
ALT Machine learning models for chemical reaction prediction are usually evaluated by endpoint accuracy, but endpoint correctness alone does not reveal how a model represents chemical transformation or why it fails. We introduce UniRxnAxis, a unified forward-retrosynthesis framework that models both tasks as direction-conditioned endpoint redistribution in a shared latent space. By aligning source states, transformed states, and encoded target endpoints, UniRxnAxis enables directional auditing of learned latent updates. This shared latent geometry allows the learned displacement (from the encoder source state Sβ to the decoder-facing transformed state Sβ) to be decomposed into an endpoint-parallel component and an endpoint-orthogonal residual, which we call emergent steering. Our evaluation shows that emergent steering is functionally important: in retrosynthesis, it recovers most of the model's utility, remains effective under random, shuffled, and prototype controls, and directly affects

ALT A popular ideation recipe treats an embedding map of existing work as a guide: pick points far from everything and decode them into ideas. I test four ways such distances could guide ideation against what appeared later. (1) Far random directions are not text: no later item reaches a raw cosine above about 0.10 with one (after centring or whitening, 22 of 24 cells stay out of reach), and an LLM decoder shown a target's neighbours, with score-only feedback, writes those neighbours in all six target sets. A neighbour-free search writes coherent text at median cosine 0.57 and 0.47 to interior gaps, beating the nearest existing item on 15 and 14 of 20, but only 0.07 and 0.06 to far points. (2) On k-means cells of a past-only map, the registered hole score mostly tracks cell size and position; three density-based gap scores forecast which cells grow fastest at or below chance (AUC 0.30 to 0.66) and add nothing to momentum in Steam and arXiv, with patents inconclusive; more isolated new Stea

ALT Modeling large multi-agent systems poses challenges for trajectory inference and parameter identification, as network size and coupling increase computational demands and partial observations require reconstructing unobserved states. We propose MAS-PINN, a neural-network architecture with a tailored training algorithm for a physics-informed learning framework that leverages the network topology and the agents' similarity to reduce computational and memory costs of simulating and identifying multi-agent systems. Its architecture represents agent trajectories through shared temporal features and agent-specific coefficients. Training alternates between learning features, fitting coupled trajectories, and estimating physical parameters. A matrix-free update exploits local interactions to fit trajectories without forming large global matrices. Experiments on Lorenz rings, LotkaβVolterra networks, and RΓΆssler oscillators show improved trajectory and parameter accuracy over baselines, includi

ALT MegNIST is a benchmark dataset for developing and evaluating machine learning methods for inner speech decoding from magnetoencephalography (MEG). The dataset comprises 12,000 trials of a single participant imagining speaking the digits zero through nine, recorded across four sessions. Each trial contains 306-channel MEG recordings sampled at 250 Hz, preprocessed using signal space separation, filtering, and downsampling. Data are provided as pre-defined training (10,000 trials), validation (1,000 trials), and test (1,000 trials) splits, with entire experimental runs assigned to each split to preserve temporal independence. The dataset is available in both BIDS format for neuroimaging analysis and as serialised tensors for machine learning workflows. Baseline classification results demonstrate that imagined digits can be decoded significantly above chance. Standardised splits, validated baselines, example code, and a PyTorch-compatible loader in the pnpl Python library enable straightf

ALT Contrastive learning beyond two modalities now scores the whole tuple with multilinear products, volumes, and fusion heads, critics that can represent the joint distribution. In practice these critics still behave like bags of pairs. Their negatives break at least two of the three pairs, a single pairwise check rejects them, and when most samples can be solved by pairs the higher-order term receives almost no gradient. To solve this problem, we propose CHORD (Contrastive Higher-Order Residuals with Decoys), which trains a residual against decoys drawn from a calibrated pairwise model, negatives that the pairwise model finds plausible, so that rejecting them often needs all three modalities. Drawing decoys costs one inner-product query over a memory bank, and the optimal residual is the log-ratio between the data and the pairwise model, so the decoy loss also bounds what that model misses. On M3-Mix, where one in ten training samples needs all three modalities, our CHORD improves accura

ALT Federated Reinforcement Learning (FRL) provides a promising way to speedup training in reinforcement learning when one has access to multiple edge devices that can operate in parallel. Recently, it has been shown that even when these edge devices have access to different dynamic modelsβsuch as cars or robots with differing system configurationsβan optimal convergence rate that has a linear speedup proportional to the number of devices is achievable. However, this result requires that the algorithm's stepsize be chosen in a manner that depends on the unknown model parameters. Also, it applies only to a discounted setting, which has been argued to fit episodic tasks better than continuing control tasks. In this paper, we obtain finite-time bounds for heterogeneous FRL with average rewards. We show that optimal convergence rate and linear speedup are achievable even with a universal stepsize choice, independent of the underlying dynamics. To achieve our result, we modify the existing one-

ALT Generative model pre-training typically optimizes a model as a whole without explicitly organizing its parameters for selective reuse. Decomposable pre-training instead learns reusable components for downstream adaptation while retaining strong generative performance. However, existing knowledge diversion methods incur substantial parameter overhead through weight factorization and rely on fixed component assignments to induce specialization, potentially compromising in-distribution generation quality. We therefore propose Specificity-Guided Knowledge Diversion (SGKDiv), which learns how generative knowledge is shared and specialized for effective generation and selective downstream reuse. SGKDiv constructs Knowledge Units from paired feed-forward channels, preserving native computational structure without SVD-based reparameterization. Calibrated condition-wise routing combines a shared usage prior with conditional adjustments, while complementary regularization shapes their allocation

ALT Large language models are increasingly exposed to model-generated text, raising the possibility that future models will be trained recursively on the outputs of their predecessors. Prior work on such self-consuming training loops has largely focused on output quality and downstream performance, leaving their effects on internal representations poorly understood. We study how representations evolve across five generations of self-consumption and whether sparse autoencoders (SAEs) remain reliable under this shift. At each generation, we compare a lineage trained on synthetic data with a matched control trained on real data under the same training settings. We evaluate how well SAEs trained before self-consumption transfer to later generations, measuring reconstruction and sparsity across SAE widths, architectures, and model layers. We then fine-tune the original SAE on each generation and use sparse probing to test whether task-relevant information remains concentrated in a small number

ALT Operator learning approximates the solution map of a parametric partial differential equation from data, enabling fast prediction at new parameter values without repeated simulation. The dominant approach fits a single global low-rank basis, typically a proper orthogonal decomposition (POD), to all solution snapshots. This fails when the snapshots span qualitatively different solution shapes, from smooth diffusive profiles to near-discontinuous shocks. We propose TEMPO (Trajectory Expectation-Maximization Mixture of POD Operators), a two-phase framework for regime-aware operator learning. In the offline phase, TEMPO recovers latent regimes via expectation-maximization, fitting a dedicated linear or nonlinear basis to each discovered regime. In the online phase, a gating network and per-regime branch networks are trained jointly, routing new inputs to the regimes at inference. We test TEMPO on Burgers' flow, Sod shock tubes, Coanda flow, Darcy flow, Navier-Stokes flow and the heterogene

ALT Patch-based transformers tokenise a time series into fixed-size, uniform windows, which commits the model to a single temporal resolution even though real signals are multi-scale. The discrete wavelet transform (DWT) is a natural alternative because it decomposes a signal across scales. A growing body of work makes the wavelet filter learnable, on the assumption that a more flexible wavelet forecasts better. That assumption has rarely been tested in isolation, because a single reported improvement usually credits three things at once. These are the use of any multi-resolution decomposition, the choice of wavelet and training the filter. We introduce a forecaster that turns a learnable-DWT per-scale energy map into variable-length patches by merging adjacent windows that share a dominant scale. Three findings follow. First, most of the improvement is there before any filter is learned. A frozen wavelet beats a fixed-patch baseline on eight of the nine datasets by 1.5% to 7.8%, and on fo

ALT Can an AI make a trained language model more mature, in the sense of a better organized latent space, by looking only at the geometry of that space? An analyst model reads a fixed-size digest of topological descriptors of the layers of a frozen target and proposes a transformation, which closed-form ridge regression writes into the feed-forward weights without gradient training. Using lightweight models, 126 runs show that category purity rose on unseen words and sentences, benchmark accuracy stayed within sampling error, and the analyst matched a category-aware heuristic with gentler edits. In a five-model round robin approach, 16 pairs raised purity and four left the target unchanged. The results provide a baseline for future work on LLM-guided, training-free model editing without using raw weights.

ALT Causal abstraction explains a frozen network by aligning its activations with the variables of a simpler, interpretable model, and checking that interventions on the two agree. Finding such an alignment currently requires a counterfactual dataset: for every pair of prompts, the answer the network ought to give once the variable has been changed. In practice, those labels are produced by an executable algorithm for the behavior, written before studying the model supposed to implement it. However, the criterion these methods optimize only leverages the answer an intervention should produce. Here, we develop an approach that lets the investigator define that relationship directly: the answers the variable can produce, and which of them an intervention should yield. We call this interface a semantic program and introduce Programmable Semantic Interventions (PSI): a low-rank activation subspace learned so that interchanging it changes the model's answer as the program declares, with the exp

ALT These lecture notes are based on an introductory course on neural networks and deep learning that the author taught in the School of Mathematical Sciences at Queen Mary University of London in 2023/24 and 2024/25. The author, however, is a physicist, and it shows: wherever possible, ideas are explained with physical intuition rather than heavy formalism. Readers with a background in physics or mathematics who want a self-contained, introductory route into modern machine learning may therefore find these notes particularly friendly. The notes start from maximum likelihood, least squares, linear regression and (stochastic) gradient descent. They then cover neural networks, backpropagation, the vanishing gradient problem, batch and layer normalization, overfitting and regularization, adversarial examples, and the kernel method. The second half turns to natural language processing: word embeddings, recurrent networks and LSTMs, the attention mechanism and Transformers, and large language m

ALT Masked diffusion models sample the tokens of one round independently. Planned decoders first generate a coarse plan, then fill in the tokens in parallel. We ask when a plan saves rounds, how many it saves, and how fine it must be. We study this on broadcast trees: the leaves are the tokens, the internal nodes are the plan, and the edge correlation ΞΈ sets how strongly tokens depend on the plan. Errors are measured by the forward Kullback-Leibler divergence Ξ΅, and the lower bounds hold for every deterministic value-adaptive rule, also with learned conditionals. We show that a decoder without the plan needs of order min{n,(n/Ξ΅)Ξ±} rounds below the reconstruction threshold, where Ξ±=log b/(4log(1/ΞΈ)) rises to 1/2 at the threshold, and that the exponent stays below 1/2 above it (binary channel: 0.02β€ΞΈβ€0.98). With the plan, logb n+1 rounds suffice. In the binary model below the threshold, a plan with units of m tokens needs of order logb(n/m)+min{m,(n/Ξ΅)Ξ±} rounds, so it gives logarithmic depth

ALT Classical molecular similarity is central to chemical information systems, yet molecule-language models trained on paired structure-text data do not necessarily organize molecular embeddings according to graded structural relationships. Here we introduce Structure-Aware Guided Alignment (SAGA), a training-only regularizer that uses Centered Kernel Alignment to align neural molecular representations with fingerprint-based Tanimoto similarity, without adding parameters or requiring fingerprints at inference. In MolCA and Atomas, SAGA improves molecule-text retrieval, shifts non-exact retrieval toward closer structural analogues, and increases average scaffold-split property-prediction performance, while molecular description and text-to-molecule generation are improved or preserved in single-run ablations. These results show that established chemoinformatics priors can make molecule-language representations more chemically faithful while retaining their multimodal utility.

ALT Transformers have demonstrated the ability to perform in-context learning, that is, to adapt to new tasks from a few examples without updating their parameters. Existing theory has largely focused on emulating gradient-based algorithms for convex statistical problems or establishing general function approximation guarantees without exploiting the underlying statistical structure. We study in-context parameter estimation and trajectory prediction for parametric ordinary differential equations (ODEs), where estimation is generally nonconvex. Under suitable identifiability and regularity conditions, we construct a transformer that estimates unknown system parameters from noisy trajectory observations through a two-stage approximate maximum likelihood procedure: a coarse search provides an initialization within a locally strongly convex region, followed by gradient descent with local linear contraction. For prediction, we compose this estimator with a neural approximation of the ODE soluti

ALT Conformal Seasonal Pools (CSP) showed that a forecaster with no training at all, built only from past values at the same point in the season and from past forecast errors, significantly outperformed DeepNPTS, Amazon's learned non-parametric forecaster. We propose Conformal Expert Pools (CEP), a training-free successor. CEP replaces the single seasonal-naive forecast by a pool of simple forecasters, builds its predictive distribution from their past forecast errors, and adapts both the pool and the errors to the series being forecast. No model is trained and no parameter is learned across series. On the CSP benchmark of 232 datasets and 24,502 forecast windows, CEP is better than CSP on 78% of datasets by CRPS, on 84% by MSIS and on 83% by Winkler score; on a typical dataset its CRPS is 14% lower and its MSIS 23% lower (both p < 1e-22, paired Wilcoxon test), and coverage of 95% intervals rises from 90.1% to 92.2%. CEP has a lower CRPS than DeepNPTS on 88% of datasets, than NPTS on 89% a

ALT In-context learning is usually analyzed as if the examples in the prompt were sampled before the learner is chosen. In deployed systems they are not: examples are retrieved, filtered, re-ranked, edited, and repeatedly tested after seeing the query and the data store. We formalize this gap by treating prompt construction as adaptive data analysis through a conditional stochastic-kernel calculus. Our main quantity is the context capacity, the mutual information or approximate max-information between the data store and the final prompt transcript. For any frozen in-context learner and any adaptive context pipeline with capacity ΞΊ, we prove a finite-sample context generalization bound of order β(ΞΊ/k), where k is the number of selected examples whose losses form the reported score. We prove a matching lower bound, whose constant is exact when log k\llΞΊ\ll k, and derive two PAC-Bayesian retrieval rules: a first-order Gibbs rule that charges data-dependent candidate gates their prior mass, an

ALT Self-driving laboratories must learn to control a process from tens to hundreds of expensive experiments, too few for deep reinforcement learning (RL). We learn the Q-function directly with a Gaussian process (GP), which avoids learning the dynamics, and show that the GP posterior supplies the learning rate of Q-learning. If each Bellman target is treated as a noisy measurement of the Q-function, the learning rate that minimises the mean squared error is the Kalman gain, and applying it to every observed transition is equivalent to GP regression. Heteroscedastic Gaussian Process Q-learning (HGPQ) sets the noise variance of each target with a heteroscedastic GP, which learns how unreliable the targets are across the states and inputs. Targets from noisy operating conditions, or from regions where the value function has large gradients, then move the estimate less than targets from low-noise regions, so HGPQ learns quickly where the data are reliable without being misled where they are n

ALT Test-time scaling systems often choose how much to sample, which prompt to use, and which verifier or aggregation rule to trust after repeatedly inspecting a calibration set. Standard validation error bars are not valid after this adaptive selection. We model a test-time scaling recipe as a randomized predictor with a random cost and prove a time-uniform PAC-Bayes certificate that holds simultaneously over all calibration times, all posterior mixtures over recipes, and all data-dependent stopping rules. Thus the same event covers the recipe search that precedes deployment, not merely the fixed recipe finally reported. For a finite grid of budgets and verifiers, the penalty is the familiar log M/n, which in the worst case no certificate valid after selection can avoid, while non-uniform priors and localized posterior mixtures can give sharper certificates for structured searches. We specialize the result to self-consistency and show a complementary margin law: plurality voting improves
