RightSignal · Wednesday, November 4

RoboColosseum: testing robot AI in thousands of blind trials at 124 sites

Tan Sang Nguyen, Yi Ru Wang, Shailesh Shailesh, et al. · Submitted 2 Nov 2026 · 20 upvotes yesterday

Preprint · not peer reviewedTrending

The authors built a network of locally run robots to compare robot control models head to head in blinded trials, across four robot types from arms to a humanoid.

Key finding: Across 5,699 comparisons, general AI agents rivalled specialised robot models, but all struggled with open-world tasks. On the humanoid, no model finished any of five tasks.

Why it matters: Robot demos are easy to cherry-pick. This shows rankings change a lot between robots and after fine-tuning, so one benchmark result says little.

Read on arXiv

Systems

SpecPilot: AI agents tune a faster way to run language models

Kunjun Li, Wenhao Chai, Yihao Liang, et al. · Submitted 2 Nov 2026 · 11 upvotes yesterday

Preprint · not peer reviewedTrending

Speculative decoding lets a small model draft text that a big model checks. SpecPilot runs agents in a loop that look for the biggest gaps in a setup, try fixes and check every output token stays the same.

Key finding: Without training or quality loss, it raised the speed-up over normal decoding from 4.46x to 10.87x for one request, and from 1.94x to 3.93x with batching.

Why it matters: Off-the-shelf serving code leaves a lot of speed unused, and each deployment needs different tuning.

Read on arXiv

Near-zero word vectors silently stall training in some newer model designs

Xin Liu, Boyuan Sun, Yuxin Wang, et al. · Submitted 3 Nov 2026 · 5 upvotes yesterday

Preprint · not peer reviewedTrending

The authors found that rarely seen tokens can shrink to almost zero during training. In gated models such as Mamba, these blow up gradients until a common clipping step overflows and throws away the whole batch's update.

Key finding: In one Telugu training run, about 20% of steps were lost. Over half of the untied gated checkpoints they checked had such rows. They propose two fixes.

Why it matters: It is a quiet bug that wastes compute without crashing, in widely used tools such as PyTorch and Hugging Face Trainer.

Read on arXiv

AI safety

Caper: try a coding agent's command in a sandbox before allowing it

Weiheng Wu, Junkai Chen, Zhensu Sun, et al. · Submitted 29 Oct 2026 · 1 upvotes yesterday

Preprint · not peer reviewedTrending

Instead of judging a shell command by its text, Caper runs it in a snapshot first, shows an approver what actually changed, then keeps it or rolls it back with a reason.

Key finding: Harmful outcomes fell to 7.8% versus at least 35.3% for six existing approaches on a benchmark where harm hides inside tools, while blocking 1.0% of safe tasks.

Why it matters: Coding agents increasingly run commands unattended, and a harmless-looking command can call harmful code.

Read on arXiv

PrimeSim: an open video model that simulates robots following commands

Travis Davies, Wentao Ma, Lihao Guo, et al. · Submitted 3 Nov 2026 · 7 upvotes yesterday

Preprint · not peer reviewedCode releasedTrending

Starting from a video generator, the team trained on 30,110 hours of demonstrations from 22 robot platforms so it follows commanded actions and keeps several camera views consistent.

Key finding: It runs at 10.1 frames per second, stays stable over long runs and leads four of five test suites against open rivals. Weights and code are released.

Why it matters: A trustworthy simulator lets robot software be trained and tested without risking real hardware.

Read on arXiv

Your RightSignal

Topics in News

RoboColosseum: testing robot AI in thousands of blind trials at 124 sites

SpecPilot: AI agents tune a faster way to run language models

Near-zero word vectors silently stall training in some newer model designs

Caper: try a coding agent's command in a sandbox before allowing it

PrimeSim: an open video model that simulates robots following commands

That’s every paper.