arXiv:2405.06043v2 Announce Type: replace Abstract: Algorithms which learn environments represented by automata in the past have had complexity scaling with the number of states in the automaton, which can be exponentially large even for automata recognizing regular expressions with a small description length. We thus formalize a compositional language that can construct automata as transformations between certain types of category, representable as string diagrams, which better reflects the description complexity of various automata. We define complexity constraints on this framework by having them operate on categories enriched over filtered sets, and using these constraints, we prove elementary results on the runtime and expressivity of a subset of these transformations which operate deterministically on finite state spaces. These string diagrams, or "string machines," are themselves morphisms in a category, so it is possible for string machines to create other string machines in runtime to model computations which take more than constant memory. We prove sufficient conditions for polynomial runtime guarantees on these, which can help develop complexity constraints on string machines which also encapsulate runtime complexity.
Science Journals
arXiv:2606.24124v1 Announce Type: new Abstract: Multi-step reasoning with Chain-of-Thought (CoT) prompting remains fragile: logical errors or hallucinations in early steps silently propagate, producing confident but incorrect conclusions. This paper presents VeryTrace, a zero-shot verification-and-repair framework that formalizes natural-language reasoning traces into a structured, compilable representation. VeryTrace introduces a Domain-Specific Language (DSL) that (i) makes step dependencies explicit, (ii) mechanizes quantitative content as executable expressions, and (iii) structures semantic inferences via deduction schemas. Our hybrid verifier combines deterministic checks for computational correctness, dependency resolution, and constraint satisfaction with targeted LLM audits for non-mechanizable semantic judgments, enabling step-level error localization and repair. Across three diverse domains-competition mathematics (AIME 2025), robotics planning (LLM-BabyBench), and kinship reasoning (CLUTRR), VeryTrace improves accuracy over zero-shot baselines on state-of-the-art LLMs without requiring domain-specific training or in-context examples, demonstrating that formalized trace verification achieves both precision and generalization.
arXiv:2606.24128v1 Announce Type: new Abstract: Diffusive-convective (DC) staircases in the ocean are observed across a wide range of settings, but their formation, structure, and persistence are not fully understood. Theories for DC staircases are reviewed to identify mechanisms governing their development and evolution. Staircase evolution through layer merging and possibly interface splitting, including the relationship to background turbulence, is assessed. Oceanographic examples illustrate the variety of settings in which DC staircases are found, and how they can persist under weak turbulence but are disrupted when turbulence becomes sufficiently strong. Key open questions are identified, highlighting the challenge of linking small-scale processes to the large-scale coherence and persistence of DC staircases in the ocean.
arXiv:2606.24174v1 Announce Type: new Abstract: We demonstrate that the Talbot length of periodic wave packets, long believed to be solely determined by their periodicity, is strongly affected by the source coherence in the nonlinear propagation regime. We reveal that reducing the source coherence -- and consequently the speckle size of a periodic field -- shortens the Talbot length and significantly improves the quality of the recurrent Talbot images, despite a fixed source periodicity. This effect arises from coherence-mediated nonlinear mode coupling, which alters a phase synchronization condition undergirding the wave packet revivals. Our findings expose a hidden role of coherence in governing Talbot revivals of random waves in the nonlinear regime, crucially informing the understanding and application of the Talbot effect in realistic media.
arXiv:2412.08147v2 Announce Type: replace Abstract: Pareto fronts are useful to find good task-mixing strategies for multitask finetuning, but they are also costly to compute. To reduce costs, recent works have used existing model merging methods to help train cheap surrogate models to estimate the Pareto fronts. However, no work has yet considered designing new model-merging methods to directly, and provably, improve the quality of Pareto fronts. Here, we fill this gap by proposing a new Bayesian approach called Variational Model Merging. In this approach, existing model-merging methods are obtained as special cases of "posterior-merging" when Gaussian posteriors are used and new model-merging strategies can be derived by using non-Gaussian posteriors. Our main theoretical result is to show that more flexible posteriors necessarily yield better estimates of Pareto fronts. For instance, a Pareto front estimate obtained by merging full-Gaussian posteriors is expected to be better than that obtained by using isotropic Gaussian posteriors. We validate the theory through extensive empirical results on vision and language transformers where better Gaussian families consistently yields better or comparable Pareto fronts. Our work is a rare instance where Bayesian ideas are used to improve Pareto analysis.
arXiv:2503.00069v2 Announce Type: replace Abstract: Recent progress in large language models (LLMs) has focused on producing responses that meet human expectations and align with shared values - a process coined alignment. However, aligning LLMs remains challenging due to the inherent disconnect between the complexity of human values and the narrow nature of the technological approaches designed to address them. Current alignment methods often lead to misspecified objectives, reflecting the broader issue of incomplete contracts, the impracticality of specifying a contract between a model developer, and the model that accounts for every scenario in LLM alignment. In this paper, we argue that improving LLM alignment requires incorporating insights from societal alignment frameworks, including social, economic, and contractual alignment, and discuss potential solutions drawn from these domains. Given the role of uncertainty within societal alignment frameworks, we then investigate how it manifests in LLM alignment. We end our discussion by offering an alternative view on LLM alignment, framing the underspecified nature of its objectives as an opportunity rather than perfect their specification. Beyond technical improvements in LLM alignment, we discuss the need for participatory alignment interface designs.
arXiv:2602.12501v2 Announce Type: replace Abstract: The long-time evolution of decaying homogeneous turbulence is a fundamental building block of the subject. We investigate the problem by using a comprehensive suite of Direct Numerical Simulations. The simulations cover initial Taylor microscale Reynolds numbers $Re_\lambda$ from $30 \text{ to } 145$, with multiple independent realizations obtained at each $Re_{\lambda}$ to ensure statistical robustness. The energy spectrum is initialized with the Birkhoff-Saffman (BS) form (with $E(k)\sim k^2$ for small $k$) in one case, and the Loitsianskii-Kolmogorov-Batchelor (LKB) form (with $E(k)\sim k^4$ for small $k$), in another. Simulations are performed for unprecedented durations, of the order of 200,000 initial eddy-turnover times in some instances. For both BS and LKB, the turbulent kinetic energy $En$ shows, after an initial transient, unambiguous power-law decay, $En\sim t^{-n}$, with nearly constant decay exponents $n$, whose values are consistent with past theoretical results (and thus not universal). We compute various length scales, second-order structure functions, and the spectral form at large wavenumbers; we note that an initially set $-5/3$ slope disappears quickly, while a perceptible $-1$ power region appears. In particular, we compare the present findings with predictions from the recent theory for decaying turbulence developed by Migdal 2026 Philos. Trans. R. Soc. A 384, 20250032. (doi:10.1098/rsta.2025.0032). The agreement for the BS case is excellent except for the large-wavenumber spectrum. A general discussion and assessment of results is provided in terms of the putative universality of energy decay. A main conclusion is that the energy decay is significantly influenced by ``boundary effects", and that universality likely manifests only when those effects are removed. Alternatively, it may be more useful to discuss the universality of enstrophy decay.
arXiv:2606.24517v1 Announce Type: new Abstract: In this paper, we analyse the existence of solutions and finite element approximation of a steady-state two-dimensional induction heating problem. One of the main difficulties of the problem is its right-hand side which, at a first sight, is only integrable. Using a priori regularity results for the PDEs involved it is shown that the natural weak formulation of the problem can be justified. Then, we study the finite element approximation and prove that the standard Galerkin FEM converges in convex domains and under suitable conditions on the mesh. We improve on this result by applying the recently-proposed bound-preserving method (BPM) to the heat equation, and show that this method converges to a solution of the problem under less stringent conditions on the domain and the mesh. As these analyses are carried out without any assumption on regularity of the solutions, then the convergence of the finite element method also proves existence of solutions. Several numerical experiments confirm the theoretical results, and showcase the improvement provided by the use of the bound-preserving method over the standard finite element method.
arXiv:2606.23995v1 Announce Type: new Abstract: Recent work has established that regularized policy gradient methods such as PPO, when used in self-play, can match or exceed specialized game-theoretic algorithms for solving two-player zero-sum imperfect-information games. The uniform distribution has emerged as a strong policy regularization target for this purpose, but it regularizes equally toward all actions regardless of their viability. We introduce EMAgnet, which instead regularizes toward an exponential moving average (EMA) of the last-iterate policy's parameters, providing an adaptive regularization target that evolves with the agent's improving strategy. We evaluate EMAgnet on both standard two-player zero-sum benchmarks and modified benchmarks with exploration challenges and large numbers of strictly dominated strategies. Relative to PPO self-play with uniform-magnet regularization under both linear and power-law annealing schedules, EMAgnet achieves lower exploitability in the majority of tested environments, with consistent performance gains across games containing strictly dominated strategies.
arXiv:2606.23742v1 Announce Type: new Abstract: Physical neural networks promise low-power machine learning by computing directly with analogue device physics, but most architectures force nonlinear device responses to act as scalar weights. Inspired by Kolmogorov-Arnold networks, we place trainable nonlinear functions on the connections, making each physical connection a learnable computational element. Realising these functions as analogue band-pass filters on field-programmable analogue arrays, we find that the benefit is task-dependent and follows from the smoothness of the physical basis: the networks represent smooth, continuously valued targets, including robotic kinematics, continuous control, and photovoltaic maximum-power-point tracking, with far fewer nodes and connections than multilayer perceptrons, but offer no parameter-efficiency advantage on classification-like decision boundaries. Trained networks transfer to hardware across approximately 35,000 connections with quantified fidelity, and a dedicated CMOS implementation is projected to operate at approximately 30 microwatts. A memristive realisation reproduces the same behaviour in simulation, indicating that the advantage comes from placing trainable nonlinearity on connections, rather than from a particular device.
arXiv:2606.23797v1 Announce Type: new Abstract: Graph and multi-agent orchestration frameworks make production large language model (LLM) workflows practical, but they do not by themselves solve conversational continuity when users maintain several interdependent objectives. This conceptual systems paper focuses on the high-complexity end of that design space, where goals can be suspended, resumed, revised, and invalidated by actions in other goals. We introduce the Goal-Oriented Dialogue Runtime (GODR), a framework-neutral design pattern that treats goals, task frames, lifecycle state, invalidation rules, and resumption contracts as first-class runtime objects while delegating bounded execution to graph runtimes, agents, tools, or application programming interfaces (APIs). GODR is not proposed as a replacement for workflow graphs in simple guided processes; it is intended for complex, multi-domain, interruptible conversations where objective continuity cannot be recovered reliably from agent identity, chat history, or execution-graph position alone. The paper formalizes the problem, proposes runtime objects and architecture-selection criteria, and frames evaluation as an agenda for future empirical validation rather than as a measured performance claim.
arXiv:2606.24141v1 Announce Type: new Abstract: We study community detection in the two-block stochastic block model under the setting where multiple independent graph samples drawn from the same distribution are available. Building on a recently simplified spectral algorithm that preserves the independence of adjacency matrix entries throughout, we show that averaging $m$ independent samples before applying spectral partitioning reduces the error bound $\gamma$ exponentially in $m$: specifically, one can find a $\gamma$-correct partition with probability $1 - o(1)$ whenever $\frac{(a-b)^2}{a+b} \geq \frac{C}{m} \log \frac{2}{\gamma}$, improving the single-sample requirement by a factor of $m$. The key technical contribution is a multi-sample analogue of the spectral norm bound on the noise matrix, which propagates through the Davis-Kahan subspace angle analysis to yield the improved recovery guarantee. We provide experimental validation across a range of graph sizes ($n$ up to $1000$) and sample counts ($m$ up to $9$), demonstrating that the derived bounds are sharp and that even two or three samples yield dramatic improvements in recovery accuracy. Our results offer a rigorous theoretical foundation for graph data augmentation strategies used in modern graph representation learning.
arXiv:2606.24143v1 Announce Type: new Abstract: On-policy distillation (OPD) trains a student on its own rollouts guided by teacher feedback and is becoming increasingly important for large language model (LLM) post-training. Like reinforcement learning (RL), however, OPD faces an on-policy systems bottleneck, as rollouts can dominate training time for reasoning workloads. Asynchronous training pipelines can alleviate this bottleneck by decoupling rollout generation from learner updates, but doing so introduces stale-policy data. While prior work has studied stale data in asynchronous RL, its effects in OPD remain underexplored. We present the first systematic study of staleness in asynchronous OPD, focusing on a practical setting where teacher feedback is implemented through local KL losses and full-vocabulary teacher logits are too expensive to store or transfer, necessitating finite teacher-score caches. We first show that KL direction changes the stale-data problem: teacher-weighted forward KL is more robust to stale rollouts, whereas student-weighted reverse KL is vulnerable. Second, for this vulnerable reverse-KL case, we study whether methods designed to stabilize asynchronous RL can mitigate OPD staleness. In our experiments, they do not improve over a simpler OPD-specific surrogate: recomputing the reverse-KL signal under the current student at learner time. Third, we analyze how finite teacher-score caches create a bias-variance tradeoff for sparse and sampled reverse-KL OPD estimators. This motivates multi-sample Monte Carlo (MC), which preserves MC correctability while reducing one-sample variance. Finally, we present and open-source AsyncOPD, a fully asynchronous OPD training pipeline built from these estimator choices. Experiments show that AsyncOPD improves training throughput by $1.6\times$ to $3.8\times$ over strict synchronous training while reaching comparable accuracy.
arXiv:2511.07788v2 Announce Type: replace Abstract: Developing chip-scale photonic platforms for the visible spectrum is essential for translating next-generation optical technologies. Here, we demonstrate that tantalum pentoxide Ta2O5 on insulator is a highly promising material framework for these applications, supporting high-Q resonator architectures across the visible spectral range. Utilizing routine and scalable thin-film deposition and fabrication methodologies, we realize Ta2O5 waveguide and resonator layouts compatible with photonic integrated circuit (PIC) architectures. Through a comparison with numerical simulations, we characterise compact racetrack resonators, evaluating their performance in both the green and blue spectral regions. In addition, we fabricate large-scale ring resonators exhibiting record-high quality factors for Ta2O5 in the visible spectrum, exceeding 7.0 x 10^4. These results highlight the immense potential of this platform for visible photonic systems and establish a robust, straightforward baseline for future visible integrated optics.
arXiv:2511.17388v3 Announce Type: replace Abstract: Position information is essential for language modeling. In softmax transformers, Rotary Position Embeddings (\textit{RoPE}) encode positions through \textit{fixed-angle} rotations, while in linear transformers, order is handled via input-dependent (selective) gating that decays past key-value associations. Selectivity has generally been shown to improve language-related tasks. Inspired by this, we introduce \textit{Selective RoPE}, an \textit{input-dependent} rotary embedding mechanism, that generalizes \textit{RoPE}, and enables rotation in \textit{arbitrary angles} for both linear and softmax transformers. We show that softmax attention already performs a hidden form of these rotations on query-key pairs, uncovering an implicit positional structure. We further show that in state-space models and gated linear transformers, the real part manages forgetting while the imaginary part encodes positions through rotations. We validate our method by equipping gated transformers with \textit{Selective RoPE}, demonstrating that its input-dependent rotations improve performance in language modeling and on difficult sequence tasks like copying, state tracking, and retrieval.
arXiv:2512.11121v2 Announce Type: replace Abstract: Pre-trained image restoration models often fail on out-of-distribution (OOD) real-world degradations. Adapting to these domains is challenging as real-world data lacks paired ground truth, and unsupervised methods often require unstable architectural changes. We propose Generative Manifold Distillation (GMD), which reframes domain adaptation as geometric manifold alignment. GMD operates in a strictly unpaired setting, requiring only low-quality (LQ) target observations. By leveraging the flow-matching dynamics of a frozen text-to-image foundation model, GMD projects off-manifold restorations onto the natural image manifold to generate high-quality pseudo-targets. To ensure stability, a quality-gated manifold filter rejects off-manifold samples, while source-anchored trajectory regularization prevents error accumulation. Ultimately, GMD distills a powerful generative prior into an efficient restoration network. Experiments demonstrate that GMD seamlessly adapts to new distributions using only LQ inputs, drastically improving perceptual quality with zero architectural modifications or added inference latency.
arXiv:2512.13331v2 Announce Type: replace Abstract: This work addresses the Assembly Line Rebalancing Problem driven by cycle-time changes in manual assembly systems where multiple workers operate in parallel within the same station. A multi-objective optimization model is proposed that incorporates task reassignment, worker allocation, ergonomic evaluation, and explicit spatial feasibility through work-area constraints. The formulation minimizes deviations from the current configuration while promoting balanced workload and ergonomic conditions among workers. The main contribution is the extension of assembly line rebalancing to multi-worker settings with explicit spatial constraints. Computational experiments on synthetic instances demonstrate that the model consistently generates feasible reconfigurations, highlighting its potential as a decision-support tool for industrial rebalancing in flexible production environments.
arXiv:2512.15067v4 Announce Type: replace Abstract: The rapid growth in wireless infrastructure has increased the need to accurately estimate and forecast electromagnetic field (EMF) levels to ensure ongoing compliance, assess potential health impacts, and support efficient network planning. While existing studies rely on univariate forecasting of wideband aggregate EMF data, multivariate narrow-band EMF forecasting is needed to capture the inter-operator and inter-frequency variations essential for proactive network planning. To this end, this paper introduces EMFusion, a conditional diffusion-based EMF forecasting framework that integrates diverse contextual factors, such as time of day, season, and holidays, while providing uncertainty-aware probabilistic forecasts. The proposed architecture features a residual U-Net backbone enhanced by a cross-attention mechanism that dynamically integrates external conditions to guide the generation process. Furthermore, EMFusion integrates an imputation-based sampling strategy that treats forecasting as a structural inpainting task, ensuring temporal coherence even with irregular measurements. Unlike standard point forecasters, EMFusion generates empirical probabilistic prediction intervals from the learned conditional distribution, providing uncertainty-aware probabilistic forecasting rather than simple point estimation. Numerical experiments conducted on the multivariate narrow-band EMF datasets demonstrate that EMFusion with the contextual information of working hours outperforms the baseline models with or without conditions. The proposed EMFusion outperforms the best baseline by 23.85% in continuous ranked probability score (CRPS) and 13.93% in normalized root mean square error.
arXiv:2606.24191v1 Announce Type: new Abstract: Task specialization can improve the efficiency of multi-robot systems (MRSs). Previous works have investigated the emergence of task-specialist robot controllers through evolutionary optimization and have argued that task specialization is more likely to evolve when subtask behaviors are readily available as building blocks. However, the available evaluation budget must be distributed across all subtasks, whereas a single generalist behavior can exploit the entire budget for its own optimization. We present a cost-benefit analysis of evolving task-specialist versus generalist behaviors in a foraging scenario here. In a physics-based robotics simulator, we study the total evaluation budget required to evolve task-specialist behaviors that outperform generalist behaviors across MRS sizes. We show that with increasing MRS size, a lower total evaluation budget is sufficient to evolve specialists that outperform generalists.
arXiv:2503.21450v4 Announce Type: replace Abstract: AI-assisted protein design has emerged as a critical tool for advancing biotechnology, as deep generative models have demonstrated their reliability in this domain. However, most existing models primarily utilize protein sequence or structural data for training, neglecting the physicochemical properties of proteins.Moreover, they are deficient to control the generation of proteins in intuitive conditions. To address these limitations,we propose CMADiff here, a novel framework that enables controllable protein generation by aligning the physicochemical properties of protein sequences with text-based descriptions through a latent diffusion process. Specifically, CMADiff employs a Conditional Variational Autoencoder (CVAE) to integrate physicochemical features as conditional input, forming a robust latent space that captures biological traits. In this latent space, we apply a conditional diffusion process, which is guided by BioAligner, a contrastive learning-based module that aligns text descriptions with protein features, enabling text-driven control over protein sequence generation. Validated by a series of evaluations including AlphaFold3, the experimental results indicate that CMADiff outperforms protein sequence generation benchmarks and holds strong potential for future applications. The implementation and code are available at https://github.com/HPC-NEAU/PhysChemDiff.
arXiv:2606.24157v1 Announce Type: new Abstract: The space $\mathcal{P}_2(\mathbb{R}^d$) of probability measures with finite second moment carries a natural geometry: the quadratic Wasserstein distance W_2 makes it a complete metric space and, following Otto, a (formal) Riemannian manifold whose geodesics are the optimal-transport interpolations. On this manifold, the gradient flow of the free energy F(rho) = KL(rho || \pi) is exactly the Fokker-Planck equation, and its implicit-Euler discretization is the JKO scheme. This is the geometry underlying diffusion models: the forward process descends the free energy, and each denoising step realizes one JKO step, which recovers DDPM, DDIM, NCSN/SMLD, and Energy Matching; this is one scheme, not separate theories. The same manifold supports a second variational principle. Its geodesics - the minimum-action curves of the Benamou-Brenier formula - are precisely the optimal-transport paths that Flow Matching learns. Fixing both endpoints and following the geodesic, generation becomes a deterministic ODE along a straight line, hence far fewer sampling steps. Placing both families of models on one manifold makes their relationship exact: diffusion follows a free-energy gradient flow, an initial-value problem; optimal-transport Flow Matching follows a Wasserstein geodesic, a boundary-value problem. The two reach the same endpoints along different paths.
arXiv:2606.24161v1 Announce Type: new Abstract: Foundation models trained on biased datasets often rely on spurious correlations between target labels and non-causal attributes, resulting in poor generalization on minority groups. Bias mitigation remains challenging due to two fundamental issues. First, when group labels are unavailable, existing group-unsupervised methods typically infer spurious attributes implicitly from model behavior, making it difficult to identify spurious factors that are semantically aligned with real-world biases. Second, even with pseudo spurious supervision, most existing debiasing methods follow a single-branch design that operates within a single shared feature space, where target and spurious attributes are intrinsically entangled. To address the first challenge, we introduce Confidence-guided Bias Concept Mining (CBCM), which leverages diffusion-disentangled, semantically grounded concept representations to identify reliable spurious attributes without attribute annotations. To address the second challenge, we propose Dual-branch Cross-projection Debiasing (DCD), a prompt-tuning framework that separates target and spurious representations into two branches and explicitly removes spurious information through cross null-space projection while preserving target-relevant semantics. Extensive experiments on four benchmark datasets show that our method achieves state-of-the-art worst group accuracy among group-unsupervised approaches, while tuning at most 0.22% of the model parameters. The source code is available in the supplementary materials.
arXiv:2512.21721v2 Announce Type: replace Abstract: We study discrete-time consensus dynamics on multilayer networks in which each layer evolves via a time-varying doubly stochastic interaction matrix, and inter-layer coupling is introduced through two mechanisms: (i) distribute-then-average and (ii) average-then-distribute. These define conservative redistribution processes that preserve total mass across all layers and can be viewed as stochastic averaging driven by products of time-inhomogeneous stochastic matrices with structured coupling. For both mechanisms, we construct quadratic Lyapunov functionals that form nonnegative supermartingales, yielding almost sure convergence. The analysis combines martingale arguments with dissipation identities and connectivity properties of induced interaction graphs. Under recurrent connectivity conditions on subgraphs of the time-varying interaction structure, we prove asymptotic consensus to the global average determined by the initial total mass. This provides a unified framework for multilayer averaging dynamics, extending classical consensus results for products of stochastic matrices to settings with explicit inter-layer coupling. As corollaries, we specialize the general framework to the multilayer garbage disposal dynamics, thereby establishing convergence guarantees under natural connectivity conditions on the underlying graphs.
arXiv:2601.02272v2 Announce Type: replace Abstract: In a staged, beam-driven, plasma wakefield accelerator, electrons are accelerated in a sequence of plasma stages, each powered by a driver electron bunch. Between each stage, a magnetic chicane is used to dispose of the spent driver and to inject the new fresh one, while transporting the witness bunch. We discuss the effect of synchrotron radiation in the interstage section on the accelerating gradient and on the final energy reach of such a machine.
arXiv:2601.06395v3 Announce Type: replace Abstract: Large language models (LLMs) are increasingly multilingual, yet open models continue to underperform relative to proprietary systems, with the gap most pronounced for African languages. Continued pre-training (CPT) offers a practical route to language adaptation, but improvements on demanding capabilities such as mathematical reasoning often remain limited. This limitation is driven in part by the uneven domain coverage and missing task-relevant knowledge that characterize many low-resource language corpora. We present \texttt{AfriqueLLM}, a suite of open LLMs adapted to 20 African languages through CPT on 26B tokens. We perform a comprehensive empirical study across five base models spanning sizes and architectures, including Llama 3.1, Gemma 3, and Qwen 3, and systematically analyze how CPT data composition shapes downstream performance. In particular, we vary mixtures that include math, code, and synthetic translated data, and evaluate the resulting models on a range of multilingual benchmarks. Our results identify data composition as the primary driver of CPT gains. Adding math, code, and synthetic translated data yields consistent improvements, including on reasoning-oriented evaluations. Within a fixed architecture, larger models typically improve performance, but architectural choices dominate scale when comparing across model families. Moreover, strong multilingual performance in the base model does not reliably predict post-CPT outcomes; robust architectures coupled with task-aligned data provide a more dependable recipe. Finally, our best models improve long-context performance, including document-level translation. Models and code have been released on [Huggingface](https://huggingface.co/collections/McGill-NLP/afriquellm) and [Github](https://github.com/McGill-NLP/AfriqueLLM).