RightSignal · Thursday, November 12

AI agents can read only the parts of web pages they need and use far fewer tokens

Zedong Liu, Jiaan Wu, Xinyang Ma, Le Xu, Kai Wang, Yuanchao Hu · Submitted 2026-11-07 · 10 upvotes yesterday

Preprint · not peer reviewedTrending

AI agents often read whole web pages, documents or tool outputs when they need only a little. The authors built SparseRead, a layer between the agent and its sources that tracks what the agent needs and limits how much raw text it sees, with a fallback to the full view.

Key finding: Across six models and five scenarios, it cut total tokens by up to 92.9% and time by up to 89.0%, keeping or improving quality in 97% of settings.

Why it matters: Reading less makes agents cheaper and faster without changing the model itself.

Read on arXiv

A teacher model helps compressed AI agents fix the mistakes they make

Genming Cui, Xiangyu Li, Zixu Hao, Xin Ding, Yunxin Liu, Ting Cao · Submitted 2026-11-10 · 23 upvotes yesterday

Preprint · not peer reviewedTrending

Storing an agent model's weights in fewer bits makes it cheaper but makes it miss task details. In SteerDistill, a full-precision teacher guides the compressed model to revise failed answers, and the model is then trained to make those corrections on its own.

Key finding: Across five agent benchmarks and three models, it beat the best compressed baseline by 1.67 to 2.18 points and stayed within one point of the uncompressed model.

Why it matters: It helps cheap, compressed agents keep most of the reliability of full-size ones.

Read on arXiv

Reinforcement learning teaches a model to study a textbook before solving proofs

Lingkai Kong, Yuzhe Gu, Haiteng Zhao, Zijian Wu, Youqing Fang, Zhouqi Hua · Submitted 2026-10-29 · 22 upvotes yesterday

Preprint · not peer reviewedTrending

The authors gave a model a tool to navigate maths textbooks, learn the relevant ideas and then write proofs for exercises, with solutions hidden. They rewarded good study behaviour with reinforcement learning.

Key finding: Proof accuracy of Intern-S2-Preview-35B rose from 51.5% to 79.5%, beating versions trained without textbooks or with fixed retrieval.

Why it matters: It suggests models can learn to learn from long documents on the fly, not just look up facts.

Read on arXiv

GTAGym tests AI agents on real missions in the video game GTA V

Jiajun Chen, Shicheng Fang, Jiahao Lu, Yuxin Wang, Xipeng Qiu · Submitted 2026-11-09 · 11 upvotes yesterday

Preprint · not peer reviewedTrending

The authors built an environment where agents play GTA V tasks from screenshots and keyboard-mouse actions, with results checked automatically from the game's state. It can also generate new tasks for training.

Key finding: The best setup reached 70.9% accuracy on one benchmark branch, but flying, evading and intercepting stayed hard.

Why it matters: Automatic checks make a complex game usable for both testing and training agents.

Read on arXiv

Systems

A serving system that adapts to AI agents cuts waiting time on big GPU clusters

Hao Kang, Weili Xu, Ziyang Li, Zheng Du, Hanchen Li, Tian Xia · Submitted 2026-11-09 · 2 upvotes yesterday

Preprint · not peer reviewedTrending

Serving systems that split prompt reading and answer writing onto separate machines struggle with agent workloads. Evolving Infra lets machines switch roles at runtime and keeps the memory caches agents will reuse.

Key finding: On up to 256 GB300 GPUs, it cut mean request latency by up to 4.1x and time to first token by up to 14x compared with Nvidia's Dynamo.

Why it matters: Agents are a growing share of AI demand, and they reuse context in ways current systems handle poorly.

Read on arXiv

Your RightSignal

Topics in News

AI agents can read only the parts of web pages they need and use far fewer tokens

A teacher model helps compressed AI agents fix the mistakes they make

Reinforcement learning teaches a model to study a textbook before solving proofs

GTAGym tests AI agents on real missions in the video game GTA V

A serving system that adapts to AI agents cuts waiting time on big GPU clusters

That’s every paper.