← Llull Lab· Paris research laboratory
As of 6 November 2026 · Updated three times a dayLlull Lab · Paris

AI & Robotics News

Research and policy news on artificial intelligence and robotics, updated three times a day from the public feeds of research laboratories, universities and EU institutions, with a daily selection of new research papers. Each item has a short summary and a link to the source.

This is the edition of 6 November 2026. See the latest news.

News

Updated three times a day
European Commission6 November 2026

Commission preliminarily finds TikTok’s ‘For You’ feed in breach of Digital Services Act for harmful ‘rabbit holes’ effect

The European Commission has reached a preliminary finding that TikTok's For You feed breaches the Digital Services Act. The Commission says the recommender system keeps serving similar personalised videos as users engage, including material on mental health and body image, which it considers harmful for minors and vulnerable adults. The findings cover both risk assessment and mitigation measures.

Read the original at European Commission ↗
Amazon Science5 November 2026

Hamed Karimi wins ECML PKDD Test-of-Time Award

At ECML PKDD in Naples, Amazon senior applied scientist Hamed Karimi and coauthors Julie Nutini and Mark Schmidt received the 2026 Test-of-Time Award for a 2016 paper. According to Amazon, the work used the Polyak-Łojasiewicz condition, a 1963 inequality, to show that standard gradient methods converge rapidly toward optimal solutions even when many solutions are equally good.

Read the original at Amazon Science ↗
Apple Machine Learning Research5 November 2026

Faster Retrieval for Better, More Efficient RAG

Apple machine learning researchers describe work on maximum inner product search, the step in which a query vector is matched against millions of stored key vectors in recommendation, search and retrieval-augmented generation. The post explains that exhaustive search becomes costly as key counts reach millions and dimensions reach thousands, and it presents an approach aimed at faster retrieval.

Read the original at Apple Machine Learning Research ↗
IBM Research5 November 2026

How MEMS inertial sensors can rival GPS

IBM Research describes how inertial sensors built with microelectromechanical systems might compete with GPS for navigation. The article attributes this to progress in self-calibration methods, which it says are giving renewed relevance to an older form of navigation that had fallen out of favour.

Read the original at IBM Research ↗
IBM Research5 November 2026

ORNL, IBM, AMD launch co-design platform for QEC

IBM Research reports that Oak Ridge National Laboratory, IBM and AMD have launched a co-design platform for quantum error correction. According to the announcement, the U.S. Department of Energy's Quantum Science Center at Oak Ridge leads the effort to connect IBM quantum hardware with the laboratory's high-performance computing clusters, which run on AMD technology.

Read the original at IBM Research ↗

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.RO5 November 2026

Failure Prediction and Classification in Vision-Language-Action Models Via Calibrated Latent Probing

The authors present VLA-Tug, a lightweight probe on the latent layers of a vision-language-action backbone that flags upcoming failures with bounded miss and false-alarm rates and predicts the failure mode. They report earlier detection than learned baselines, bounds that stay valid more often on unseen tasks, and a released set of 3,180 annotated episodes.

Read the original at arXiv ↗
02
arXiv · cs.AI5 November 2026

Learning the Task or Learning the Verifier? Rethinking Criteria Exposure in Embodied Self-Improvement

The authors study self-improving embodied coding agents that are shown a benchmark's success criteria. They report that this exposure quickly raises scores under the given criteria, but the gain vanishes when the criteria are slightly altered and skills transfer poorly. They then propose instruction-faithful criteria as a more reliable measure.

Read the original at arXiv ↗
03
arXiv · cs.RO5 November 2026

HITCH: Teaching a Vision-Language Model to Operate a Humanoid through a Whole-Body Harness

The authors introduce HITCH, a harness that lets a vision-language model query 3D points and command a humanoid's pelvis, torso, hands, walking and gaze through a learned whole-body tracker. After fine-tuning on teacher trajectories, they report success on unseen benchmark episodes and on a real G1 that carries a box to a goal without real training data.

Read the original at arXiv ↗
04
arXiv · cs.RO5 November 2026

Social Humanoid: Socially Responsive Human-Humanoid Interaction

The authors present Social Humanoid, a hierarchical system in which a tokenized sequence model generates reactive whole-body motion for a Unitree G1 from a human partner's movement, executed by a tracking controller. They build a 260-hour interaction benchmark via adaptive retargeting and report coordinated behavior in simulation and on hardware.

Read the original at arXiv ↗
05
arXiv · cs.RO5 November 2026

RoboARCade: Benchmarking Agentic Robot Control

The authors introduce RoboARCade, a MuJoCo benchmark of long-horizon manipulation tasks for agents that act directly as robot policies across three embodiments. They report that leading agents stay below half success, mostly fail through poor strategy or imprecise execution, often wrongly claim success, and improve when given a single demonstration in context.

Read the original at arXiv ↗
06
arXiv · cs.AI5 November 2026

The Geometry of Simulated Perspectives: Dissociating Representation and Behaviour in a Multimodal Language Model

The authors examine how persona prompts reshape hidden-state geometry in a multimodal model rating visual interestingness across 100 images and 53 conditions. They report that prompt framing changes internal representations more than persona identity, while persona content drives behavioural ratings more, suggesting representational distance need not match behavioural relevance.

Read the original at arXiv ↗
07
arXiv · cs.AI5 November 2026

A Behavioural Framework for Predicting and Understanding Loss of Control in Frontier Artificial Intelligence Systems

The authors propose a behavioural framework that splits loss-of-control risk into monitoring evasion, misaligned motive, and harm-enabling capability, measured through 13 aspects and 29 tasks merged into one LoC score. Testing 13 frontier models, they report that this score tracked failure rates in contained long-horizon settings (Pearson r = 0.84), and that targeted fine-tuning and monitoring lowered failures.

Read the original at arXiv ↗
08
arXiv · cs.AI5 November 2026

Benchmarks as Software: Terminal-Bench and the Case for Continuous Benchmarks

The authors argue that agentic benchmarks build up maintenance debt as models find correct answers that verifiers reject and infrastructure shifts. Applying a continuous-validation workflow to Terminal-Bench 2.0, they found problems in about 30% of it and released fixes as Terminal-Bench 2.1, which raised accuracy over 10 percentage points and shifted ranks by up to three (Spearman rho = 0.958).

Read the original at arXiv ↗
09
arXiv · cs.RO5 November 2026

GPWM: Modeling the World as 3D Particle Motion

GPWM is a generative model that forecasts how 3D scenes change by producing particle and rigid body motion, using one architecture for rigid, deformable, cloth, and granular materials. Trained on the 25.5-million-frame ParticleWorlds corpus, it reportedly lowers prediction error by 31% on average versus the strongest prior baseline across six domains and supports contact-rich planning.

Read the original at arXiv ↗
10
arXiv · cs.AI5 November 2026

PharmAgents2: Towards Autonomous End-to-End Drug Discovery

PharmAgents2 is an autonomous drug discovery system that pairs Paper2Skill knowledge extraction with Self-Scaling Discovery parallel agents, assessed on PharmAgents-Bench v0.1 with 121 questions and three end-to-end tasks. The authors report it beat baselines in most categories and, in real studies, proposed a new allosteric site on LpxH yielding a confirmed inhibitor scaffold.

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.