Multi-token prediction (MTP) accelerates language model decoding by drafting several future tokens in one forward pass, but the speculation horizon must be fixed before drafting and the best length varies with context. This paper frames horizon selection as an online survival problem: it predicts per-position rejection hazards to forecast how long a draft prefix will survive exact verification, then adaptively picks the horizon maximizing expected tokens per second. The resulting controller, AMTP, requires no retraining, leaves the target verifier unchanged, and comes with a sublinear throughput-regret bound against the best horizon at each visited context.

Evaluating robotic foundation models still relies on expensive closed-loop rollouts in simulation or reality. This paper proposes Sensorimotor Alignment ($S^2$), a training-free offline metric inspired by the Platonic Representation Hypothesis, which scores how well a model's sensory embeddings align with the geometry of recorded end-effector trajectories. The authors find positive correlations between $S^2$ and task success across vision-language-action models, raw vision-language models, world action models, and real-robot policies, and argue the metric can cheaply rank models and diagnose representation-level behavior.

This paper studies P-trainable neural networks, where a proportion P of hidden neurons are trainable and the rest stay fixed as random features, bridging random feature methods and fully trainable networks. The authors establish two-layer approximation bounds showing that a small trainable fraction can capture most of the gain when initialization mismatch is large, yielding a phase transition in rates of the form $\mathcal{O}(((P/\kappa_1^*+(1-P)/\kappa_0^*)m/\log m)^{-\alpha})$. They verify these assumptions explicitly for spherical biased-ReLU features and present numerical experiments showing analogous phase transitions in training losses across architectures with up to 15 layers.

AI systems increasingly make consequential decisions, but fully auditing every output is infeasible at scale. This paper studies selective auditing: sequentially deciding when to trust an AI decision versus escalate it to an expert, using only a scalar confidence score. The authors frame this as an online constrained optimization problem and derive algorithms that guarantee the fraction of unaudited errors stays below a user-specified threshold α without any distributional assumptions. The core contribution is an optimally efficient algorithm that provably minimizes audit effort for both adversarial and i.i.d. data streams.

Dataset-use auditing for text-to-image models is difficult because fine-tuned models absorb collection-level visual style without reproducing individual training images. SpectralCanary addresses this by blending a fixed canary image into the low-frequency Fourier content of every collection image before release, preserving captions and avoiding auxiliary generation or per-image optimization. A lightweight verifier then detects the transferred spectral signature in a suspect model's black-box outputs. The method achieves strong detection rates across multiple architectures and domains with minimal fidelity loss, establishing low-frequency traces as a practical auditing mechanism.

Standard softmax and cross-entropy hide two floating-point bugs: Softmax Collapse rounds the target gradient to zero once logit margins cross a precision cliff, while Softmax Cancellation produces phantom gradients when a student network approaches its teacher. Barron et al. rewrite both functions in log-space using softplus, expm1, and log1p to recover accurate, zero-sum gradients across bf16, fp32, and fp64. They tie these numerical issues to the Slingshot phenomenon in grokking and derive a clean equilibrium for logit margins under AdamW with weight decay.

This paper asks whether transformers that generate intermediate reasoning steps can discover new algorithms beyond the teacher they are trained to imitate. It designs a two-stage training protocol for in-context linear regression—first supervised by gradient descent, then by ground-truth weights—and proves that a one-layer linear self-attention transformer learns the Heavy-Ball method. The result matters because it gives a rigorous case where chain-of-thought reasoning internalizes acceleration rather than simply memorizing the supervised update.

Looped language models reuse the same parameter block many times to improve quality without inflating weight memory, but each recurrence still stores its own KV-cache, so inference memory grows linearly with depth. This paper studies when and why the final recurrence's KV-cache can be shared across earlier recurrences—a property the authors call KV-shareability—finding that contractive residual dynamics are necessary but not sufficient. They introduce a two-stage recipe (stochastic-depth pretraining plus a lightweight post-training corrector called BeNice!) that reliably enables near-lossless KV-sharing and composes with existing cache-compression techniques.

Local anomaly detection in images with multimodal inlier distributions—such as medical scans, faces, and X-ray baggage—remains challenging for industrial-defect methods that rely on heuristic reconstruction errors or feature distances. LACToS introduces a principled alternative by scoring superpixels via statistical typicality under the conditional inpainting posterior of an inlier diffusion model, made tractable by grouping pixels into superpixels and using a single denoising posterior as a cheap proxy. This shift from point estimates to distribution-aware testing matters because it directly addresses the false positives that arise when normal images contain diverse, context-dependent content.

CRANE addresses the lack of post-hoc counterfactual explanation methods for multimodal classifiers, where applying unimodal techniques independently produces incoherent, chimeric explanations that violate cross-modal constraints. The framework retrieves real training instances from the target class via several nearest-neighbour strategies, most notably Multimodal Consensus Retrieval (MC-R), which intersects per-modality neighbour sets to preserve joint distributional coherence. By operating without model internals for most strategies and grounding explanations in observed data, CRANE aims to deliver plausible, actionable counterfactuals across heterogeneous inputs including text, images, tables, and time series.

On-policy self-distillation trains language models by distilling privileged self-supervision over the model's own rollouts, yet its privacy risks have remained almost entirely unexplored. This paper presents the first study of membership inference attacks against such systems, showing that the training imprint is sparse and concentrates on uncertainty-sensitive reasoning tokens rather than spreading uniformly across outputs. The authors propose OP-Print, a unified gray-box and black-box attack framework that either aggregates privileged teacher-student residuals directly or infers them from restart-based output-basin stability. Their empirical results suggest that on-policy self-distillation introduces a distinct, previously overlooked privacy leakage surface that conventional MIA heuristics fail to capture, raising important concerns for privacy auditing of modern post-training pipelines.

High-dimensional nonparametric density estimation is notoriously plagued by the curse of dimensionality. This paper proposes to factorize the joint distribution autoregressively into a product of one marginal and $d-1$ conditional densities, estimating each factor with parallel gradient boosting (PGB), a fast tree-based multi-quantile regression method. The authors prove that for subadditive divergences the joint convergence rate is governed by the hardest factor, and that when conditional estimators adapt to sparsity the optimal rate relates to the treewidth of an underlying dependence graph. Together, the theory and experiments suggest that flexible autoregressive factorization offers a practical off-the-shelf alternative to kernels, copulas, and random forests for moderate-to-high dimensional continuous data.

Scientific inference often requires evaluating and optimizing complete trajectories, yet existing methods typically learn local one-step dynamics or train black-box generators that lack a reusable path objective. This paper proposes action learning, the task of recovering a stochastic action from trajectory observations, and introduces LEAP, which learns a path-space correction relative to a tractable reference process via noise-contrastive trajectory comparisons. By constructing a differentiable objective on complete paths, LEAP captures nonadditive temporal dependencies and enables downstream reuse for ranking, constrained inference, and sampling without retraining.

The paper presents structure-preserving reduced-order models for two high-speed aeroelastic challenge problems—a clamped panel at Mach 1.92 (RC-19) and a cantilever plate at Mach 5.85 (HyMAX). Because the experimental packages lack time-resolved responses or repeated runs, the author proposes a learned model that retains the exact physics-based potential energy, fits generalized aerodynamic and damping forces by least squares, and trains only a dissipation matrix engineered to remain positive definite. An exact discrete-gradient integrator then guarantees an energy balance to machine precision regardless of the trained parameters, aiming to show that physical structure, not large datasets, can carry predictive capability in chaotic flutter regimes.

Sequence compression is a critical bottleneck in efficient language and multimodal models, yet most architectures rely on simple pooling or MLP projections that ignore inter-token correlations. This paper reframes compression as a test-time ridge regression problem: input tokens define a regression whose fitted coefficients become the compressed vector. The resulting TTR-Compressor solves exactly in closed form via a small Cholesky factorization, introduces only $2d$ learnable parameters, and yields strong gains on long-context retrieval and visual token merging.

Graph soft prompting methods such as G-Retriever claim to inject relational graph knowledge into LLMs via learned continuous tokens, yet it remains unclear whether these tokens actually carry graph-specific information. This paper identifies and systematically investigates a pathological failure mode called token collapse, in which graph-derived tokens become nearly identical across inputs, effectively degenerating into generic soft prompts. Through causal interventions across three benchmarks, multiple graph encoders, and two LLM families, the authors show that strong task accuracy can conceal a token interface that the model does not meaningfully use, and they propose a diagnostic protocol to evaluate graph-token alignment beyond simple accuracy.

Apple Silicon dominates local LLM inference, yet nearly every published optimization is designed and benchmarked on NVIDIA server GPUs under batched serving workloads. This paper asks which speedups survive when moved to single-request, unified-memory Apple machines, evaluating four families—KV-cache pruning, speculative decoding, ported CUDA kernels, and post-training quantization—across four Apple devices with matched RTX 4090 controls. The central finding is that a reported speedup is not an intrinsic property of a technique but a joint function of technique, runtime, hardware contract, and workload regime.

RaSFNet tackles LiDAR semantic segmentation by fusing 3D point features with 2D frustum features in a projection-based pipeline. Its three proposed modules—RAGDE for ray-aligned geometric encoding, STAM for azimuth-aware global context via a scan-serpentine Mamba, and RCPFF for range-constrained cross-representation fusion—are designed to resolve projection conflicts and preserve 3D spatial correspondence. The method reports $70.2\%$ mIoU on the SemanticKITTI validation set and $77.7\%$ on the nuScenes validation set.

Modern nuclear and particle physics detectors such as the sPHENIX TPC produce increasingly large, sparse data volumes that strain conventional reconstruction. This work evaluates FM4NPP, a ~200M-parameter Mamba-based foundation model pretrained on TPC spacepoint sequences via self-supervised $k$-Next-Nearest-Neighbor prediction, on downstream track finding and particle identification. The central claim is that a single frozen backbone with lightweight adapters can outperform a traditional Cellular-Automaton seeder at low momentum and match or exceed a Gaussian-Process-Regression $dE/dx$ baseline for PID, potentially extending physics reach while simplifying analysis chains.

ReDence tackles the problem of detecting modern social engineering scams that go beyond credential phishing by performing adaptive, multimodal investigations of websites and their artifacts. Its core idea is a hybrid architecture that pairs a fine-tuned 4B-parameter agentic Investigator (Qwen3.5-4B) for reasoning and reporting with a specialized 4B-parameter Decider (Tev1) for fast binary and multiple-choice decisions, cutting latency while preserving depth. This matters because SE scams caused billions in losses in 2025 and existing detectors either lack interactive evidence gathering or are too slow to operate at scale.