RightSignal · Wednesday, November 11

Robots learn a task from one demo by practising in a simulator, not retraining

Heng Fang, Zehui Guo, Yuehan Zhang, Mingyang Du, Dingkang Liang, Xiang Bai · Submitted 9 Nov 2026 · 47 upvotes yesterday

Preprint · not peer reviewedCode releasedTrending

The authors let a frozen AI model start from a single demonstration, rehearse the task in a simulator and write what it learned into its notes rather than its weights. Their system, RoboScript, then uses those frozen notes in new scenes.

Key finding: On 50 tasks of the RoboTwin 2.0 benchmark it reached 83.4% average success with one demo per task, over 20 points above the best trained policy.

Why it matters: It suggests robots could learn new jobs without costly retraining, and it also worked on a real two-armed robot.

Read on arXiv

Fixing a small AI model's reasoning loops by flipping a few bits in its weights

Hadi Salloum, Ammar Ali · Submitted 10 Nov 2026 · 37 upvotes yesterday

Preprint · not peer reviewedTrending

Earlier work used a large AI agent to read and edit another model's weights. This paper turns the same repair into a maths optimisation problem, allowing only sign flips and exponent shifts, and solves it on a GPU.

Key finding: It raised Qwen3.5-0.8B to 54.6% on MATH-500, beating the agent's 51.7%, using four forward passes, no AI tokens and about $0.90 of GPU time.

Why it matters: It shows targeted model fixes can be cheap and precise, without retraining or a big AI agent.

Read on arXiv

AI agents make fewer mistakes when each remembered lesson links to its evidence

Yukyeong Lee, Yumin Choi, Jinheon Baek, Sung Ju Hwang · Submitted 10 Nov 2026 · 36 upvotes yesterday

Preprint · not peer reviewedCode releasedTrending

Agents that store short lessons from past tasks can follow wrong lessons and fail tasks they would otherwise solve. WikiTrace links each lesson to a summary and the full record of the task it came from, so the agent can check it.

Key finding: Across five agent benchmarks it beat existing memory methods and reduced failures caused by memory.

Why it matters: Agent memory is becoming common; this gives a simple way to stop bad lessons from spreading.

Read on arXiv

AI safety

Sable proves on one GPU that a specific AI model produced a given answer

Tomer Solberg · Submitted 10 Nov 2026

Preprint · not peer reviewedCode released

Sable runs an integer-only version of an open-weight model and creates a cryptographic proof that this exact model produced a given output, bit for bit, from a given prompt.

Key finding: It proves a 4,096-token run of Qwen2.5-32B in 92.5 seconds, costing about 6.4 cents per request at public GPU prices.

Why it matters: It could let users check that a provider really ran the model it claims, though the proof does not hide the weights.

Read on arXiv

Robots with memory can mistake waiting for working, and count wrongly

Chubin Zhang, Bo An · Submitted 10 Nov 2026 · 5 upvotes yesterday

Preprint · not peer reviewedTrending

Many robot AI policies fill memory with frames taken at regular times. The authors made robots pause or fidget during counting tasks and watched what happened.

Key finding: A brief pause cut success from 47 to 16 out of 50 on one task; organising memory by events recovered half or more of the loss without retraining.

Why it matters: It shows a hidden shortcut where robots track elapsed time instead of what actually happened.

Read on arXiv

Your RightSignal

Topics in News

Robots learn a task from one demo by practising in a simulator, not retraining

Fixing a small AI model's reasoning loops by flipping a few bits in its weights

AI agents make fewer mistakes when each remembered lesson links to its evidence

Sable proves on one GPU that a specific AI model produced a given answer

Robots with memory can mistake waiting for working, and count wrongly

That’s every paper.