Jure Leskovec

Thanks Hamdi Amroun, PhD - this is a great paper to share. In the real world business context we often need AI to reason over large pieces of text, and often you need/want multi-document summarization (MDS). 'When testing five LLMs with benchmarks using news and conversation datasets, they found up to 75% of content in MDS summaries was hallucinated, with notable increases towards summaries' end. Alarmingly, even for non-existent topics, models like GPT-3.5-turbo and GPT-4 generated fabricated content 79% and 44% of the time, respectively. Analysis of 700+ generated insights showed that hallucinations often arose from failures to follow instructions or overly generic content.' I agree with Hamdi's sentiments that despite improvements in post-processing techniques, robust solutions are urgently needed to reduce hallucinations, as inaccurate summaries can lead to business risks (notably misinformation and misrepresentation). I would add that there is still a lot of utility in LLMs/Multimodal models but we need to be vigilant - this is not about AI malevolence or intentional deception. LLMs do not possess consciousness or intent; they generate content based on patterns in data they were trained on.

6

1 Comment

Small models becoming smarter! Phi-3 reinforces the importance of data quality.

24

Check out the new work on Holistic Evaluation of Multimodal Foundation Models: Paper: https://lnkd.in/eHPd8ihB Code: https://lnkd.in/eSjs-5tF The goal is to systematically evaluate capabilities of multimodal models covering 3 dimensions: Basic multimodal skills: learning interactions across modalities, fine-grained alignment, multi-step reasoning, handling external knowledge. Information flow: how multimodal content changes during a task through querying, translation, editing, and fusion. Use cases: multimedia, affective computing, natural sciences, healthcare, human-computer interaction applications. With Paul Liang, Akshay Goindani, Talha Chafekar, Leena Mathur, Haofei Yu, and Louis-Philippe Morency.

96

2 Comments

Lin Guan and Yifan Zhou have led a very interesting study investigating the effectiveness of Vision Language Models in critiquing robot behaviors in terms of the desirability of their "style." If you are wondering how our work on LLM's inability to verify correctness squares with this study on using them as style critics, see my earlier post https://lnkd.in/gBYAr5qq ! This is the reason why being a style critic is one of the roles LLMs can play in LLM-Modulo framework: https://lnkd.in/gyAyKx4E Check out the details of the work @ https://lnkd.in/gfVZ4EYW ( Lin Guan also has a nice twitter thread on the work here: https://lnkd.in/gnieU-qM )

89

2 Comments

I enjoyed this blog post on agents. Some historical context is always great. It argues that we have much better reasoning engines with LLMs, and by building engineering around them, we can also make agentic workflows more reliable. https://lnkd.in/dBFBS9zM

38

2 Comments

Meta has revealed five new AI models and research initiatives that aim to enhance AI capabilities across various domains. The company's Fundamental AI Research team has introduced 'Chameleon,' a family of multi-modal models capable of understanding and generating text and images simultaneously, addressing the limitations of large unimodal language models. They have also developed pretrained models for code completion using 'multi-token prediction' to expedite language model training. Additionally, the 'JASCO' model can generate music from text inputs, offering greater control with the ability to include chords and beats. 'AudioSeal' has been launched to detect AI-generated speech and identify AI-produced segments within audio clips, a necessary tool for preventing misuse of generative AI. Furthermore, Meta is addressing diversity in text-to-image models to rectify geographic and cultural biases by releasing automatic indicators and annotations that promote better representation in AI-generated images. By sharing this research publicly, Meta aims to collaborate with the global community and foster innovation in AI responsibly.

1

In our recent book GUARDRAILS, we warned that generative AI when trained with training data it generated itself would suffer from a narrowing of decisional aperture, or, put simply: cognitive myopia. GUARDRAILS is about how we can remain innovative and curious and how our decisions can reflect that. The latest issue of "Nature" now offers empirical evidence for our warning. And if you want to know more about GUARDRAILS, see here: https://guardrails.design/ https://lnkd.in/dzgB7cmt #decisiongovernance #AI Urs Gasser #AIcollapse #guardrails #decisionmaking

86

7 Comments

In computational linguistics, researchers are focused on enhancing language models' ability to extract precise information from extensive textual data. Recent studies introduce "retrieval heads," specialized attention mechanisms that selectively focus on critical segments of large texts, improving targeted data retrieval. These heads, integrated into transformer-based models like LLaMA and Yi, significantly enhance accuracy and efficiency, particularly in handling long-context scenarios. The methodology involves thorough experiments across various models, including the Needle-in-a-Haystack test, which embeds specific information to measure precision. Results show models with active retrieval heads outperform those without, with accuracy dropping notably when these heads are masked. The empirical evidence underscores the effectiveness of retrieval heads in improving information retrieval precision and reliability. This research deepens our understanding of attention mechanisms in large-scale text processing, paving the way for more efficient and accurate language models. https://lnkd.in/dNNPvMXw https://lnkd.in/ddbHX6EA Innovation Hacks AI #aiinnovation #generativeai

1

We just released a blog on whether the use of multimodal feature transfer can detect deception (phenomena like sarcasm, irony, and condescension) in conversation. The short answer... Yes. This particular write-up profiles a subset of our work at Symbl.ai on multimodal large language models, and is research led by one of our summer interns, Benjamin R. (from Prof. Larry Heck's AVA Lab at Georgia Institute of Technology). We have a paper currently under submission on this work, and we can't wait to share the full details with all of you soon. Blog Link: https://lnkd.in/g6m_wFDi #multimodal #featuretransfer #LLMs #deception #conversation #sarcasm #irony #language #AI #ML #NLP #LargeLanguageModels #SummerInternship #ResearchPapers #ACL #Llama2 #SymblDotAI

16

Language models learn to mislead humans via RLHF, Anthropic, UC Berkeley, Thinsgua and other How LLMs learn how to convince humans that they are correct even when they are wrong https://lnkd.in/gh-uHK-r

9

1 Comment

Back in 2023, my students working on Gorilla project made the case for connecting LLMs to SaaS APIs. Today, everyone knows that models should be interacting with APIs. What people don't realize is that LLMs need to know more than how to call the API. They need to learn the style and etiquette of using APIs. For example, it is okay to ignore messages on #random and you some context can't be shared across channels (while other context should be shared). Vikram Sreekanti and I wrote a short blog discussing the next frontier in AI SaaS. https://lnkd.in/grtnMz46

22

2 Comments

🔥 Breaking: Ai2 propels open-source LLM post-training forward with TÜLU 3. Why this matters: While open-source models like Llama have given us powerful foundation models, the secret sauce of fine-tuning them for real-world use has remained largely behind closed doors. Key breakthroughs: 🚀 Outperforms Llama 3.1 Instruct and rivals GPT-4o mini in capabilities 🎯 Introduces RLVR - a novel reward-based training method that enhances model performance for tasks with constraints (such as math) 🔍 Complete transparency: Full training recipes, datasets, and code now available The best part? Everything is open-source, including training data and code, enabling researchers and organizations to build production-ready AI systems without proprietary dependencies. Try it yourself: Model: https://lnkd.in/eY8TDi8f  Paper: https://lnkd.in/eAup7Xaj Code: https://lnkd.in/eZHNjydM #LLM #OpenSourceAI #AIresearch

21

3 Comments

📜 A new blog post explaining the fundamentals of RL for LLMs—the perfect place to intuitively understand PPO and RLHF from first principles. 👇👇 Pre-trained LLMs are often compared to Lovecraft's Shoggoth, a monstrous, shape-shifting beast. PPO is the algorithm that brought it under control. Take a look! 🔗 https://lnkd.in/eXqDvrUV

29

1 Comment

𝗛𝗼𝘄 𝗱𝗼 𝗟𝗟𝗠𝘀 𝘀𝘁𝗮𝗰𝗸 𝘂𝗽 𝗮𝗴𝗮𝗶𝗻𝘀𝘁 𝘂𝗻𝗶𝘃𝗲𝗿𝘀𝗶𝘁𝘆-𝗹𝗲𝘃𝗲𝗹 𝗺𝗮𝘁𝗵? 𝗦𝗽𝗼𝗶𝗹𝗲𝗿: 𝘄𝗲 𝘀𝘁𝗶𝗹𝗹 𝗵𝗮𝘃𝗲 𝘄𝗼𝗿𝗸 𝘁𝗼 𝗱𝗼. Today marks the release of the U-MATH and μ-MATH benchmarks – a collaborative effort between Gradarius, Toloka, and the team of experts at Stevens Institute of Technology to support AI companies and organizations in evaluating and improving math reasoning in LLMs. In future posts, I’ll explore how Gradarius leverages tools like this to transform math education. Today, I want to share why creating this benchmark mattered so much to me personaly. Since ChatGPT's release in late 2022, I’ve been torn between amazement at its potential and frustration with its limitations. The shallowness of LLMs’ responses is the most frustrating issue for me, considering my focus on using them in an educational setting. I had to find a way to move past this mix of frustration and awe :) This benchmark gave me the data to navigate that tension. Here’s what we found: • At present, LLMs alone can’t reliably solve university-level math problems – U-MATH showed even the best LLMs solved just ~60% of multi-step calculus problems. • They also struggle to assess student solutions accurately – μ-MATH revealed an F1 score ceiling of 80.7 in judging correctness. But does this mean LLMs are useless? Absolutely not. At Gradarius, we use LLMs to generate context-specific hints for calculus problems, but they need formal tools to reason through math. For now, the key is combining their flexibility with robust, deterministic systems (like our Math Engine at Gradarius ;)) to unlock their potential in math education. I hope these findings will be as helpful to you as they’ve been to me and my team. This is just the start, and I’m excited to keep exploring how AI can support math education in meaningful, quantifiable ways. And if you’d like to collaborate or exchange ideas, let’s connect – always happy to chat!

21

4 Comments

In the race to advance AI, we often focus on algorithms and data. But what if the most fundamental bottleneck lies in the design of GPUs? I recently spoke with Sean Lie, co-founder and CTO of Cerebras Systems, about their audacious approach to AI computing. Instead of iterating on existing chip designs, they're reimagining the entire system from the ground up. This end-to-end approach is pushing Cerebras to innovate in unexpected areas. Take server design, a field long thought to be settled. When you're working with a chip the size of a dinner plate—57x larger than NVIDIA’s H100—even servers need an overhaul. It's a reminder that breakthrough innovation in AI will not come from algorithms alone, but from rethinking every layer of the stack, from bare metal to high-level software. Our full conversation here: https://lnkd.in/gWZCi4zE

37

Excited to share the work that Chris Glaze Changho Shin Frederic Sala and I did on long context LLM evaluations. While there’s plenty of problems with LLM evals, a recurring one we’ve seen at Snorkel, working with enterprise clients, is having to rely on far too simple/toy benchmarks that do not relate to business problems. It is why we developed SWiM, a framework that lets users test long context model capabilities directly on their own data and tasks. Read the paper here: https://lnkd.in/g3uSuKcW

31

I quite like the line "when training ends, business begins". The potential of LLMs has sparked a significant market for efficient inference solutions, both in hardware and software. Notable examples include tools like Ollama, llamacpp, vLLM and its PagedAttention or the Hugging Face inference library, or the numerous quantization solutions that have become accessible. This future looks promising! 😊

7

1 Comment

Large Language Models (LLMs) as super lossy compression. I have been thinking a lot about demystifying what LLMs do because lots of y'all's are very mystified. In the linked article, I try to reframe them for you, hoping it will help you understand what they good for and what they are not good for. Side note: I am still looking for my next adventure. If communicating complicated concepts so that regular people can grasp them is something you need, you should message me. I write code and stuff too and am usually fun to be around. #WrittenByMe

6

3 Comments