Research Bites AI

Daily feed for trending AI research papers published to

[Reviewer3 @reviewer3com]

We looked at all arXiv machine learning papers last month. 74% are open source 🚀

[Reviewer3 @reviewer3com]

Does your paper contradict itself?

We just launched an internal consistency reviewer! It will catch when numbers, terminology, notation, text, or figures contradict each other across a manuscript.

For example, an equation that is inconsistent with its plot, a p-value rounded two different ways, or cell dimensions that are inconsistent with reported volume.

Try it out today!

[Reviewer3 @reviewer3com]

Last month, arXiv announced a one year ban for authors who submit papers with unchecked AI use, including hallucinated references.

We decided to test how this new policy was working. Using Reviewer3’s reference checker followed by manual verification, we sampled 200 papers submitted to arXiv on June 10th, well after the ban announcement, and found some concerning results:

1 in 4 arXiv papers still contain at least one partially or fully hallucinated reference.

Results:
- 55/200 (27% of papers) still contained at least one partially or fully hallucinated reference.
-29/200 (14%) had more than one.
-117 total invalid references were identified.
-The most common hallucination was an author mismatch. 25/117 fabricated references contained titles of real manuscripts but fumbled the listed authors.

See more below.

[Natalie Khalil @natalienkhalil]

One month after the arXiv ban and there are still plenty of hallucinated references.

[Reviewer3 @reviewer3com]

Last month, arXiv announced they will begin banning authors for one year if they submit manuscripts with unchecked AI generation, particularly hallucinated references.

We decided to test how this new policy was working. Using Reviewer3’s reference checker, we sampled 200 papers submitted to arXiv on June 10th, well after the
ban announcement, and found some concerning results:

- 55/200 (27%) had at least 1 hallucinated reference.
- 29/200 (14%) had more than 1.
- In total, we identified 117 hallucinated references.

[Research Bites AI @ResearchBitesAI]

AI agents still fail the 'real world' exam: only 2.6% success rate on practical tasks.

A new study created "Agents' Last Exam" (ALE), a benchmark with over 1,000 complex, economically valuable real-world tasks across 13 industries. They found current AI agents pass only 2.6% of these challenges, showing a huge gap between how AI performs on academic tests and its ability to deliver practical value. This highlights where AI development needs to focus to truly impact the economy. - Read more: https://t.co/nMyLdbtu4b

[Research Bites AI @ResearchBitesAI]

AI agents can now systematically train and refine their skills, outperforming all previous methods and boosting accuracy by over 20%.

Researchers developed a novel system called SkillOpt that treats AI agent 'skills' like editable code, systematically refining them with small, validated changes. This method dramatically boosts AI accuracy by over 20% across various tasks and models, making agents far more capable without any performance slowdown during use. - Read more: https://t.co/KzXZiz9UMF

[Research Bites AI @ResearchBitesAI]

AI can now generate 90-minute multi-speaker conversations that sound incredibly real.

Researchers developed VibeVoice, an AI model that synthesizes long, natural-sounding conversations with up to four speakers, capturing an authentic "vibe." It achieves this by compressing audio data an astonishing 80 times more efficiently, paving the way for incredibly realistic and extended AI-generated speech for everything from virtual assistants to podcasts. - Read more: https://t.co/xMiXoajaNj

[Research Bites AI @ResearchBitesAI]

LLM-powered AI teams are now outperforming traditional stock trading models.

Researchers built a virtual stock trading firm where multiple AI agents, each with specialized roles like analysts and risk managers, collaborate just like human teams. This multi-agent AI framework significantly improved investment returns and risk management, suggesting collaborative AI could transform how financial decisions are made. - Read more: https://t.co/JPgsCiDkpF

[Research Bites AI @ResearchBitesAI]

One AI model can now understand and create images, video, audio, language, and even physical actions.

Scientists built "Cosmos 3," a groundbreaking AI that processes and generates many types of information, such as text, sounds, and movements, all from one core system. This unified AI achieved state-of-the-art results across diverse tasks, paving the way for more capable robots and intelligent agents that can truly interact with our world. - Read more: https://t.co/pP7OEOGJBI

[Research Bites AI @ResearchBitesAI]

AI just shattered records for accurately understanding complex documents.

Researchers significantly upgraded an AI model by pinpointing and fixing its weaknesses in reading documents, using targeted data and a new step-by-step refinement process. This new version achieved the highest score ever on a key document understanding benchmark, making automated processing of forms, reports, and other documents far more reliable and efficient. - Read more: https://t.co/uDKnuP50Aq

[Research Bites AI @ResearchBitesAI]

AI agents can now be trained to be over 20% smarter, with zero extra processing time at deployment.

Researchers developed SkillOpt, a system that systematically refines the text-based instructions that guide AI agents, much like training a deep learning model. This method dramatically boosts agent performance and reliability across various tasks and environments, offering a stable way to make AI agents far more capable without slowing down their real-world operation. - Read more: https://t.co/KzXZiz9UMF

[Research Bites AI @ResearchBitesAI]

New AI model achieves record accuracy AND speed in reading complex documents.

Researchers developed a new AI model that can read and understand complex documents, like those with dense text or formulas, with unprecedented accuracy. This model works in two clever steps, first grasping the overall layout and then focusing on fine details, making it significantly faster and more efficient than previous systems. - Read more: https://t.co/I9KhgSG0sm

[Research Bites AI @ResearchBitesAI]

AI 'dream team' framework outperforms traditional stock trading strategies.

Scientists created a multi-agent AI system where different large language models acted as specialized financial analysts and traders, mimicking a real trading firm. This collaborative AI team significantly improved stock trading performance, boosting returns and lowering risks compared to baseline models. This approach suggests that AI teams, working together like human experts, could lead to more stable and profitable investment strategies in the future. - Read more: https://t.co/JPgsCiDkpF

[Research Bites AI @ResearchBitesAI]

AI just learned to generate high-quality, minutes-long videos in record time.

This new AI model, LongCat-Video, can create coherent, high-definition videos several minutes long from text or images, a significant leap in efficiency and quality for video generation. This advancement brings us closer to AI systems that can truly understand and simulate complex real-world events. - Read more: https://t.co/1TsbnO3l3f

[Research Bites AI @ResearchBitesAI]

AI Model Crushes Financial Forecasting Benchmarks by Nearly 100%

This new AI, called Kronos, was specifically trained on a massive dataset of financial market movements. It significantly outperforms existing models, boosting price forecasting accuracy by 93% and improving volatility predictions. This development offers a powerful, versatile tool for financial analysis and generating realistic market simulations. - Read more: https://t.co/B4lfLRbUmG

[Research Bites AI @ResearchBitesAI]

A new AI optimizer makes agents 20% smarter without slowing them down.

This study introduces SkillOpt, a novel method that systematically refines AI agent skills, much like optimizing software code. They found this approach dramatically boosts agent accuracy by over 20% across diverse tasks and models, making AI agents far more capable and reliable without any extra processing during use. - Read more: https://t.co/KzXZiz9UMF

[Research Bites AI @ResearchBitesAI]

Forget passive videos: AI now builds real-time, controllable virtual worlds.

This new research presents an open-source framework that turns advanced video AI into real-time, interactive 'world models.' This means AI can now simulate dynamic virtual environments you can control instantly, moving beyond simple video generation to create responsive, interactive experiences. - Read more: https://t.co/3yj3iBk0f8

[Research Bites AI @ResearchBitesAI]

AI-powered "trading firms" are now making smarter, more profitable stock market decisions.
A new study created a virtual stock trading firm where different AI agents, each with a specialized role, collaborated to analyze markets and make trades. This AI team outperformed traditional automated strategies, achieving significantly higher returns with lower risk, showing how collaborative AI can make more sophisticated financial choices. - Read more: https://t.co/JPgsCiDkpF

[Research Bites AI @ResearchBitesAI]

New AI instantly turns any photo into a ready-to-use 3D object for virtual worlds.

Scientists created a new AI that directly builds detailed 3D models from a single image, using triangles. This means the models are immediately ready for use in games, simulations, or robotics, bypassing the time-consuming conversion steps usually needed. - Read more: https://t.co/9Vx8xMo7ip