RightSignal · Friday, November 6

VFlash: Fast, Accurate Visual Planning via Flash Primitive Decoding

Yu-Cheng Chou, Shihao Wang, Jin Huang, et al. · Submitted 5 Nov 2026 · 77 upvotes yesterday

Preprint · not peer reviewedTrending

Robot and driving models often plan with points, boxes and paths, but vision-language models write these one number at a time. VFlash uses a small helper model that predicts a whole shape at once.

Key finding: It reduced prediction errors on 2D and 3D robot path tasks. Early tests suggest about 25x faster generation for 3D motion forecasting, and a robot model using its plans reached 70.2% success versus 62.6%.

Why it matters: Faster planning matters for robots that must react in real time.

Read on arXiv

Harness-S: Learning from Sibling Harnesses for Generalizable Agent Training

Kangning Zhang, Wenxiang Jiao, Bowen Fang, et al. · Submitted 3 Nov 2026 · 51 upvotes yesterday

Preprint · not peer reviewedTrending

An agent 'harness' is the software around a model that runs its tools. When an agent fails a training task, this method retries it with slightly different harnesses and trains on the successful attempts.

Key finding: On AppWorld, SWE-Gym and DBBench, the trained agent did much better than normal training, even when it went back to the original harness.

Why it matters: It suggests agents can learn skills from better tooling and keep them without that tooling.

Read on arXiv

Random Is Almost Enough? Stable and Low-Cost Benchmark Compression via BestRand

George Ling, Shanshan Zhong, Lijin Yang, et al. · Submitted 31 Oct 2026 · 48 upvotes yesterday

Preprint · not peer reviewedCode releasedTrending

Testing AI models on full benchmarks is slow and costly, so people test on a small subset. The authors show that random subsets usually work well but sometimes give large errors, and propose BestRand to avoid those rare bad picks.

Key finding: Across more than 70 benchmarks, BestRand cut the risk of large errors by 50.2% and average error by 16.4%, at about half the setup cost of rival methods.

Why it matters: Cheaper, reliable testing helps smaller labs compare models.

Read on arXiv

AI safety

SAE Interprets MoE: Sparse AutoEncoders Reveal That MoE Experts Exhibit Human-Interpretable Fine-Grained Specialization

Anyuan Zhuo, Shaohua Zhang, Yuan Lin, Pinyan Lu · Submitted 5 Nov 2026 · 4 upvotes yesterday

Preprint · not peer reviewedCode releasedTrending

Mixture-of-experts models send each word to a few of many 'expert' sub-networks. The authors use 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: Experts specialise more finely 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 is a step toward checking and controlling it.

Read on arXiv

Systems

Muir: Linear Attention with Sensitivity-Allocated Gating and Matched-Filter Readout

Tushaar Gangavarapu, Gil Tabak, Clemens JS Schaefer · Submitted 5 Nov 2026 · 1 upvotes yesterday

Preprint · not peer reviewedTrending

Linear attention stores context in a fixed-size memory so long inputs stay cheap. Muir changes how that memory is erased, written and read, and adds a stable way to train it.

Key finding: From 166M to 1.34B parameters it lowered perplexity by 2.5–3.6% versus a Transformer, trained 2.7x faster at 32K length, and kept near-perfect retrieval out to 1 million tokens.

Why it matters: Cheaper long-context models could make very long documents affordable to process.

Read on arXiv

Your RightSignal

Topics in News

VFlash: Fast, Accurate Visual Planning via Flash Primitive Decoding

Harness-S: Learning from Sibling Harnesses for Generalizable Agent Training

Random Is Almost Enough? Stable and Low-Cost Benchmark Compression via BestRand

SAE Interprets MoE: Sparse AutoEncoders Reveal That MoE Experts Exhibit Human-Interpretable Fine-Grained Specialization

Muir: Linear Attention with Sensitivity-Allocated Gating and Matched-Filter Readout

That’s every paper.