RightSignal · Saturday, October 17

FDE-Gym: AI agents struggle to deliver real customer projects

Dongxu Yang, Hongda Yan, Renxin Wang, et al. · Submitted 15 Oct 2026 · 282 upvotes yesterday

Preprint · not peer reviewedTrending

The authors built 95 tasks adapted from real customer projects across 18 sectors. Agents must talk to simulated stakeholders, find unstated requirements and change working systems. Scores reflect how much they improve the customer's own success measure.

Key finding: The best of 13 model setups fully solved only 14 of 95 tasks. Finding requirements nobody wrote down was the main obstacle.

Why it matters: Most agent tests give clear instructions. Real work rarely does, and this shows how big that gap still is.

Read on arXiv

Waypoint-OPD: teaching smaller models only the lessons a teacher learned consistently

Bingxi Hou, Guochao Jiang, Guofeng Quan, et al. · Submitted 15 Oct 2026 · 665 upvotes yesterday

Preprint · not peer reviewedTrending

When a teacher model improves through training, a student can copy what changed. This method checks a midway snapshot of the teacher and only passes on changes that pointed the same way early and late in training.

Key finding: It scored best among the compared methods in all six maths and code settings tested, for one extra scoring pass.

Why it matters: It is a cheap way to make smaller, faster models better at reasoning.

Read on arXiv

OpenGCUA: an open toolkit for AI agents that operate a computer

Hanzhang Zhou, Long Li, Quyu Kong, et al. · Submitted 15 Oct 2026 · 315 upvotes yesterday

Preprint · not peer reviewedCode releasedTrending

The team released an open setup for agents that use a computer both by clicking on screen and by writing code. It also hands some decisions from fixed rules to the model itself.

Key finding: Across eight models it raised average partial scores by 19.6 points and cut steps per task by about half compared with clicking alone.

Why it matters: The best computer-use agents are closed. An open version lets others study and improve them.

Read on arXiv

AI safety

LEASH: a small AI agent calls in a stronger one only when it's unsure

Qingni Wang, Changdae Oh, Yuzhe Yang, et al. · Submitted 15 Oct 2026 · 6 upvotes yesterday

Preprint · not peer reviewedTrending

A cheap model does the work and watches its own internal signals for signs it resembles past failures. When it does, a stronger model takes over, and an observer keeps watching that model too.

Key finding: With one small model it matched an always-on top model's success rate on a computer-use benchmark while cutting cost by 19.2%.

Why it matters: It offers a cheaper way to run agents while still keeping an eye on them.

Read on arXiv

Teams of AI agents learn to organise themselves by bidding for turns

Guowei Xu, Zhenting Qi, Amani Kiruga, Himabindu Lakkaraju, Yilun Du · Submitted 15 Oct 2026 · 10 upvotes yesterday

Preprint · not peer reviewedCode releasedTrending

Instead of a boss agent assigning work, each agent bids to take part, and the bids decide who acts and who gets credit. Agents learn both their actions and their bidding.

Key finding: Learned self-organisation beat other ways of coordinating, and larger teams coped better when agents dropped out.

Why it matters: It suggests groups of agents can coordinate without a single point of control.

Read on arXiv

Your RightSignal

Topics in News

FDE-Gym: AI agents struggle to deliver real customer projects

Waypoint-OPD: teaching smaller models only the lessons a teacher learned consistently

OpenGCUA: an open toolkit for AI agents that operate a computer

LEASH: a small AI agent calls in a stronger one only when it's unsure

Teams of AI agents learn to organise themselves by bidding for turns

That’s every paper.