This is the edition of 7 November 2026. See the latest news.
News
Updated three times a day
OpenAI·6 November 2026
OpenAI introduces BioBasicsBench, a benchmark of 100 tasks built to measure how AI agents interpret raw experimental data and deal with routine wet-lab science work. The company presents it as a way to assess agent performance on ordinary laboratory biology problems.
Read the original at OpenAI ↗
MIT News·6 November 2026
MIT researchers, together with three scientists from the chemical company BASF, report a machine learning method for predicting molecular properties when little data exist, such as ligand affinities for a newly found drug target. They say it outperforms existing methods at narrowing down candidate molecules. The work appeared in the Journal of Chemical Information and Modeling, and the model is open source.
Read the original at MIT News ↗
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 Preprints are not peer reviewed.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
The authors present OopsieRecover, which applies abductive inference to a few corrective robot demonstrations to recover their intent, encoding it as dense spatial representations and conditional costs that push a VLA policy away from undesirable actions. Across 6 tasks, they report that a handful of corrections let the policy overcome failures in novel object poses, scenes and related tasks.
Read the original at arXiv ↗
02
arXiv · cs.AI5 November 2026
The authors introduce MCPHallu, a benchmark of 358 tasks on 36 production MCP servers, each built with a structural trigger targeting a reasoning, execution or memory failure subtype. Evaluating fourteen LLM agents, they report weakest results on infeasible goals, where agents often keep calling tools or claim success instead of recognizing impossibility.
Read the original at arXiv ↗
03
arXiv · cs.AI4 November 2026
The authors connect information-theoretic empowerment with capability-based human power in Markov games through goal revelation, the mutual information between a human's latent goal and the trajectory, establishing a chain of bounds between them. They use the chain to describe failure modes, such as assistants maximizing empowerment through influence that is irrelevant to the human's goals.
Read the original at arXiv ↗
04
arXiv · cs.RO5 November 2026
The authors propose STEP, a one-stage training scheme for perceptive humanoid locomotion that combines bidirectional masked prediction between depth and proprioception with an action-conditioned predictor of the next fused state. They report better terrain traversal in simulation and under degraded vision, along with deployment on a 23-DoF Noetix E1 humanoid.
Read the original at arXiv ↗
05
arXiv · cs.RO5 November 2026
The authors describe AffectiveVLA, a pipeline for social touch on a Unitree G1 humanoid in which an RGB intent recognizer prompts a VLA policy, suppresses motion when confidence is low, and works alongside compliance profiles and a voice-triggered safe stop. They report that intent gating raises GR00T's macro-F1 from 0.75 to 0.90.
Read the original at arXiv ↗
06
arXiv · cs.AI5 November 2026
The authors set up an experimental framework that varies information disclosure policies in markets of LLM agents engaged in trade, bargaining and double auctions, while keeping rules and models fixed. Across ten models, they report that the effects of finer disclosure depend strongly on the model, and that full disclosure in auctions caused market breakdown for several models.
Read the original at arXiv ↗
07
arXiv · cs.AI5 November 2026
The authors present FINCOSMOS, an open-world evaluation in which agents gather web data and build backtests for a strategy that a simulated user reveals only on request, with programmatic checks and veto rules for grading. Across 150 sessions with four coding agents and a tool-less model, they report that no session was accepted.
Read the original at arXiv ↗
08
arXiv · cs.AI5 November 2026
The authors build an analytical framework using an invariant box and an exit-time argument to study token dynamics in deep Mamba models, proving exponential convergence and showing that the first token attracts later tokens. They link this to clustering and attention sinks and report agreement with experiments on Falcon-Mamba-7B.
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.