arXiv:2608. 11212v1 Announce Type: new Abstract: Top-k Mixture-of-Experts (MoE) routing is discontinuous, so a deployment-motivated numerical disturbance -- simulated 4-bit KV-cache quantization read by a protected BF16 gate -- pushes tokens across decision boundaries and flips which experts fire.
arXiv:2607. 25532v1 Announce Type: new Abstract: Consider a model trained at a single hospital to predict patient recovery, where the measured feature $X$ bundles the patient's true health signal ($C$) with a systematic artefact from that hospital's equipment ($S$).
arXiv:2608. 15787v1 Announce Type: cross Abstract: Two Mixture-of-Experts (MoE) forward passes can share every weight yet route the same token through different experts.
The paper investigates test‑time adaptation for medical image segmentation, showing that a fixed adaptation horizon can harm many individual cases. It introduces prediction fragmentation—a measure of disagreement between the source model and the adapted mask—to predict harmful adaptation without extra labels or backward passes. Using a case‑level router based on this metric, the authors reduce harmful adaptation on cardiac MRI from 58.7% to 20% while maintaining accuracy.
The paper introduces a method for deciding whether to adapt a frozen segmentation model at test time, arguing that a fixed adaptation horizon conflates two distinct decisions: how far to adapt and whether to adapt at all. By measuring disagreement geometry—called prediction fragmentation—between the source model and the adapted mask, the authors predict harmful accepted area (HA) without extra labels or backward passes, achieving strong correlation across three medical benchmarks. A case‑level router built on this metric reduces HA significantly while maintaining or improving Dice scores, and the approach generalizes across architectures and domains.
The paper investigates when auxiliary context can genuinely improve multi‑modal time series forecasting. It identifies two necessary dataset‑level conditions: the target must not be dominated by a last‑value shortcut (low autocorrelation) and the context must provide additional information beyond history (non‑zero conditional mutual information). Experiments on a large mixture‑of‑experts model and several fusion mechanisms show that only when both conditions hold does context routing yield a substantial reduction in mean‑squared error; otherwise its contribution collapses to a capacity floor.
arXiv:2607. 28308v1 Announce Type: new Abstract: Sparse mixture-of-experts (MoE) language models route each token to multiple experts, suggesting a geometric account of their benefit: co-selected experts should contribute distinct representation directions.
arXiv:2606. 10703v1 Announce Type: new Abstract: Interpretability methods routinely use population-level summary statistics over observed model behaviour to license claims about the effects of targeted interventions on specific computations; in Pearl's terms, they treat rung-1 associational evidence as if it supported rung-2 interventional conclusions, a move whose validity is rarely tested.
Sparse mixture-of-experts (MoE) language models route each token to multiple experts, suggesting a geometric account of their benefit: co-selected experts should contribute distinct representation directions. Existing evidence often conflates route coherence, candidate quality, and candidate-by-context interaction.
Signal‑Routed Temperature Scaling (SRTS‑BCE) is a 10‑parameter, argmax‑preserving calibration method that separates calibration objectives from adaptive capacity. It cross‑fits a correctness‑risk score over six logit statistics and assigns a top‑label BCE temperature to each of three risk groups, generalizing TvA‑TS when K=1. Experiments on fine‑tuned CIFAR‑100 and ViT‑B/16 show that SRTS‑BCE reduces ECE from 1.65 to 0.96 with a small calibration budget, outperforming higher‑capacity SMART+BCE when only 250 examples are available, and revealing a budget‑dependent ranking reversal on Swin‑T.
The paper introduces GRIP, an algorithm‑agnostic framework for machine unlearning in Mixture‑of‑Experts large language models. GRIP enforces hard geometric constraints on router updates, projecting gradient changes into the null space of the retain set’s routing matrix to prevent routing manipulation. Two variants—training‑time stochastic projection and post‑training analytical correction—show significant improvements in routing stability, retain accuracy, and resistance to white‑box adversarial recovery across two MoE models.
arXiv:2608. 14650v1 Announce Type: new Abstract: Existing adaptive-inference and world-action-model systems use cheap-stage outputs or predicted futures to allocate additional computation.