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.

Hybrid retrieval—fusing lexical BM25 with dense neural retrievers—is the default recommendation in modern search stacks, usually justified by aggregate benchmark gains. This paper challenges that consensus by diagnosing what it calls the "fusion tax": the frequency and severity with which fusion hurts individual queries relative to the better of its two components or to a deployable dense-only baseline. Through exhaustive experiments across seven BEIR corpora, five encoders spanning 22M–596M parameters, and multiple fusion functions, the authors show the tax is material, pervasive, and undiminished by encoder scale, and they derive practical guidance for when to fuse and which method to choose.

Time series databases store billions of points, yet answering natural-language shape queries like "plunge and then recover quickly" remains hard because large language models must bridge free-form language to raw numeric windows. TT-SAX proposes a minimal redesign of SAX that adds a second track of trend letters measured across neighboring segments, letting an LLM compile shape queries into regular expressions that can be matched efficiently at database scale. The paper is significant because it reframes symbolization design around LLM usability rather than classical distance or compression metrics.

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.

PiHand tackles the synthesis of physically plausible dexterous hand-object interactions by learning a latent interaction prior from simulated mocap-tracking experts. Its three-stage pipeline—expert RL tracking, VAE distillation, and bounded residual latent exploration—enables zero-shot manipulation of unseen objects and trajectories, and even transfers to tasks like catching and bimanual handover that are rare or absent from the training data. The core thesis is that embedding interaction physics directly into the latent space outperforms both raw action-space RL and hand-only priors in sample efficiency and success rate.

AnyGameRenderer tackles the programmer-art problem of agent-built games by converting crude gameplay video into stylized, high-fidelity output via a rank-128 LoRA on the 33B MiniMax-H3 ref2va model. It accepts plain video inputs---gray blockouts, colored primitives, semantic maps, or wireframes---without engine G-buffers or per-object labels, and aims to preserve entity count, timing, and position across genres from tower defense to kart racing. The authors assemble roughly 10k paired clips from three sources and introduce an entity-level benchmark of 753 annotated entities to measure correspondence, improving temporal match over the zero-shot base.

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.

MixPT tackles the trade-off between computational cost and scene-wide interaction in point cloud transformers. It interleaves fixed-order window attention blocks with periodic global attention blocks, using spatial pooling to keep global attention feasible when the number of tokens is large. The authors show that this hybrid schedule lowers per-forward latency compared to both pure window and pure global baselines while maintaining competitive accuracy on indoor and outdoor 3D segmentation benchmarks.

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.

This paper tackles cross-session indoor navigation, where an agent must reach a prior-session goal despite scene changes, sensor differences, and partial overlap. The key idea is a sparse topometric graph whose nodes compress local 3D geometry around traversable anchors and whose edges encode traversable distance, enabling both memory-efficient storage and learned graph matching for alignment. The approach matters because it promises to slash map memory by over two orders of magnitude while outperforming dense grids, appearance maps, and semantic graphs on standard benchmarks.

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.

This paper attacks the real-time deployment bottleneck for polyp segmentation by proposing a \sim77K-parameter Lightweight Frame Compensator (LFC) inserted before a frozen backbone. Under $K=3$ frame skipping, LFC fuses three consecutive frames via spatio-temporal attention into a single compensated image, claiming a $2.05\times$--$2.55\times$ throughput gain while recovering most accuracy loss. The core insight is that global-attention Transformers suffer the most from skip-frame jitter, making them the prime beneficiaries of lightweight temporal fusion. Unfortunately, the bulk of the evidence rests on pseudo-labels rather than expert-annotated video ground truth, which undermines the clinical certainty of the reported gains.

Existing vision-language models struggle with fine-grained visual perception, often hallucinating objects or missing fine details. WeVisOne-VL tackles this by unifying object-level and pixel-level tasks—detection, segmentation, matting, OCR, and more—inside a single autoregressive VLM built on Qwen3-VL. It introduces dedicated coordinate tokens and a lightweight 9.6M-parameter pixel decoder triggered by an <OBJ> token, enabling dense outputs without a separate segmentation backbone. Trained on roughly 100M curated images via a three-stage curriculum and mixed-task RL, it aims to move VLMs from broad understanding toward reliable, deployable perception.

Local feature matching underpins SfM and visual localization, but standard matchers process every image pair independently, which is inefficient and ignores multi-view cues available in star-graph neighbourhoods. MVLoMa addresses this by jointly matching a centre image against all neighbours in a single forward pass, reusing frozen geometry-aware tokens from a multi-view foundation model such as DA3-Giant and adding star-structured cross-attention so the centre aggregates evidence from all neighbours. It preserves the conventional sparse pairwise outputs required by RANSAC while improving strict-threshold pose AUC and cutting matcher compute by roughly $2.2\times$ at $25$ views.