
ALT Recurrent memories let language models process long sequences with a fixed-size state. Linear memories such as DeltaNet and KDA store one matrix and take one delta-rule step per token. The product of a chunk of such steps has an exact WY form, so they train fast. Deep memories such as TTT-MLP and Titans store a small MLP, which can hold maps that no matrix can, but their update has no WY form. Toaster bridges the two by building on a known fact: a gradient step on a linear layer is a delta-rule write. It builds a two-layer nonlinear memory from two linear memories with an activation between them. We use target propagation to give the hidden memory its values by projecting the output error back through the second memory. Given a chunk's cross-layer signals, approximated once per chunk, each layer can then run on an unmodified delta-rule kernel. Our default runs this kernel on the output layer and updates the hidden layer once per chunk. Running it on both layers gives a better model in

ALT Clustering on attributed graphs can produce demographically biased clusters, particularly when communities align with protected groups. Graph transformers can amplify this bias through same-group attention, while global attention limits scalability. We propose FairPGT, a partitioned graph transformer for fair graph clustering. FairPGT restricts attention to METIS partitions and connects them through routing tokens, which divides the quadratic cost of attention by the number of partitions. A graph-aware self-training target guides clustering, with a provably bounded deviation from the standard DEC target. Fairness is imposed at two levels: a mutual-information term promotes demographic parity in cluster assignments, while a normalized cross-group attention ratio (nCGAR) guides attention toward group neutrality through self-distillation. We prove that the nCGAR correction moves attention toward neutrality and establish a condition preventing overshooting. Across seven benchmarks, FairPGT

ALT Overparameterized learners can fit noisy training labels perfectly and still predict well, which neither classical complexity measures nor information-theoretic generalization bounds explain. We develop an information-theoretic framework that separates two things an interpolating learner does with its sample: storage, the information the trained model retains about its training labels, and expression, the information about the sample that reaches its predictions at fresh inputs. We first characterize the ideal model: for a loss l, the least complex l-sufficient representation is the Bayes act, whose entropy is the minimal sufficient complexity; this generalizes, and corrects, an earlier "lossless maximal compression" view of boosting. We then show that every interpolator stores exactly nH(Y|X) bits of label noise, so storage, and bounds built on it (including CMI bounds), cannot tell interpolators apart. Expression can: excess risk is bounded above and below by the expressed informatio

ALT Agents can expand the covariate space of randomized experiments by retrieving fast-moving local context missed by slow-changing survey baseline covariates, helping characterize who benefits most from an intervention. Verifying such covariates is difficult: individual treatment effects are unobserved, retrieval leakage can pull post-treatment information into nominally pretreatment summaries, and encoder leakage can inject post-treatment pretraining knowledge into representations even after temporal filtering of inputs. Guided by a two-channel information bound separating summary provenance from encoder weights, we propose a three-layer benchmark protocol: (i) retrieval admissibility via pre-treatment date filtering; (ii) heterogeneity validation scored rank-weighted average treatment effects (RATEs); and (iii) encoder admissibility checked with chronologically vintaged encoders, future-fact probes, and weight perturbations. Across 11 datasets anchored by Uganda's Youth Opportunities Pr

ALT Wearable EEG inference is constrained not only by model size, but also by the channel-dependent sensing front end that must process high-dimensional inputs before compact downstream computation. We present EEG2MCU, a sensing-to-execution co-design framework that treats electrode count as a hardware design variable through task-aware montage selection. EEG2MCU retains a compact time-frequency representation and generates only the INT8 regions required by each first-layer tile, avoiding full input materialization. Halo reuse and access-aware placement across Flash, SRAM, and tightly coupled memory further reduce redundant generation and data movement. On a Cortex-M7 MCU, EEG2MCU reduces peak SRAM by 11.38x relative to the baseline, including a 1.71x reduction from the 62-to-19 electrode montage, while reducing latency by 2.57x and energy by 2.48x. It retains 94-96% of full-montage accuracy, demonstrating practical high-dimensional EEG inference under tight embedded-resource constraints.

ALT Diffusion modeling is a popular method of sampling from a complex, high-dimensional distribution p using learned approximations of its (noised) score functions. Generating a single sample requires simulating the reverse diffusion process, which requires many queries to the learned scores. Thus, a priori, the learning errors can easily compound. Indeed, in existing theoretical analyses, the sampling error (as measured in total variation distance to p) accumulates polynomially with the horizon of the stochastic process. This is a consequence of bounding the error of the output law by the error of the law of the entire reverse diffusion path. In this paper, we demonstrate that by circumventing the "path-space" analysis, the error accumulation of the diffusion sampler can be mitigated when p satisfies a PoincarΓ© inequality. Moreover, we achieve this result as a special case of a broader framework that applies to any distribution p, but relaxes the error metric from total variation distance

ALT How much the teacher talks and when are among the most informative and costly-to-measure signals of teaching. Yet, classroom audio is among the most sensitive data an AI system can process: recordings capture minors and fall under GDPR, preventing raw audio from leaving the recording device. We present a teacher-student segmentation pipeline designed to meet exactly this constraint: label-free at inference, running on a recording device, and requiring no per-session calibration. The pipeline chains a lightweight voice activity detector, compactly pretrained speaker embeddings, agglomerative clustering, and a small logistic regression that assigns the teacher role based on session-relative structure rather than voice identity. We show that our structural cue transfers across classrooms, languages, and acoustic conditions. In leave-one-teacher-out cross-validation on 292 recordings from 63 teachers across English and German corpora, our pipeline achieves teacher F1 of 0.81 and 0.87, resp

ALT Black-box optimization evaluates candidates through simulations, rollouts, or experiments, often without gradients. Learned optimizers are usually judged by whether they outperform existing methods. We ask a complementary question: what information does a learned optimizer use from its search history? We introduce JEPA-PS (JEPA Policy Search), based on a Joint-Embedding Predictive Architecture. It is pretrained on random rollouts using parameter vectors and scalar returns. Its latent-prediction objective predicts the next point in an optimization trajectory rather than its reward or gradient, allowing us to examine what was tried and how well it performed. At test time, JEPA-PS reads a short evaluation history and proposes a candidate in one forward pass, without task descriptions, reward shaping, or test-time gradients. We evaluate it on 15 control environments against 12 optimizers, including evolutionary methods, Bayesian optimization, CMA-ES, and an LLM optimizer. JEPA-PS is compet

ALT Path-based attribution methods like Integrated Gradients (IG) are widely used to interpret neural networks, yet imperceptible input perturbations can drastically alter their attribution maps while leaving model predictions unchanged. We introduce Entropy-Regularized Integrated Gradients (ERIG), a path-based attribution method that intrinsically attenuates gradient irregularities along the integration path using maximum-entropy softmax weighting with temperature parameter Ο. We quantify the effect of the entropy-regularized weighting on attribution stability by using the Lipschitz properties of the model and its gradients. We decompose the attribution change under an input perturbation into a model-response term, whose bound is Ο-independent, and a weight-variation term, whose bound decreases at rate O(1/(2Ο)). Their ratio R(Ο) characterizes the trade-off between gradient information and attribution stability. We further add a new dimension to attribution evaluation by adapting classica

ALT Online continual learning (OCL) requires learning from a data stream whose arrival rate is independent of model training speed. When training cannot keep pace, computationally expensive gradient-based methods update less frequently, limiting adaptation. Analytic methods reduce this overhead, but their frozen backbones cannot adapt representations to changing distributions. We propose DuoCL, a dual-path OCL framework whose fast analytic path accumulates feature statistics from every incoming minibatch while its slow path periodically adapts the backbone. The slow path adjusts its sample count to the available compute budget without interrupting fast-path learning. Feature alignment compensates for drift in the accumulated classifier statistics. Across four datasets and three computational budgets, DuoCL improves online accuracy by an average of 9.67 percentage points over the strongest gradient-based baseline in each setting. DuoCL consistently outperforms analytic learning under distri

ALT We study variance-dependent reinforcement learning in episodic time-inhomogeneous linear Markov decision processes (MDPs) with feature dimension d, horizon H, and K episodes. While recent works have established variance-dependent regret with Γ(1) deterministic-regime overhead for linear mixture MDPs (Zhao et al., 2023) and worst-case nearly minimax optimal regret Γ(dβHΒ³ K + dβ· HβΈ) for linear MDPs (He et al., 2023),, existing variance-aware analyses for linear MDPs (Li and Sun, 2023; Huang et al., 2023) rely on a coarse concentration step whose resulting regret guarantees retain a polynomial Γ(KΒΉ/β΄) additive term even in deterministic environments. We propose VA-LSVI-UCB++, a computationally efficient algorithm that achieves the variance-dependent regret bound Γ(dβ(Hmin{Var_KVar_KΟ})+dβ΅Hβ΄), where Var_K and Var_KΟ denote the cumulative conditional variances of the optimal value function and the played policies' value functions along the realized trajectory. Our result adapts to the minim

ALT In multi-agent systems, a receiver agent often waits for a helper agent's message (text or its latent KV cache), which can improve performance but adds latency. We study when a receiver agent can skip the message safely (the final answer rarely changes) and beneficially (the time saved exceeds the cost of deciding). Our mechanism reads the receiver agent's confidence without the message and uses a finite-sample test to certify a threshold at which, with high probability, at most 5% of skipped answers change; if none passes, the system keeps communicating. We test it on OpenBookQA, ARC-Challenge, and MMLU-Pro with three Qwen model pairs, on cross-family agents, and on SQuAD and GSM8K. Four insights emerge. First, oracle headroom is weak evidence that messages help: control inputs without answer-specific content, matched to change answers as often as real messages, reproduce a median 76% of the oracle's gain. Second, whether a message can be skipped depends on how often and how predictab

ALT Geometric convolutions on Riemannian manifolds depend on the geometry of the underlying domain, making stability under metric perturbations a fundamental question in geometric deep learning. This paper studies this question for parallel transport convolution through methods from differential geometry and analysis. On a fixed compact manifold, we establish that, for a continuous input signal and a continuous reference kernel supported in a sufficiently small normal neighborhood, the convolution output depends continuously in LΒ² on the Riemannian metric in the CΒ²-topology, under the specified localization and transport-path conventions. The analysis develops perturbation estimates for geodesics, parallel transport, Jacobi fields, and exponential maps, and uses these estimates to compare the transported kernels. A central difficulty is the cut locus, which obstructs uniform comparison of the transport construction over the entire manifold. We address this difficulty by combining uniform e

ALT Official development assistance fell by 23.1 per cent in 2025, and the decision of who is helped first when there is not enough for everyone is moving into software. A leading published machine-learning targeting system excluded 53 per cent of eligible households as deployed. If the rule of distribution ends up in code, it should carry the judgement of many cultures, and the machine should execute while a person judges. AllotNet is an interpretable mixture-of-experts conditional logit for the allocation judgements elicited by Allot, a pre-registered six-language instrument: eight linear taste experts over seven ethical criteria, weighted by a gate over participant context, with a damped residual, an encoder for household intake, an outcome head and a budget-constrained planner that builds a review queue and never rejects anyone. A model of human judgement has no answer key on real data, so we wrote one first: a planted synthetic population of 5.6 million sessions and 104 million respon

ALT Speculative decoding accelerates language-model generation by verifying multiple draft tokens in a single target-model pass. However, better draft predictions and longer accepted continuations do not necessarily produce faster generation. To examine this gap, we present FACT (Fixed-Budget Allocation and Cross-Step Training), an experimental framework for studying how drafter training, candidate allocation, and execution conditions jointly affect acceptance and latency. Using a lightweight feature-prediction drafter with a frozen Llama-3-8B-Instruct target under greedy decoding, we first examine how training translates into accepted tokens. Backpropagating through predicted-feature rollouts with greater weight on early steps improves acceptance by 15-24% over uniformly weighted detached training, yet average rollout proxy cross-entropy does not reliably rank drafters by acceptance. Beyond prediction quality, the placement of draft candidates also matters: with a fixed 256-node budget an

ALT Selecting less computation in a neural operator need not reduce inference time. We study this gap using conformal calibration to choose Fourier modes or channel width per input subject to a fixed error budget. A triangle-inequality envelope bounds all candidates simultaneously; a direct-score rule handles non-monotone candidate errors and, with an input-independent score, calibrates a static baseline under the same marginal guarantee. Under frozen-model and exchangeability assumptions, both rules bound the probability of issuing a prediction outside the budget and abstain when calibration is infeasible. On Burgers and Darcy flow across three training seeds, mode selection reduces active spectral support without a measurable latency reduction, because full-grid transforms and channel mixing dominate the cost. Narrowing channels does reduce Darcy forward cost, but reference-based selection must first compute the dense prediction. An input-only binary width selector avoids that pass and r

ALT We show that the importance of a LoRA adapter's initial input-subspace direction is regime-dependent, and we characterize this dependence in theory and experiment. This refines the recent view that spectral initializers help through update magnitude alone. Direction matters most when labeled data are scarce relative to the representation dimension and the activation subspace used for initialization captures supervised task structure. In this regime, the adapter's row span remains sensitive to its starting point, making task-aligned initialization especially useful. A spiked-covariance perturbation analysis characterizes the finite-sample stability of the estimated subspace. A first-step analysis of LoRA training explains why the subspace's alignment with the task matters and motivates an alignment coefficient Ο that closely tracks gains across classification tasks. We test this account with PivotLoRA, which initializes adapters from activation PCA at prediction-relevant positions and a

ALT Plasticity and stability in continual learning are typically studied at the entire-model level and in a task-agnostic manner, yet prior work has reported divergent conclusions across settings. Inspired by cognitive distinctions between declarative and procedural memory, we instead study how plasticity and stability vary across transformer modules and task types. In a controlled synthetic environment with clean measurement, we show a double dissociation: query-key (QK) and value-output (VO) projections are the most plastic on procedural tasks, whereas MLPs are the most plastic on declarative memorization. Stability, however, does not simply mirror plasticity: when learning a new procedure, updating QK retains over half of the stored declarative memory, while updating MLPs retains far less and updating VO loses roughly four-fifths. Targeted module reset-and-recovery experiments explain this asymmetry: plasticity tracks which module is best suited to a task's primary computation or storag

ALT Landmark diffusion (Roseland) computes a diffusion-map-style spectral embedding from the singular value decomposition of a degree-normalized data-to-landmark kernel matrix, at a cost far below that of full diffusion maps. It assumes, however, that all data are available at once. We introduce incremental Roseland (iRoseland), which casts the landmark affinity as a row-growing matrix whose SVD can be updated with a kernelized Sequential KarhunenβLoΓ¨ve (SKL) scheme. Degre correction is handled by a first order expansion about a reference mean kernel row and so that the landmark-side eigenbasis may be recovered from a single landmark-by-landmark eigendecomposition at each step. The presented method requires only a single pass over the data and runs in bounded-memory with state independent of the number of points processed. A forgetting factor lets the method track drifting distributions, and a sample-size-invariant embedding keeps outputs comparable across time. Experiments show agreement

ALT On-policy distillation (OPD) trains a student language model to match a teacher model based on its own generations. When using OPD, one must choose whether to apply it at the sequence or token level and which divergence to minimize. However, despite the wide adoption of OPD in practice, the formal understanding of which OPD objective is appropriate for a given setting is limited. Toward addressing this gap, we theoretically analyze how the choice of OPD objective governs which tokens are prioritized, in what order, and how efficiently. Our analysis reveals that sequence-level OPD exhibits a future agreement bias, which myopic token-level objectives avoid. This bias induces a reverse learning order, in which sequence-level OPD first matches the teacher at later token positions, while at earlier positions it can favor tokens that are improbable under the teacher, and thus often undesirable. As for the choice of divergence, we prove that it strongly influences the amenability of OPD to gr
