RightSignal · Tuesday, November 10

An AI research assistant that tracks what it knows to decide what to test next

Weida Wang, Benteng Chen, Yigeng Jiang, et al. · Submitted 2026-11-09 · 32 upvotes yesterday

Preprint · not peer reviewedTrending

Sage-Taiyi keeps track of its evidence, assumptions and open questions, and uses that to choose research targets and reasoning methods. It was tested on a benchmark of 265 tasks from 50 papers across ten fields.

Key finding: It scored 54.0 overall against 47.9 for the strongest rival, AI Scientist-v2. In chemistry work with experts, its suggestions raised one reaction's yield from trace amounts to 84%.

Why it matters: Its ideas were checked in real experiments, not only on a benchmark.

Read on arXiv

Cleaned-up research code helps AI agents rerun machine learning experiments

Yubo Wang, Rui Meng, Jaehyun Nam, Tomas Pfister, Jinsung Yoon · Submitted 2026-11-05 · 22 upvotes yesterday

Preprint · not peer reviewedTrending

The authors built an agent workflow that turns a paper's released code into a tidy package with setup steps, launch scripts and a guide. They tested it on 1,029 tasks from 36 papers.

Key finding: Reproduction success rose from 76.4% to 92.6%, and API spending fell by 19.7% on average across four models.

Why it matters: Writing down how an experiment was actually run makes results easier to check.

Read on arXiv

Systems

A serving engine answers branching AI agent questions up to 9.5 times faster

Bobo Li, Jiajia Song, Hao Fei, Wynne Hsu, Mong Li Lee · Submitted 2026-10-30 · 22 upvotes yesterday

Preprint · not peer reviewedTrending

Agents often ask a model several related yes-or-no style questions that share the same start. Flide packs these branches into one stream so the shared part is computed once, while each branch gives exactly the same answer as if run alone.

Key finding: It answered queries up to 9.5 times faster than the popular engines vLLM and SGLang.

Why it matters: Faster, exact answers matter for agents that make many quick decisions.

Read on arXiv

Systems

Training small AI models from a teacher tolerates cheaper, less exact settings

Cong Yu, Ziting Wang, Songlin Jiang, et al. · Submitted 2026-11-08 · 7 upvotes yesterday

Preprint · not peer reviewedCode releasedTrending

On-policy distillation trains a student model on its own answers, guided by a teacher. Systems cut costs with lower-precision maths and out-of-step updates, which cause mismatch. The authors measured how much mismatch training can handle.

Key finding: It tolerates a lot before quality drops. Combining 8-bit precision with asynchronous updates reached the same quality 1.62 times sooner.

Why it matters: It gives teams a simple check for which cost-saving settings are safe to use.

Read on arXiv

Systems

An AI agent tunes large-model training setups and speeds them up to 2.75 times

InfraRSI Team, Yuchuan Tian, Haocheng Mei, et al. · Submitted 2026-11-08 · 2 upvotes yesterday

Preprint · not peer reviewedTrending

InfraRSI takes plain instructions and tries settings and code changes for training large AI models, checking each change against fixed rules to avoid cheating. It learns from failed runs, including out-of-memory crashes.

Key finding: Across four workloads it reached speedups of up to 2.75 times and raised average hardware use from 17.4% to 30.0%.

Why it matters: Better use of expensive GPUs lowers the cost of training models.

Read on arXiv

Your RightSignal

Topics in News

An AI research assistant that tracks what it knows to decide what to test next

Cleaned-up research code helps AI agents rerun machine learning experiments

A serving engine answers branching AI agent questions up to 9.5 times faster

Training small AI models from a teacher tolerates cheaper, less exact settings

An AI agent tunes large-model training setups and speeds them up to 2.75 times

That’s every paper.