompt-injection attacks embedded in retrieved content. We introduce RAG-PIBench, a benchmark for RAG-style prompt-injection detection containing 4,876 contextual examples across frozen train, validation, and protected-test splits. Using a leakage-aware construction pipeline and strict evaluation protocol, we compare keyword-based, semantic-reference, TF-IDF, and transformer-based detectors. DistilBERT achieves the best protected-test performance (F1 = 0.896, PR-AUC = 0.968), while TF-IDF SVM and logistic regression remain competitive. Our results demonstrate the value of leakage-aware benchmark design and strong sparse baselines for reliable prompt-injection detection in RAG systems.

Handcrafted preprocessing is widely employed in automated dermoscopic analysis to suppress imaging artifacts and enhance lesion visibility. Nevertheless, its actual contribution to modern real-time models remains unclear, particularly when evaluation protocols do not adequately control correlations among images of the same lesion. This study presents a leakage-controlled, lesion-disjoint evaluation of dermoscopic preprocessing and augmentation for joint multi-class lesion classification and instance segmentation using a fixed nano-scale YOLO26 segmentation model (YOLO26n-seg). From HAM10000 (10,015 images), quality control yields 10,013 valid image-mask pairs from 7,468 unique lesions, partitioned into mutually exclusive sets by lesion identity. With the architecture, resolution, training budget, and evaluation protocol held fixed, we compare minimally processed images plus online augmentation against offline class balancing, DullRazor-CLAHE preprocessing, and raw-processed hybrid views, over three random seeds. On the lesion-disjoint test set, the raw baseline achieves a mask mAP$_{50:95}$ of $0.5636 \pm 0.0234$, a Dice score of $0.9356 \pm 0.0024$, and a macro-F1 score of $0.6917 \pm 0.0202$. Offline augmentation does not improve the mean performance, while the combined and hybrid strategies reduce both class-aware segmentation and classification accuracy. At only 2.69 million parameters, the model runs at approximately 50 frames per second. Under a leakage-controlled, lesion-disjoint protocol with all non-input factors held fixed, minimally processed dermoscopic images combined with standard online augmentation deliver a better accuracy-efficiency trade-off than increasingly complex deterministic preprocessing, which yields no consistent joint benefit across three seeds on HAM10000.

This article is a geometric rediscovery of the singular value decomposition, with a further claim: the construction it builds is the machinery behind much of machine learning. The same argument that answers an idle question about ellipses is the algorithm behind principal component analysis, kernel methods, and PageRank, and it is not only the results that transfer but the proofs themselves, run as procedures. The usual introduction states $A = U\Sigma V^T$ and justifies it via the spectral theorem applied to $A^T A$. This is correct but unilluminating, since it assumes a powerful theorem to reach a result that is, in the end, about ellipses. Part I reverses the order. A linear map sends the unit circle to an ellipse; one asks which input directions map to its axes, and finds, example after example, that they are perpendicular. In the plane this can be watched: rotate a frame, track how far its images are from perpendicular, and a sign change forces a frame where they are exactly perpendicular, which is also where the map stretches hardest. Maximizing the stretch and recursing generalizes this to n dimensions, with singular values falling out in order, and the construction proves the spectral theorem rather than assuming it. Part II puts each construction to work: maximize-and-recurse becomes the power method and PageRank; the lemma locating the maximizer becomes the stopping rule of gradient descent; the duality between $A^T A$ and $A A^T$ becomes the transport at the heart of kernel PCA. Each connection is stated with its boundary, saying what the decomposition supplies and where another idea takes over. Prerequisites are the standard sophomore sequence, and the worked examples are small enough to check by hand.

Tabular foundation models (TFMs) can classify the nodes of a graph without training on it, by reading node and neighborhood features as table rows next to labeled context rows. Work in this line reports predictive performance, not conformal coverage or prediction-set size. To our knowledge, we give the first reliability study of the setting, with TabICL as the TFM and half of each graph as labeled context. As for any predictor fixed before calibration, a frozen in-context predictor makes split conformal prediction exactly valid in finite samples, with no training, validation fold, or tuning on the target graph. An audit across ten graphs then shows that the training-free TabICL posterior has lower expected calibration error (ECE) than GCN with temperature scaling (GCN+TS) on nine of them. Its mean ECE over the ten graphs is 0.019, about 35 percent below the 0.029 of GCN+TS. We also introduce HG-DAPS, a training-free diffusion score whose homophily gate reads only the in-context labels, so the guarantee still holds. Relative to adaptive prediction sets (APS), it reduces mean set size by 5.8 to 17.1 percent on six homophilous graphs and changes it by under 1 percent on four heterophilous ones. On two binary, class-imbalanced graphs, a pre-registered trap case shows that gating on raw rather than adjusted homophily lowers coverage among low-homophily nodes by 0.27 and 0.12. Marginal coverage stays at the nominal 0.90 and masks this drop.

Reinforcement learning (RL) has become a standard tool for post-training language models on reasoning tasks, where the policy is updated by reward feedback while exploring the space of responses. Despite its empirical success, theoretical understanding of RL post-training remains limited, in particular of why on-policy exploration combined with a neural reward model is effective. In this paper, we address this question by modeling the reward as a hierarchical function on the response space: the reward consists of infinitely many local components, each of which becomes relevant only after the preceding ones have been resolved. We show that a natural Transformer-based actor–critic algorithm, which alternates between sampling from the current KL-regularized policy, fitting a Transformer critic to the observed rewards, and updating the policy, achieves the minimax optimal rates in the query budget and in the regularization strength up to logarithmic factors, and is minimax optimal for a fixed number of prompts. In contrast, we prove that sampling from the fixed reference distribution, as in offline reward modeling, can limit regret decay to a logarithmic rate. These results show that on-policy exploration progressively zooms in on the region where the reward is concentrated, and quantify its benefit for RL post-training.

Streaming video-language models must decide not only what to answer, but whether the evidence needed for the current question has arrived. Existing systems learn that decision as a separate trigger; we ask whether an unmodified model already computes it. We show that frozen VideoLLMs carry a linearly readable evidence-readiness signal, labelled from timestamped evidence rather than from model output. It decodes in all seven models of a shared byte-identical evaluation (AUROC 0.733-0.905 under the strictest not-ready sampling, where a fitted clock is near chance), and a probe fitted without any of a benchmark family's footage still reads that family. It is question-conditioned: on byte-identical windows, changing only the question reverses the readout on 66.1% of pairs, while every question-blind control is at chance by construction. The model can answer incorrectly and still encode readiness: AUROC remains 0.722 among wrong answers. Readiness also beats uncertainty estimators and their supervised combination on latency-matched answer selection, and tracks independent human judgments more closely than confidence. Released streaming triggers are also linear readouts, yet a trained trigger read on its own base model's activations is approximately orthogonal to readiness and decodes it far less accurately than a probe. We turn the readout into Readiness Gating, an answer-timing policy that improves accuracy by up to +9.75 pp at matched video duration with negligible computational overhead. How much it gains varies with the accuracy headroom the task makes available: across 26 configurations the gain tracks that headroom, and an intervention that moves it over identical pixels moves the gain with it.

Activation steering has gained popularity as a lightweight inference-time debiasing technique for large language models. However, prior work reports that steering vectors generalise poorly, with unintended effects on model performance and limited transfer to new datasets. Our work analyses what the debiasing direction used for activation steering actually encodes, in order to shed light on its inconsistent performance. We study the linear debiasing direction obtained by contrasting the activations of anti-biased and biased prompts, and evaluate it as a steering intervention across bias and general knowledge benchmarks. We find that this direction is dominated by model confidence, pointing from regions of high to low-probability tokens in activation space rather than encoding a meaningful representation of model bias. Steering along it does reduce measured bias, but this is a consequence of reducing model confidence: on QA benchmarks we find that this steering drives the model to abstain from answering, with a side effect of improving fairness metrics. Our experiments show that model confidence is the dominant separating factor between biased and anti-biased prompts in hidden space, indicating that isolating a linear representation of bias which is disentangled from model confidence is difficult and steering-based debiasing results should be interpreted with care. In short, steering appears to reduce bias, not by correcting the model's underlying preferences, but by making it less confident, even on tasks unrelated to bias.

While HCI increasingly examines AI-safety for youth, the literature lacks a comprehensive view of what risks have been identified, how they are addressed, and whether proposed protections work in-practice. We systematically reviewed 100 empirical HCI studies involving children and youth interacting with or exposed to AI across schools, homes, care settings, and public services. Using the YAIR taxonomy for risks and the MIT Mitigation Taxonomy for countermeasures, we map which risks have been identified, whether each risk is addressed by countermeasure(s), and whether each countermeasure for that risk is implemented and even evaluated. The risk-countermeasure mapping shows that most risks are matched only with proposed/ideated countermeasures; few countermeasures have been implemented, and fewer still evaluated; and existing evaluations often measure technical performance rather than protection from harm. We identify where coverage is absent, where safeguards remain untested, and propose concrete directions for HCI research to strengthen youth AI-safety.

Sequential test-time training adapts a memory network through successive updates, each computing an inner-loop gradient based on the network's previous state. Intuitively, this state dependence should allow each update to account for what the memory has already learned and better incorporate new information. However, we find that this expected advantage does not consistently materialize in nonlinear memories: a fixed-base parallel TTT baseline outperforms its serial counterpart. Our exploratory experiments point to a key underlying difficulty: nonlinear memories can be harder to optimize than linear ones within a single pass over the sequence. To alleviate this optimization difficulty, we introduce DeltaTTT, which replaces joint inner-loop optimization of a two-layer memory network with layerwise learning. Each layer is assigned a local prediction target and updated through a state-dependent delta rule. This formulation retains a nonlinear readout while enabling chunkwise parallel computation. Experiments on DeltaNet and LaCT backbones show improvements in language modeling and retrieval over their recurrent baselines.

Concept bottleneck models (CBMs) make predictions inspectable and intervenable by routing them through human-interpretable concepts, but originally required concept annotations. Annotation-free variants remove this requirement, but typically use large concept vocabularies, static at both training and inference, producing bottlenecks larger than any task or prediction needs and harder to inspect. We propose AnyBottle, a single recipe for building compact, task-specific CBMs. AnyBottle assumes only a frozen backbone and an unsupervised concept pool, such as a sparse autoencoder. A black-box teacher trained on the same backbone then guides selection: each round adds the concept that best explains the bottleneck's current failures, with candidates restricted to regions of teacher/student disagreement. Trained with nested dropout over this selection order, the final bottleneck predicts accurately from any concept prefix, so inference spends fewer concepts on inputs it is confident about early and more on hard ones. Since no stage is modality-specific, a new domain and task requires swapping only the backbone and concept pool. Across six vision and two text datasets and two teacher paradigms, AnyBottle yields bottlenecks with fewer concepts and higher concept consistency than annotation-free baselines, while staying close to the black-box reference. Overall, AnyBottle shows that going annotation-free need not mean going large: a small, discovered vocabulary can be as expressive as a much larger, fixed one.

Whether alignment gets easier or harder as models grow is often argued from isolated findings, as if alignment were one property. We treat it as a family of measurable scaling relations: for each risk category r, the alignment burden needed to hold a fixed safety target is modeled as B_r(N)=a_rN^alpha_r, with N a capability proxy; against a budget proportional to N, scaling helps if alpha_r<1, keeps pace if alpha_r~1, and accumulates alignment debt if alpha_r>1. We give three operationalizations of burden and distinguish observed, audited and true alignment. A toy model, in which corrections consume capability headroom, makes the consequences explicit. We prove that the largest exponent among corrected risks, not an average, sets the long-run regime; that above 1 any policy holding headroom above a floor must grow super-exponentially; that, for burdens that are positive mixtures of power laws, fits on small models underestimate large-scale exponents; and that an audit that uncovers hidden failures without false positives never underestimates true alignment. We propose a pre-registrable protocol and apply reduced versions of it twice. A preregistered reanalysis of public adversarial-training data for Pythia classifiers finds that the compute needed to bring attack success under 10% grows as N^0.60. A preregistered pilot on Qwen2.5 0.5B-72B finds exponents of -0.05 for truthfulness and 0.48 for stated dispositions (both scaling helps under its reduced rule, though local slopes approach 1 at the top; replicated on Qwen3 0.6B-14B), while sycophancy (0.89, or 0.83 with two seeds added at 72B) and a planted backdoor are undetermined: the backdoor is removed quickly when its trigger is known but survives blind safety training at four of five sizes. We release four browser games that play these laws (www.aisafety.fun). We make no claim about which regime holds for current frontier models.

Probabilistic load forecasting has been widely studied for power-system operation and planning, but customer- and transformer-level forecasting introduces a distinct scalability challenge. At these levels, load uncertainty is strongly affected by customer behavior, weather, and mixed load composition, making it difficult for a single shared model to capture heterogeneous patterns. Using separate probabilistic models can improve local accuracy, but becomes costly to train, store, update, and validate at scale. To address this challenge, we develop a scalable customer-aware forecasting framework that learns common demand behavior through a shared model while adapting only a compact subset of parameters. Rather than using an independent model for each load or assigning each load to a specialized model, the proposed design learns a small bank of low-dimensional adaptation components and allows each load to combine them according to its forecasting characteristics. This preserves shared knowledge across customers while providing sufficient flexibility for heterogeneous and mixed load compositions. Experiments on 590 load profiles from the SMART-DS dataset show consistent improvements in deterministic accuracy and probabilistic quality over statistical, neural-network, Transformer-based, and pretrained time-series baselines, while retaining low storage and inference costs.

In the field of Explainable AI (XAI), counterfactual (CF) explanations interpret a model's decision by suggesting the changes to the input that would lead to a more favourable outcome. To be useful in practice, such an explanation should change few features and change them as little as possible, properties known as sparsity and proximity. We observe that existing methods remain limited in this respect, especially for numerical features, whether they are model-agnostic and amortised, or gradient-based with full access to the model. In this paper, we propose FlowCF, a model-agnostic generative method that frames CF generation as sparse transport from the factual to the target class. We solve this transport with flow matching, which we extend to mixed feature types with a novel mixed flow operator, and exploit the resulting geometry to optimise for sparsity through a gating network that minimises the number of features the transport changes. Extensive experiments on six benchmark datasets demonstrate that FlowCF produces the best numerical sparsity and proximity, changing 29% of the numerical features where the best baseline changes 89%, at 70% smaller displacement, while remaining comparable on the other desiderata.

Understanding the principles behind Shampoo has recently guided the development of more effective neural network optimizers. These methods learn a preconditioner by optimizing the Frobenius or Kullback-Leibler (KL) divergence against the gradient second moment. In this work, we investigate how the choice of divergence shapes preconditioning, which remains unclear and blocks further improvements. To do so, we develop a unified Bregman divergence framework that connects all popular divergences, allowing us to study them jointly. Through empirical spectral analysis of gradient second moments, we examine how divergence choice shapes Kronecker approximation and interacts with finite-sample error in preconditioning. We find that some divergences can better compensate for finite-sample underestimation of the empirical second moment, helping explain the differing behavior of their corresponding Shampoo variants. We further validate this explanation through GPT-2 pretraining experiments. By connecting divergence choice to practical training behavior, we believe our framework provides principled guidance for understanding the foundations of, and further improving, Shampoo.

Geometric alignment scores based on Gram determinants provide a compact way to model higher-order consistency among modalities, yet how such scores respond to modality degradation is poorly understood. This paper asks whether the response of a multimodal geometric score is determined primarily by the magnitude of the perturbation-induced displacement. Using frozen cohorts from MSR-VTT (N=878) and DiDeMo (N=980), we apply controlled video blur and audio noise and analyze the response in the relational geometry on which the score is defined. Displacement magnitude explains at most 15% of the out-of-sample variance in the absolute response, and magnitude-matched pairs respond systematically differently, so scalar magnitude does not organize the response. The closed-form first-order expansion of the Gramian volume yields the Directional Geometric Response (DGR): the projection of the displacement onto the local volume gradient, which jointly captures the clean operating point, displacement magnitude, and displacement direction. The absolute first-order DGR term explains the observed response with out-of-sample R^2 of 0.838-0.969, matched-magnitude ranking accuracies of 0.864-0.963, and response-sign accuracies of 0.909-0.989, whereas the tested direction-free alternatives remain weak or unstable under the corresponding evaluation protocols. A pre-specified gain-normalization candidate, V/(g_V+eps), fails its predictability and clean-order gates. DGR uses the observed degraded-state displacement and is therefore an explanatory quantity, not a deployment-time predictor: geometric response depends on where the representation operates, how far degradation moves the relational geometry, and in which direction it moves.

Hybrid architectures combining linear sequence models with softmax attention provide an effective balance between efficient long-context modeling and precise token retrieval. Existing designs such as Native Hybrid Attention (NHA) combine compressed long-term states with sliding-window attention, but their exact attention is restricted to a fixed local window. In this work, we introduce Prefix-State Hybrid Block Attention (PHBA), which replaces local sliding-window attention with top-k block-sparse retrieval and couples each retrieved block with a compact prefix state summarizing its preceding context. The prefix states are constructed by a gated linear recurrence at block boundaries and retrieved together with the corresponding token blocks, allowing the model to combine precise long-range evidence with compressed historical context within a unified layer. We further develop a hardware-aware Triton implementation that streams routed token blocks and prefix states without materializing large intermediate tensors. Experiments show that PHBA improves long-context and retrieval performance over strong linear and hybrid baselines while retaining efficient training and inference.

Proactive power management systems reduce processor dynamic power through runtime power prediction and power-aware scheduling. Accurate, stable and low-overhead digital on-chip power meters (OPMs) are crucial for improving the prediction quality. Recent studies have explored various modeling methods, including using linear models, decision trees, and multi-layer perceptrons (MLPs) to construct OPMs. However, most current approaches train models end-to-end without analyzing the physical interpretability of features, affecting their ability to generalize to unseen workloads. Grounded in the design principles of synchronous digital VLSI circuits, X-OPM introduces a robust feature engineering framework that uses tree-based models to capture feature interactions and linear models for prediction. It also incorporates a human-in-the-loop workflow to balance model accuracy against modeling effort. Evaluated on a commercial C906 vector processor, X-OPM consistently achieves $R^2 > 0.93$ across all workloads with sampling window size set below $8$ cycles. In contrast, state-of-the-art methods including APOLLO, COBIT, and standard MLPs fail to generalize across all test cases. Layout with commercial EDA tools shows that X-OPM incurs an area overhead below $0.1\%$, which is on par with lightweight tree-based and linear models, and significantly smaller than MLP-based models.

Machine learning can accelerate molecular discovery by designing molecules and planning experiments. However, many scientific challenges demand molecules with very rare properties, and in this sparse setting, existing algorithms offer little gain over random guessing. We propose a method to efficiently search large regions of molecular space using algorithmically controlled stochastic synthesis. Rather than design, make and test individual molecules, we design and make complex mixtures, test them as a pool, then deconvolute the molecule-activity map. We optimize synthesis to encode maximal information. Theoretically, this approach can reduce the number of experiments required to find the optimal molecule among $d$ candidates from $\mathcal{O}(d)$ to $\mathcal{O}(\log d)$ or $\mathcal{O}(1)$. In simulation, on estimated protein fitness landscapes, it finds active molecules with an order of magnitude fewer experiments than existing Bayesian optimization methods.

Few-shot meta-learning traditionally formulates task adaptation either as analytical gradient descent through unrolled computational graphs or as metric-based distance comparisons over flattened 1D fea- ture vectors, which either incur costly test-time backpropagation or discard native 2D spatial geometry. In this work, we propose METALEARNNCA, a decentralized framework that achieves few-shot adapta- tion through the dynamical interaction of coupled Neural Cellular Automata (NCAs) without computing analytical gradients during inference. MetaLearnNCA decomposes task adaptation into an Active- NCA, which executes task inference conditioned on a continuous 2D spatial memory grid termed the spatial program, and a learned Meta-NCA, which acts as a decentralized cellular optimizer by diffusing spatial error residuals across local neighborhoods to dynamically update this program. METALEARN- NCA is competitive against canonical meta-learners in-distribution (96.12% on Omniglot) with Out-Of- Distribution transfer gains on MNIST, KMNIST, and Fashion-MNIST transfer across 10 independent testing seeds across 1-, 5-, and 10-shot regimes (e.g., surpassing Prototypical Networks by +10.54% on 10-shot MNIST and a +3.87% gain on 10-shot Fashion-MNIST over FOMAML). Our results establish that robust, gradient-free learning-to-learn can emerge from decentralized cellular dynamics on non-von Neumann substrates.

In many-query scenarios, data-driven surrogate models provide an efficient alternative to high-fidelity solvers for simulating physical systems governed by Partial Differential Equations (PDEs). In this context, the Latent Dynamics Network (LDNet) has recently demonstrated remarkable performance in predicting the response of spatio-temporal systems, combining Neural Ordinary Differential Equations with nonlinear dimensionality reduction. However, the original formulation assumes a fixed initial condition, limiting its applicability to many real-world applications where a system evolves from varying starting states. In this work, we overcome this limitation while keeping the end-to-end training procedure of the original LDNet and its encoder-free nature, which preserves its intrinsic independence from spatial resolution and grid topology. We infer the initial latent state directly from a small set of early-time observations, treating latent-state initialization as an adaptation problem, and investigate two strategies: an auto-decoding formulation and a meta-learning approach in which the initial latent state acts as a task-specific context variable. We demonstrate the accuracy of the proposed methods across diverse physical phenomena, spanning advection-diffusion, fluid dynamics, and solid mechanics. Meta-learning markedly accelerates latent-state inference and induces smoother, better-conditioned optimization landscapes, and spontaneously organizes the latent space into a structured representation that reflects physically meaningful features of the underlying dynamics. The coordinate-based decoder enables training from spatially subsampled data while recovering high-resolution solution fields at inference. The resulting approach provides an efficient and resolution-independent surrogate modeling framework for many-query simulations of time-dependent PDEs with varying initial conditions.

Long-context inference and Retrieval-Augmented Generation (RAG) handle evidence selection at vastly different scales, from a single long prompt to an entire corpus. We ask whether a single model-internal mechanism can select evidence across this range. We introduce UNifying REtrieval And Long-Context with a Single Model (UNREAL), a model-native evidence selection framework to span corpus retrieval and long-context inference. UNREAL encodes chunks and derives retrieval queries directly from the frozen LLM's internal representations. It adds fewer than 500K trainable parameters and leaves the backbone unchanged. On a 3B-token, 21M-chunk Wikipedia index, all four dense and hybrid UNREAL backbones outperform state-of-the-art retriever-reranker systems. The best model raises recall from 49.1% to 73.2% on HotpotQA, from 31.7% to 60.1% on 2WikiMultiHopQA, and from 8.8% to 14.4% on MuSiQue. Applied to long-context tasks, the same selection mechanism removes distractors before generation, raising NoLiMa accuracy from 1.0% to 24.83% at its maximum context length of 128K tokens, and LV-Eval's F1 score from 49.97% to 54.66% at 256K. UNREAL also reduces FLOPs and time-to-first-token relative to full-context inference from roughly 32K tokens onward, with larger gains as context grows. Together, these results establish model-internal evidence selection as a common foundation for corpus retrieval and evidence-sparse long-context inference.

Retrieval-augmented generation (RAG) is a widely used approach for grounding large language models (LLMs) in external knowledge. However, configuring a pipeline is an expensive hyperparameter optimization problem over many interacting choices, from chunking and embedding model to reranking and generation. Existing optimizers, from greedy search to Bayesian optimization, reduce each trial to an aggregate score and search without modeling why a configuration performed as it did, even though the retrieved chunks already provide evidence about whether each failure occurred during retrieval or after it. We introduce Agentic AutoRAG, an LLM-agent optimizer for multi-objective RAG hyperparameter optimization with retrieval-versus-generation failure attribution. It proposes configurations scored on a frozen exam from the corpus: after each trial a Diagnoser attributes each failed question to retrieval or generation, and a Proposer, grounded in a knowledge base of model rankings and pricing, selects the next configuration, weighing accuracy against cost to trace a Pareto frontier. On three multi-hop QA benchmarks it reaches higher LLM-judge accuracy than every baseline we compare, and within its first 10 trials it matches or beats the statistical baselines' full 30-trial judge accuracy. In its cost-aware mode on a real-world healthcare corpus it reaches a median exam accuracy of 77%, above the strongest baseline's 71.5%, at about 58% of that baseline's cost per query, and it matches that 71.5% at about 22% of the cost.

We establish a polynomial sample complexity separation between symmetry-aware and symmetry-agnostic feature learning. We study growing-rank multi-index models with high-dimensional Gaussian covariates in $\mathbb{R}^d$ and $r=\Theta(d^\delta)$ teacher directions forming a cyclic symmetry orbit, where $0<\delta<1/2$. We compare three ways of exploiting this structure: architectural weight sharing, data augmentation over the full symmetry group, and learning without access to the symmetry. In particular, we analyze a symmetry-tied convolutional network, an untied network, and the same untied network trained with full-group data augmentation, using spherical online SGD with correlation loss. For a class of polynomial links with information exponent $p\ge3$, we prove matching sample complexity bounds up to logarithmic factors: the tied and augmented learners achieve weak directional recovery in $\widetilde{\Theta}(d^{p-1})$ samples, whereas the symmetry-agnostic learner requires $\widetilde{\Theta}(rd^{p-1})$. For the pure quadratic Hermite link, the same separation holds for weak recovery of the teacher subspace, with sample complexities $\widetilde{\Theta}(d)$ and $\widetilde{\Theta}(rd)$, respectively. Thus, full-group data augmentation matches the sample efficiency of architectural weight sharing, and both provide a polynomial advantage over training without symmetry. For $p\ge3$, the proof reveals a two-stage mechanism: fluctuations at initialization select one direction in the teacher orbit, after which localized growth amplifies its overlap to the weak recovery scale while competing overlaps remain near their initialization scale.

Large language models routinely answer questions that cannot be answered from the information given, and in dialogue they answer before enough has been said. Unanswerability is linearly decodable from hidden states, but it is unclear which of its forms share a representation and whether the signal is useful in dialogue. We contribute a turn-labeled multi-turn benchmark (423 conversations, 1,661 labeled turn-states) and an evaluation harness with a simulated user who answers clarifying questions, and use them with six datasets and six open-weight LLMs to test how far probes for unanswerability carry. Probes transfer robustly between datasets that share a ground of unanswerability: missing information in math (AUROC 0.77-0.97) and in a passage (SQuAD 2.0<->MuSiQue, 0.77-0.90). Probes for epistemic "known-unknowns" transfer poorly to math, but this separation weakens under lexical controls and changes with layer and coordinate system, so it remains unresolved. Single-turn probes fail zero-shot to detect when a conversation becomes answerable; in-structure probes recover it, but no better than a bag-of-words classifier. A gate on the calibrated probe, with no model fine-tuning, fires on underspecified turns far more precisely than chance, and its end-task success comes within 0.08 of a gate given the true labels. Yet across four models it does not reliably beat vanilla generation or prompted consolidation. The remaining gap lies mostly in how models use a clarification, not in detection.