Class-incremental graph learning suffers from catastrophic forgetting because models are repeatedly fine-tuned on new classes, causing severe recency bias where earlier sessions are misclassified as recent ones. MAEM addresses this by freezing a dedicated Mixture-of-Experts expert per session and routing test nodes via the reconstruction error of session-specific masked autoencoders, eliminating the need for task identities or a shared trainable router. To prevent unbounded model growth, it merges least-used experts via Git Re-Basin alignment followed by topology-aware pseudo-data distillation, achieving strong replay-free performance across four benchmarks.

Fine-tuning pretrained diffusion and flow matching models for reward maximization tends to collapse modes or shift mass between distant regions. This paper introduces Weighted Adjoint Matching, which modifies the stochastic optimal control formulation of Adjoint Matching by placing a time-dependent weight $w(t)$ on the quadratic control cost---expensive early to preserve coarse semantics, cheap late to refine details and increase reward. The authors prove equivalence to maximizing reward under pathwise KL regularization, and demonstrate improved reward-diversity tradeoffs on Gaussian mixtures and MNIST.

Training good latent tokenizers for generative models remains poorly understood, especially how to scale the encoder itself. This paper proposes Relative Prediction Representations (RePR), which supervise patch features to factorize a pointwise mutual information matrix estimated by a conditional diffusion model. The result is a theoretically grounded, negative-sample-free objective that resists feature suppression and claims to be the first to demonstrate compute-optimal tokenizer encoder scaling, matching or beating pixel-space diffusion at equal FLOPs.

KANLab is a browser-based educational tool that adapts the TensorFlow Playground interface to Kolmogorov-Arnold Networks (KANs). It strips away modern performance-oriented additions like residual connections and learnable scaling so that displayed B-spline curves map directly to their trainable control points, letting users drag points and watch gradient descent correct them in real time. By making KAN internals tangible in a compact 2D setting, it aims to bridge the gap between novices and experts.

Standard reinforcement learning on verifiable rewards incentivizes language models to guess rather than admit uncertainty, because a wrong answer carries no penalty while abstention foregoes any chance of reward. This paper frames graceful failure as a three-outcome RL problem (Success \succ Fallback \succ Error), identifies why naive policy-gradient methods collapse to degenerate always-fallback policies, and proposes GracefulRL, a log-barrier-shaped objective that preserves gradient signal near saturation. The idea matters because coding and agentic mistakes can propagate into real harm, yet current RL pipelines train models to maximize success rate without regard for how they behave when success is unlikely.

Mixed-precision quantization for sparse MoE models needs more than expert rankings: NatScale introduces contrast calibration, a principled way to control how strongly sensitivity scores compete for storage. It applies a model-specific power transformation $w_i = g_i^{\beta_\tau}$ so that normalized squared weights achieve a target Jeffreys divergence $\mathcal{J}(q(\beta_\tau), u) = \tau$, preserving rank order while setting a common information-theoretic scale. Combined with joint gate/up/down error measurement and spectral structural priors, it allocates bit-widths under a global byte budget via LP relaxation and rounding.

What is the least width required for a deep neural network to universally approximate continuous functions from $\mathbb{R}^n$ to $\mathbb{R}^m$? This paper attacks the minimal-width problem by importing tools from differential topology, proving that whenever real projective space $\mathbb{RP}^{n-1}$ smoothly immerses into $\mathbb{R}^n$, any network with a continuous monotone activation must have width at least $n+\min\{n,m\}$ to approximate arbitrary maps. This yields the exact ReLU minimal width $n+m$ for the special dimensions $n\in\{2,3,4,7,8\}$ with $m\le n$, and exposes a surprising link between universal approximation and the immersibility of projective spaces.

Rotation-based post-training quantization for LLMs currently forces a choice between cheap data-agnostic transforms like Hadamard and costly learned rotations that require iterative optimization. SpectrumQuant introduces a closed-form, covariance-aware orthogonal transform derived from a first-principles objective on the Stiefel manifold: by aligning activations with their eigenbasis and applying a Hadamard or butterfly mixing component, it equalizes per-coordinate variances to minimize quantization error. The result is a method that cuts calibration costs by more than $20\times$ relative to learned baselines while preserving competitive accuracy on Llama and Mistral models.

ChemArgus tackles a real blind spot in chemistry evaluation: endpoint-only scores hide where reasoning chains break. The benchmark enforces fine-grained rubric scoring across three controlled protocols that share a single scoring unit, so gaps between full-problem, sequential-carry, and oracle-scaffolded results isolate integration, propagation, and local-reasoning failures. The design is principled, yet the paper never states the dataset size and evaluates models lacking citations or independently verifiable endpoints, leaving the empirical results impossible to validate.

JointServe targets multi-model LLM agent serving on memory-constrained GPUs, where heterogeneous expert weights and ever-growing KV caches physically contend for HBM bytes. It unifies weight shards and KV blocks as competing residency objects under one budget, exploits the agent's own tool-call stream to prefetch destination experts during caller generation gaps, and streams only incremental layer groups when a swap occurs. The work is motivated by the architectural shift from single-model tool use to planner-orchestrated pools of heterogeneous experts—a pattern visible in SWE-bench, $\tau$-bench, and OSWorld traces—and targets the private and edge deployments where full co-residency is arithmetic impossible.

Neural emulation need not require learning a substitute for known physics. This paper investigates PolyNets that compile continuous-time polynomial dynamics and discrete-time RK4 integration into a single tensor graph, benchmarking the approach on the 1D Korteweg–de Vries equation against classical CPU/GPU solvers and three trained neural models. The central finding is that when governing coefficients are known, the resulting network requires no training yet preserves numerical fidelity while achieving competitive GPU throughput.

Architectural drift is expensive to detect when the mapping between source files and architectural modules is missing or outdated. This paper frames code-to-architecture mapping as semi-supervised node classification over a directed multi-relational dependency graph, expanding a small labeled seed via iterative self-training. By holding features, training loops, and classifier heads fixed across GCN, GAT, and a graph-free MLP, the authors isolate whether dependency structure genuinely improves mapping accuracy.

MMLA-RSI proposes recursive self-improvement through two Memory-Mediated Learning Architecture routes: Reasoning-Time Training (RTT) treats the numerical policy-state updater as an object of improvement, while Reasoning-Time Memory Adaptation (RTMA) treats the memory-management rule as an object of improvement. Each admitted successor takes over its parent's role in producing the next generation's rule. The paper's main contribution is a formally bounded construction with shared interfaces and a common algorithm, though the authors explicitly state that beneficial recursive chains remain to be evaluated.

MorphingNet attacks the scalability problem in symbolic regression by turning formula search into continuous network training. Its units are pole-free rational functions and complex exponential-logarithm shapes whose parameters morph smoothly between elementary forms, so the network itself is a closed-form formula at every step. A post-training structure-recovery step rewrites the learned expression into compact elementary notation. The paper reports state-of-the-art SRBench accuracy, exact recovery of most ground-truth formulas, and GPU speedups of two orders of magnitude over genetic programming on large data.

CTIA-ODE proposes a continuous-time architecture for long-term multivariate forecasting that sidesteps rigid alignment between endogenous and exogenous variables. Its core mechanism, Continuous-Time Inter-Variable Attention (CTIA), lets a Neural ODE’s evolving hidden state query a static exogenous key-value field during integration. The framework is stabilized by a Linear Residual Augmentation trend anchor and an Outlier-Robust Hybrid Loss. The authors report state-of-the-art results across eight benchmarks with a compact memory footprint.

Discriminative models generalize to unseen samples, suggesting they capture data-manifold structure beyond their task labels. This paper asks whether that structure can be exposed as a generative signal without retraining, labels, or auxiliary generators. The authors propose the Discriminative Score Function (DSF), which turns the loss-gradient geometry of a fixed discriminative model into an image-space update rule. DSF enables unconditional generation, conditional guidance, inpainting, and zero-shot quantization calibration across classifiers, detectors, and self-supervised encoders.

Standard diffusion models rely on additive Gaussian noise, yet many imaging modalities---including PET and astronomical photon counting---produce discrete Poisson observations. This paper introduces Mixed Poisson Diffusion Models (MPDMs), a generative framework in which the forward corruption is driven by multivariate counting processes rather than Brownian motion. The core payoff is a family of posterior samplers structurally matched to Poisson inverse problems, featuring exact sampling for Poisson denoising and principled adaptations of DPS, PnP-DM, and DAPS.

NeuroDETR tackles asynchronous EEG event detection, where the goal is to identify both the class and temporal boundaries of neural events directly from continuous recordings without predefined windows. The method introduces a DETR-style architecture that uses multiscale discrete spatiotemporal vector quantization to build compact event memories, combined with an event-aware sparse attention module called FOCAL that concentrates refinement on likely event periods. Evaluated across eight diverse EEG datasets spanning P300, motor imagery, speech, and seizure detection, the authors report state-of-the-art event-level F1 scores compared to five baselines.

This paper confronts the evaluation crisis in agentic self-distillation, where a flood of recent methods claim gains over GRPO using inconsistent protocols and reused baselines. The authors conduct a unified benchmark of ten representative approaches across four tasks, three model sizes, and 132 training runs totaling over 12,000 H800 GPU-hours, measuring not only final accuracy but also training stability and wall-clock cost. Their central finding is sobering: no method consistently beats GRPO, many suffer severe hidden instability, and extra compute rarely translates into better agents.

Single-channel RF receivers cannot use spatial filtering when a desired signal overlaps with a structured co-channel interferer in both time and frequency, a regime where linear filters fail because the impairment is non-Gaussian. FlowSep-RF addresses this by learning a clean-endpoint rectified flow that maps a raw IQ mixture directly to the target waveform, enabling separation with a single neural function evaluation instead of iterative diffusion sampling. The core idea is to treat the received mixture as the source and the clean signal as the target of a near-straight flow, combining time-domain and frequency-domain features to improve reconstruction fidelity and downstream demodulation accuracy. If the gains hold under broader channel conditions, this could provide a low-latency replacement for score-based separators in practical baseband receivers.