This is the edition of 5 November 2026. See the latest news.
News
Updated three times a day
Microsoft Research·5 November 2026
Microsoft Research describes PDD Kit, a tool that helps developers turn their intentions into a formal specification in Verus. It then proves that code produced by an AI coding agent satisfies that specification for all inputs, and relies on tests where proofs cannot be completed. The team reports using it on the login logic of a Rust storefront, where verification exposed existing bugs.
Read the original at Microsoft Research ↗
OpenAI·5 November 2026
OpenAI has unveiled two initiatives, a National Community Compact and a Compute Campus Policy Blueprint, aimed at shaping how AI infrastructure gets built. According to the announcement, these efforts seek to safeguard resources in surrounding areas, generate employment, and steer the development of such facilities in a responsible direction.
Read the original at OpenAI ↗
MIT News·5 November 2026
MIT researchers describe UQ-DAAAM, a robot memory system that flags stored object descriptions likely to be wrong and revises them. It compares captions of one object from several viewpoints to estimate uncertainty, seeks extra views, and merges the captions into a corrected entry. The team reports that, against baseline methods, it raised confidence and answer accuracy about objects, locations and timing.
Read the original at MIT News ↗
OpenAI·4 November 2026
OpenAI reports building an autonomous pipeline that located, confirmed and produced fixes for 948 local privilege-escalation flaws in the Linux kernel. The company says the system covered finding the vulnerabilities, validating them and generating patches. The work concerns automated security analysis of kernel code containing weaknesses that could let a local user obtain elevated access.
Read the original at OpenAI ↗
New research papers
arXiv
Preprints from the arXiv categories Robotics (cs.RO) and Artificial Intelligence (cs.AI). Preprints are not peer reviewed.
01
arXiv · cs.RO4 November 2026
The authors present HALO, a humanoid system trained on vision, language and motion data synthesized in a thousand procedurally generated simulated homes, so data volume grows with compute instead of human teleoperation. They report zero-shot real-world transfer with a 91% success rate over 360 trials in six unmapped environments, covering skills such as sitting, crawling under tables and pushing boxes.
Read the original at arXiv ↗
02
arXiv · cs.AI4 November 2026
The authors propose a finetuning method that progressively prunes safety-critical neurons while using preference optimization to rebuild refusal behavior in other parts of the network, aiming to make alignment less sparse. Across four models from three vendors, they report greater robustness to pruning, refusal-direction ablation and black-box jailbreaks, with attack success rates cut by up to 90% for some attacks.
Read the original at arXiv ↗
03
arXiv · cs.AI4 November 2026
The authors introduce a benchmark of 56 research tasks written by 42 physicists, in which agents must survey literature, plan, run experiments and write manuscripts, graded by deterministic checks plus expert rubrics. Across 22 models in 26 configurations, they report that none passes more than 46 of 168 rollouts, attributing failures to weak rigor and limited expert intuition.
Read the original at arXiv ↗
04
arXiv · cs.RO4 November 2026
The authors propose Conservative DeepReach, which trains a neural value function toward a subsolution of the Hamilton-Jacobi-Isaacs variational inequality using a one-sided margin loss, so the learned unsafe set should contain the true one. On systems from Air3D to a 13D quadrotor, they report the fewest unsafe states misclassified as safe and the lowest closed-loop collision rates among compared methods.
Read the original at arXiv ↗
05
arXiv · cs.RO4 November 2026
Using a LeRobot platform with a visuo-tactile sensor, the authors compare tactile RGB images, depth maps and marker-flow fields derived from one stream under identical demonstrations and policy backbone. Over 720 real-robot trials, they report tactile input raising success from 57.8% to 78.9%, with marker flow the strongest single cue and cue-specific encoding helping consistently.
Read the original at arXiv ↗
06
arXiv · cs.AI4 November 2026
The authors show that serving Mixture-of-Experts models with fewer active experts per token can raise compliance with harmful jailbreak requests while barely changing MMLU accuracy, and locate most of the effect in prompt processing and the first generated tokens. Their fix, SafeStart, runs all experts for that early phase and reportedly retains at least 85% of the decoding speedup.
Read the original at arXiv ↗
07
arXiv · cs.AI4 November 2026
The authors release a benchmark of 1,000 question-answer items, created and reviewed by chemistry and embodied AI experts without AI assistance, spanning 11 image and video task types for self-driving chemical laboratories. Evaluating 28 multimodal models, they report that models follow coarse procedures in videos but struggle to ground actions in operable regions and feasible trajectories.
Read the original at arXiv ↗
08
arXiv · cs.RO4 November 2026
The authors argue that Euclidean goal costs in JEPA world models poorly separate states at long temporal distances, and train a lightweight metric with a margin ranking loss on offline latent triples while keeping encoder and predictor frozen. Across four navigation and manipulation benchmarks, they report average planning success rising from 32.7% to 63.0%.
Read the original at arXiv ↗
AI-generated summaries. Items, titles and links come from the publishers' public news feeds and from arXiv, and each title links to the original. Summaries are written and checked automatically from the source text and are not individually reviewed by a person; the original is authoritative. Research papers are preprints, not peer reviewed. Llull Lab is independent of the publishers listed. How it is made.