RightSignal · Saturday, November 7

VFlash: robots plan whole paths in one step instead of one number at a time

Yu-Cheng Chou, Shihao Wang, Jin Huang, et al. · Submitted 5 Nov 2026 · 98 upvotes yesterday

Preprint · not peer reviewedTrending

AI models that plan robot motion usually write out points, boxes and paths one coordinate at a time, which gets slow for long paths. VFlash adds a small helper model that predicts a whole shape or path at once.

Key finding: It made fewer prediction errors than step-by-step decoding on the robot trajectory tasks tested, and early measurements show about 25x faster generation for 3D motion forecasting.

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

Read on arXiv

BestRand: testing AI on a small slice of a benchmark, more reliably

George Ling, Shanshan Zhong, Lijin Yang, et al. · Submitted 31 Oct 2026 · 47 upvotes yesterday

Preprint · not peer reviewedCode releasedTrending

Running full benchmarks is expensive, so people test models on a subset. The authors show that random subsets usually work well but sometimes give big errors, and propose a cheap method that avoids those rare bad picks.

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

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

Read on arXiv

SwarmEvolve: hundreds of small AI agents search for better programs without writing text

Qile Jiang, George Em Karniadakis · Submitted 5 Nov 2026 · 6 upvotes yesterday

Preprint · not peer reviewedTrending

AI-guided program search is slow because the model writes long reasoning text every round. SwarmEvolve has many cheap agents reason internally and share their working memory directly with each other.

Key finding: On 21 maths and scientific computing problems it reached the best scores other frameworks reach in 60 minutes in a median of 3 to 8 minutes, and with the small Qwen3 8B model it improved on some results first found by AlphaEvolve.

Why it matters: It suggests useful scientific search could be done with small, cheap models instead of frontier ones.

Read on arXiv

Systems

Muir: a cheaper attention design that handles very long text without slowing down

Tushaar Gangavarapu, Gil Tabak, Clemens JS Schaefer · Submitted 5 Nov 2026 · 6 upvotes yesterday

Preprint · not peer reviewedTrending

Linear attention keeps a fixed-size memory of the text so far, making long inputs cheap. The authors find which parts of that memory update matter most, shrink the rest and spend the savings elsewhere.

Key finding: At sizes from 166M to 1.34B parameters it lowered perplexity by 2.5-3.6% compared with a standard Transformer, kept near-100% passkey retrieval up to 1M tokens, and trained 2.7x faster than a Transformer at 32K length.

Why it matters: Designs like this could let models read very long documents at a fraction of today's memory cost.

Read on arXiv

AI safety

Experts inside mixture-of-experts models really do specialise, study finds

Anyuan Zhuo, Shaohua Zhang, Yuan Lin, Pinyan Lu · Submitted 5 Nov 2026 · 6 upvotes yesterday

Preprint · not peer reviewedCode releasedTrending

Mixture-of-experts models send each word to a few of many sub-networks. The authors used a sparse autoencoder, a tool that breaks a model's internal 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, and the split is finer than broad topics: separate experts handle Java versus front-end web code, or geometry versus arithmetic.

Why it matters: Knowing what each part of a model does helps people inspect, debug and trust it.

Read on arXiv

Your RightSignal

Topics in News

VFlash: robots plan whole paths in one step instead of one number at a time

BestRand: testing AI on a small slice of a benchmark, more reliably

SwarmEvolve: hundreds of small AI agents search for better programs without writing text

Muir: a cheaper attention design that handles very long text without slowing down

Experts inside mixture-of-experts models really do specialise, study finds

That’s every paper.