New Tool Detects Drift in Language Model Safety
Researchers propose a calibrated instrument to detect when language models' certified safe behavior degrades over time.
Research Digest
Preprints and papers, rewritten in plain English.
Token-Budgeted Sampling for LLM Reasoning
A new method allocates a fixed token budget across reasoning attempts, improving test-time scaling efficiency.
Graph Attention Model Targets Early Ventricular Reentry Prediction
A new approach combines physiological knowledge with graph attention to forecast the abnormal heart rhythms that can lead to sudden cardiac death.
PolyGraph: Buffered CUDA Graphs for Adaptive Speculative Decoding
A new technique adapts CUDA Graphs to variable batch sizes and draft lengths in speculative decoding, avoiding the fixed-configuration bottleneck.
Probing Concept Recovery in In-Context Learning
A new arXiv preprint argues that prediction accuracy alone does not show whether a language model has truly recovered the underlying concept from demonstrations.
Agentic RL Needs Task-Level Scaling Laws, Not Just Token Counts
A new arXiv paper argues that scaling laws for agentic reinforcement learning must account for task diversity and synthesis strategies, not just raw token counts.
Compaction-Aware RL Improves Ascend Triton Kernel Generation
A new reinforcement learning approach targets key obstacles in LLM-based kernel generation for Ascend hardware.
Decision model for web agents weighs request budgets and failure risk
A proposed agentic reinforcement learning model treats real-world web constraints as explicit costs rather than optimizing task success alone.
DISCERN: Causal model targets shortcut learning in medical AI
A new latent structural causal model aims to make clinical multimodal models generalize across patient cohorts by separating disease evidence from background factors.
StrucChunk: Structure-Aware Retrieval for Legal RAG
A new arXiv paper proposes structure-aware chunking to keep legal arguments intact in retrieval-augmented generation.
Neural Optimal Transport for Subset Alignment
A new framework adapts optimal transport to focus on relevant subsets of a target distribution, enabling more targeted distribution alignment.
Torch-DPhys: A PyTorch-Native Differentiable Physics Framework
A new framework brings differentiable thermal-hydraulic simulation to PyTorch while remaining compatible with the MOOSE ecosystem.
How Looped Transformers Tackle Depth-Bound Algorithmic Tasks
A mechanistic study examines how weight-tied 'looped' transformers can solve multi-step algorithmic tasks without scaling up depth or chain-of-thought length.
BEM: A Bregman-Divergence Approach to Concept Erasure in Flow Matching
A new preprint proposes learning Bregman divergences to remove unwanted concepts from flow-matching text-to-image models like FLUX and Stable Diffusion 3.
Boundary-Aware Location Encoders Trained via Reverse Geocoding
A new pretraining method uses administrative boundaries from reverse geocoding to improve location encoders for geospatial prediction.
Multi-Step Inertial ADMM Proposed for Federated Learning
The paper introduces a multi-step inertial variant of ADMM aimed at improving optimization in federated learning.
Post-diagnosis layer aims to explain symbolic regression failures
A proposed search-independent diagnostic step could give symbolic regression solvers richer feedback than a single scalar fitness score.
Neural Representations Follow Temporal Trajectories, New Study Shows
A new arXiv paper examines how neural network representations evolve over time, building on the known low-rank simplicity bias.