RightSignal · Sunday, November 8

VFlash: robot planning models draw whole paths at once instead of one number at a time

Yu-Cheng Chou, Shihao Wang, Jin Huang, Chi-Pin Huang, Chiao-An Yang, Xianzhe Fan · Submitted 5 Nov 2026 · 104 upvotes yesterday

Preprint · not peer reviewedTrending

Vision-language models used in robotics write plans such as boxes and paths one coordinate at a time, which is slow. VFlash adds a small helper that predicts a whole shape in one pass and accepts it if it is valid.

Key finding: It cut prediction errors on 2D and 3D robot path tasks, ran about 25 times faster in early tests on 3D motion forecasting, and lifted a robot model's success from 62.6% to 70.2%.

Why it matters: Faster and more accurate planning matters for robots and self-driving systems that must react in real time.

Read on arXiv

BestRand: testing AI models on a small random sample of a benchmark, done more safely

George Ling, Shanshan Zhong, Lijin Yang, Hao Yang, Nanxiang Li, Zhongzhan Huang · Submitted 31 Oct 2026 · 50 upvotes yesterday

Preprint · not peer reviewedCode releasedTrending

Running full benchmarks is costly, so people test on a subset. The authors show random subsets are usually accurate but sometimes badly wrong, and propose BestRand to avoid those rare bad picks.

Key finding: Across more than 70 benchmarks, BestRand reduced high-error risk by 50.2% and average error by 16.4%, at 51.3% of the construction cost of competing methods.

Why it matters: Cheaper, reliable evaluation lets smaller teams compare models without running every test.

Read on arXiv

Muir: a cheaper kind of AI model memory that handles 10-million-word streams

Tushaar Gangavarapu, Gil Tabak, Clemens JS Schaefer · Submitted 5 Nov 2026 · 12 upvotes yesterday

Preprint · not peer reviewedTrending

Linear attention models keep a fixed-size memory of past text, so they use constant memory no matter how long the input. Muir redesigns how this memory is erased, written and read, and spends the saved parameters elsewhere.

Key finding: From 166M to 1.34B parameters it lowered perplexity by 2.5–3.6% versus a Transformer, ran to 10M-token streams without degrading after training on 4K, and trained 2.7 times faster than a Transformer at 32K.

Why it matters: Models with fixed-size memory could make very long documents and conversations much cheaper to process.

Read on arXiv

SwarmEvolve: hundreds of small AI agents share thoughts directly to solve maths problems faster

Qile Jiang, George Em Karniadakis · Submitted 5 Nov 2026 · 9 upvotes yesterday

Preprint · not peer reviewedTrending

AI-guided search for better programs is slow because each agent writes out long reasoning in text. Here many cheap agents reason internally and pass their working memory to each other instead of text.

Key finding: On 21 maths and science computing problems it reached in a median 3–8 minutes the scores other frameworks reach in 60, and with Qwen3 8B improved on some results AlphaEvolve found with larger models.

Why it matters: It suggests that small, cheap models working together can match expensive ones on some search tasks.

Read on arXiv

AI safety

Researchers show the 'experts' inside Mixture-of-Experts models have narrow, readable jobs

Anyuan Zhuo, Shaohua Zhang, Yuan Lin, Pinyan Lu · Submitted 5 Nov 2026 · 8 upvotes yesterday

Preprint · not peer reviewedCode releasedTrending

Mixture-of-Experts models send each word to a few of many sub-networks. The authors used sparse autoencoders, a tool that splits a model's signals into readable features, to see what each expert actually contributes in Qwen3-30B-A3B.

Key finding: Most features are driven by just one or two experts, with separate experts for things like Java versus front-end web code, or geometry versus arithmetic.

Why it matters: Knowing what each part of a model does helps people check and control how it behaves.

Read on arXiv

Your RightSignal

Topics in News

VFlash: robot planning models draw whole paths at once instead of one number at a time

BestRand: testing AI models on a small random sample of a benchmark, done more safely

Muir: a cheaper kind of AI model memory that handles 10-million-word streams

SwarmEvolve: hundreds of small AI agents share thoughts directly to solve maths problems faster

Researchers show the 'experts' inside Mixture-of-Experts models have narrow, readable jobs

That’s every paper.