← Back to all news

AI safety and alignment

Alignment, interpretability, red-teaming, bias and privacy: the research on what these systems do when they misbehave.

17,798 stories · RSS feed

arXiv Machine Learning
Sep 22

PAGE: Partition-Aware Gated KV-Cache Eviction

PAGE is a partition‑aware gated KV‑cache eviction method that reframes eviction as a per‑input admission decision. It uses a single label‑free scalar— the early‑to‑late drop in pairwise top‑k head agreement—to classify inputs into a capacity‑bound class (where eviction is catastrophic) and a dilution‑prone class (where eviction is safe or beneficial). By thresholding this drop, PAGE applies a base evictor only when necessary, reducing the harm rate in the capacity‑bound regime from 0.75 to 0.026 and achieving a 29× improvement across four models and benchmarks without retraining the evictor.

By Pankaj Kumar, Subhankar Mishra
benchmarkssafety
More like this →
arXiv Machine Learning
Sep 22

A Hybrid Attention Model Learning Unified Time-aware Patch Representation for Irregular Multivariate Time Series Forecasting

The paper introduces a hybrid attention model that learns a unified time‑aware patch representation for irregular multivariate time series (IMTS) forecasting. It employs a time‑aware patch encoding to embed variable‑length intra‑patch timestamps, a time bias attention mechanism to adjust for temporal misalignment and asynchronous cross‑channel dependencies, and a hybrid causal mask on a decoder‑only Transformer to balance historical context with autoregressive forecasting. The authors also curate VersaTSA, a 30 B‑observation dataset preserving native sampling sparsity, and demonstrate state‑of‑the‑art zero‑shot performance on three IMTS benchmarks while remaining competitive on regular MTS tasks.

By Zhihao Lin, Li Lin, Qi Zhang, Kaiwen Xia, Shuai Wang, Jialin Qiao
llmsragefficiencybenchmarkssafety
More like this →
arXiv Computer Vision
Sep 22

Think Like a World Model, Act Like a VLA: Distilling World-Model Representations into Compact Robot Policies

The paper proposes a method to distill world‑model representations into compact Vision‑Language‑Action (VLA) policies. By adding a single feature‑alignment term during VLA training, a frozen world model’s internal features are cached and the student policy learns to match them, eliminating the need for a generative future‑rolling component. The resulting lightweight policy runs in 32 ms on an RTX 5090, achieving high performance on LIBERO and RoboCasa‑GR1, and transfers effectively to real robotic hardware.

By Trung Dao, Sankalp Yamsani, Jaden Park, Joohyung Kim, Yong Jae Lee
roboticsmultimodalsafety
More like this →
arXiv Machine Learning
Sep 22

A Unified Benchmark for Dynamic Medical Treatment Reinforcement Learning

The paper introduces MedGym, a benchmark environment for dynamic medical treatment recommendation that models patient evolution in continuous time using Physics-Informed Neural Networks. It addresses gaps in existing reinforcement learning (RL) approaches by allowing evaluation of RL methods under irregular measurement intervals, personalized treatment responses, and safety considerations. MedGym enables direct comparison between discrete-time and continuous-time RL methods and supports clinically relevant metrics such as personalization and trajectory-level safety.

By Yuepeng Wang, Ken Kawano, Yoshihiko Fujisawa, Yongqi Zhou, Akifumi Wachi, Mehrshad Sadria, Lei Zhou, Richard Weiss, Katsuki Fujisawa, Ying Chen, Xin Liu, Kyoung-Sook Kim, Xiao Hu, Sebastien Gros, Xun Shen
reinforcement-learningbenchmarkssafety
More like this →
arXiv Computation and Language
Sep 22

Apollo Restore: A Foundation LLM for Historical Greek Optimized for Fill-in-the-Middle Restoration of Ancient Greek Texts

Apollo Restore is a 24‑billion‑parameter large language model fine‑tuned from Mistral Small to fill in gaps in fragmentary Ancient Greek texts using a fill‑in‑the‑middle objective. It achieves state‑of‑the‑art performance on short lacunae, placing the correct restoration among its top twenty candidates for 80.6% of documentary‑papyrus, 54.6% of literary‑papyrus, and 61.0% of stone‑inscription gaps, outperforming previous models by significant margins. In blind expert evaluations, 20 specialists preferred Apollo Restore over the strongest baseline and judged its performance at least as good as human restorations in 77% of cases, while also improving the published reading of the Vesuvius‑carbonised papyrus P.Herc. 1667.

By Hope McGovern, Anna Dolganov, Samuel Belkadi, Guillaume Kunsch, Dimitris Vlitas, David A. Smith
llmsfine-tuningsafety
More like this →
Simon Willison
Sep 21

Jev introduces a new shape of LLM - System One, aka Decision Models

TypeSafe AI’s new model, Jev, is a ‘System One’ or decision model that takes text or semi‑structured data as input and outputs floating‑point probabilities for yes/no, choice, or score questions, rather than text. It charges only for input tokens ($0.042/million) and offers free output, making it cheaper and faster than typical LLMs. The API lets users build a state object (string, array, or key‑value pairs) and query it with multiple questions, receiving confidence scores and probability distributions for each answer.

llmsragbenchmarkssafety
More like this →