Safety Geometry Collapse in Multimodal LLMs and Adaptive Drift Correction
Safety Geometry Collapse: multimodal LLMs fail safety transfer due to representation drift; proposes adaptive correction mechanism.
Search the full wire by company, model, lab, or keyword. Every story we have ever aggregated.
Safety Geometry Collapse: multimodal LLMs fail safety transfer due to representation drift; proposes adaptive correction mechanism.
Novel framework for routing problems with asymmetric hybrid geometries (points, lines, areas) using learned representation and decision-making.
Reddit user shares personal token consumption metrics across Anthropic accounts without analysis or comparative insight.
IVGT learns implicit 3D scene geometry from multi-view images using continuous neural representations and transformers.
Automated evaluation framework for design video generation fidelity across layout, motion, temporal, and content dimensions.
QSurv: deep learning framework for nonparametric continuous-time survival modeling using Gauss-Legendre quadrature.
Skew-adaptive conformal prediction via hyperbolic sine transform and learned gauge function for asymmetric uncertainty quantification.
Identifies premature exploitation in LLM agents; proposes Exploration Checkpoint Coverage metric and training methods for autonomous environment discovery.
LLQR+SAM combines sharpness-aware minimization with learned preconditioner for geometry-informed optimization.
Entropy-rate discretization scheduler for flow matching and Schrödinger bridges derived from conditional-marginal geometry without training.
Unified framework for metric-based machine-generated text detection addresses token-level bias via contextual relation modeling.
SAFE: variational quantum classifier with bounded observables and SAFE-AI metrics for reliability across accuracy, robustness, explainability.
PDI-Bench: Quantitative framework for auditing geometric coherence in generated video via perspective distortion and point-tracking metrics.
Tensor similarity: Weight-based metric for mechanistic interpretability that is invariant to symmetries and captures cross-layer functional equivalence.
RoSHAP proposes distributional framework and robust metric for stable feature attribution rankings amid stochastic variation.
CLOVER closes training-evaluation mismatch in end-to-end autonomous driving by learning value estimates aligned with rule-based planning metrics.
GPart achieves end-to-end isometric parameter-efficient fine-tuning via global partitioning, improving on LoRA.
Study shows LLM-based financial governance lacks behavioral compliance; proposes five rationale-level metrics and mechanical enforcement approaches.
Task-aware layer pruning improves OOD generalization but not ID accuracy in LLMs; geometric explanation via norm/distance profile divergence.
Study applies Goldstone modes from physics to analyze stable information propagation in equivariant deep networks across depth.
Cohere discusses metrics and integration strategies for achieving AI-native enterprise operations.
Method detects step-level hallucinations in LLM reasoning by monitoring hidden-state trajectory geometry during inference.
Neural ODE surrogate model for metropolitan flood forecasting using conditional latent dynamics network.
Boris Mann critiques vague agent counting as meaningless metric, comparing to spreadsheet or tab counts without context.
Study argues generative AI in education boosts performance metrics but undermines deep learning and metacognition.
In today’s data-driven world, organizations increasingly rely on video to capture critical information, yet extracting meaningful, real-time insights from... In today’s data-driven world, organizations increasingly rely on video to capture critical information, yet extracting…
Study reveals geometric coupling between routers and experts in sparse MoE models, explaining routing collapse and informing load-balancing improvements.
Theoretical analysis of geometric memorization in transformers showing embeddings encode relational structure vs. linear parameter scaling.
Study quantifies sample efficiency in Predictive Coding vs Backpropagation using target alignment metric, finding PC enables more efficient learning in small-scale experiments.
Empirical study of supervised fine-tuning on procedural skills across Qwen3.5 0.8B-4B models shows W-shaped pre-training trajectory and uniform SFT gains.