RightSignal · Friday, November 13

SO-Arena: Benchmarking the Capability Spectrum and Reliability of System One Models

Lichen Li, Ziping Dong, Sicheng Feng, Qi Li, Xinchao Wang · Submitted 2026-11-11 · 72 upvotes yesterday

Preprint · not peer reviewedTrending

A benchmark for fast 'System One' models, which answer without writing out long reasoning. It covers 11 domains and also tests reliability: whether answers stay the same when the options change, and whether a model notices when no option is valid.

Key finding: Across 21 models, the best performer overall ranked only seventh for reliability.

Why it matters: A model that scores well can still be unreliable, so both need checking before you trust its decisions.

Read on arXiv

PhysLoop: Harnessing LLMs' Physics Awareness for Self-Improving Video Generation Models

Yuxuan Chen, Xiaojie Li, Hao Chen, Juncheng Ma, Kewei Zhang, Zhongjing Du · Submitted 2026-11-12 · 41 upvotes yesterday

Preprint · not peer reviewedTrending

A setup that uses language models to guide, pick and fix generated videos so they follow physics better. The improved videos then become training data for the video model.

Key finding: The authors release 100K videos and two test sets, for generation and for repair.

Why it matters: AI video often looks real but breaks physics. This is one way to fix that with little human labelling.

Read on arXiv

Actor, Judge, or Teacher? Understanding Decision Models in Long-Horizon Agents

Haochen Liu, Qianhao Zhu, Zixuan Wang, Dingming Li, Tongbo Chen, Peng Qiu · Submitted 2026-11-12 · 29 upvotes yesterday

Preprint · not peer reviewedCode releasedTrending

The authors test whether a model that is bad at acting in multi-step tasks can still be good at judging and teaching. Judging an action after the task ends shows its consequences.

Key finding: Used as a teacher, the decision model JEV raised a small policy's success from 72.8% to 82.9% on ALFWorld and from 56.8% to 71.4% on WebShop.

Why it matters: Weak agents may still be useful for training stronger ones.

Read on arXiv

Systems

Factored One-sided Adam-Moment for Practical and Scalable SOAP

Jaemyung Yu, Byeongho Heo, Sangdoo Yun, Dongyoon Han · Submitted 2026-11-12 · 12 upvotes yesterday

Preprint · not peer reviewedCode releasedTrending

FOAM is a lighter version of the SOAP training optimizer. It drops one side of SOAP's costly maths and estimates the rest more cheaply.

Key finding: From 124M to 760M parameters, it matched SOAP and Muon on validation loss while using half of SOAP's optimizer memory.

Why it matters: Cheaper optimizers make it easier to train language models faster on the same hardware.

Read on arXiv

AI safety

Humanized but Still True? Auditing Facts and Detection in AI-Text Humanizers

Ziyuan Wang, Genliang Zhu, Xiaobo Chen · Submitted 2026-10-30 · 2 upvotes yesterday

Preprint · not peer reviewedCode releasedTrending

The authors checked whether 'humanizer' tools, which rewrite AI text to dodge detectors, keep numbers, dates, names and quotes intact.

Key finding: All 11 commercial tools changed or dropped at least one such detail in 35-65% of the documents that had one.

Why it matters: Text run through these tools may contain silently changed facts.

Read on arXiv

Your RightSignal

Topics in News

SO-Arena: Benchmarking the Capability Spectrum and Reliability of System One Models

PhysLoop: Harnessing LLMs' Physics Awareness for Self-Improving Video Generation Models

Actor, Judge, or Teacher? Understanding Decision Models in Long-Horizon Agents

Factored One-sided Adam-Moment for Practical and Scalable SOAP

Humanized but Still True? Auditing Facts and Detection in AI-Text Humanizers

That’s every paper.