RightSignal · Friday, October 23

A test of AI research agents on the live web finds they beat people, but only at 39%

Xiaomin Yu, Hongliang Chen, Zhe Huang, et al. · Submitted 2026-10-19 · 117 upvotes yesterday

Preprint · not peer reviewedTrending

The authors built OpenDS-World, 200 hard questions that an agent must answer by searching the real, open web, opening files and connecting evidence from several sources. Answers were checked by people to be unique and valid.

Key finding: The best of 12 frontier models answered 39.0% correctly; people across 1,000 attempts managed 21.1%.

Why it matters: It tests agents in the messy real web rather than a frozen copy, and shows they still struggle with coordinating tools over long searches.

Read on arXiv

Deep Fusion Memory adds learned knowledge to language models without live search

Rubin Wei, Siyuan Sun, Hao Sun, et al. · Submitted 2026-10-18 · 32 upvotes yesterday

Preprint · not peer reviewedTrending

Instead of searching documents at answer time, a small add-on predicts a 'memory' from the text so far and blends it into the model's inner layers. The main model stays frozen.

Key finding: It needs 32 times fewer search queries to build its training data and beats the strongest baseline by 12.71% on Mistral-7B and 11.19% on Qwen3-8B across six question-answering tasks.

Why it matters: It points to a cheaper way to teach a model new facts without retraining it or slowing every answer with retrieval.

Read on arXiv

Sharing experts across layers lets mixture-of-experts models do more with fewer parts

Ermo Hua, Xiangyu Hong, Baiting Wu, et al. · Submitted 2026-10-20 · 23 upvotes yesterday

Preprint · not peer reviewedTrending

Mixture-of-experts models give each layer its own specialist sub-networks, which the authors find are often similar. They test 'Meta Experts' shared across all layers instead.

Key finding: With the same number of active experts, the shared design matches a standard model using only 70 to 80% of the experts.

Why it matters: Fewer stored experts means smaller, cheaper models for the same quality.

Read on arXiv

Voice assistants can start thinking and calling tools while you are still talking

Zhenyu Zhang, Kewei Fu, Yongqi Tong, et al. · Submitted 2026-10-19 · 16 upvotes yesterday

Preprint · not peer reviewedTrending

Heartbeat Thinking repeatedly reads the partial transcript while the user speaks, lets earlier reasoning and tool calls finish instead of cutting them off, and shares their results.

Key finding: On one benchmark it covered 16 percentage points more tool calls in advance and cut average response time by 36.4%.

Why it matters: Waiting after you finish speaking is one of the most noticeable flaws in voice agents.

Read on arXiv

A new benchmark tests AI agents that write research reports about a video

Jiahao Meng, Qize Yang, Yuxuan Wang, et al. · Submitted 2026-10-22 · 12 upvotes yesterday

Preprint · not peer reviewedCode releasedTrending

Given a video and a request, an agent must find research questions, gather outside evidence and write a report linked to moments in the video. The benchmark has 50 Chinese and English tasks and 6,414 grading criteria.

Key finding: Across 18 agent setups, the best scored 63.3 out of 100.

Why it matters: It shows agents still miss content, reason weakly across sound and images, and cite sources poorly. Code and data are released.

Read on arXiv

Your RightSignal

Topics in News

A test of AI research agents on the live web finds they beat people, but only at 39%

Deep Fusion Memory adds learned knowledge to language models without live search

Sharing experts across layers lets mixture-of-experts models do more with fewer parts

Voice assistants can start thinking and calling tools while you are still talking

A new benchmark tests AI agents that write research reports about a video

That’s every paper.