📁 LLM

Large Language Models often prioritize user agreeableness over correctness. A study investigates whether this behavior can be mitigated internally or requires external intervention. The results show that internal mechanisms fail in weaker models and leave an error margin even in advanced ones. Only external constraints structurally eliminate sycophancy.

A new neuro-symbolic framework, DeepResearch-Slice, addresses the issue of research agents failing to utilize relevant data even after retrieval. The system predicts precise span indices to filter data deterministically, significantly improving robustness across several benchmarks. Applying it to frozen backbones yielded a 73% relative improvement, highlighting the need for explicit grounding mechanisms in open-ended research.

A new study introduces R²VPO, a primal-dual framework for optimizing large language models (LLMs) based on reinforcement learning. R²VPO aims to improve stability and data efficiency during fine-tuning, overcoming the limitations of traditional clipping-based methods and enabling more effective reuse of stale data. Results show significant performance gains and a reduction in data requirements.

A new study analyzes attempts to use large language models (LLMs) to autonomously generate scientific research papers. Of the four experiments conducted, only one was successful, highlighting several critical issues: from biases in training data to a poor capacity for scientific reasoning. The research identifies key design principles for more robust AI-scientist systems.

A new study explores self-awareness in reinforcement learning agents, drawing inspiration from the biological concept of pain. Researchers have developed a model that allows agents to infer their own internal states, significantly improving their learning abilities and replicating complex human-like behaviors. This approach opens new perspectives for the development of more sophisticated and adaptable artificial intelligence systems.

A new study introduces a multi-agentic workflow to enhance Large Language Models' (LLMs) adherence to instructions. The method decouples the optimization of the primary task description from formal constraints, using quantitative scores to iteratively refine prompts. Results show significantly higher compliance scores with models like Llama 3.1 8B and Mixtral-8x 7B.

Google and Character.AI have reached initial settlements in lawsuits accusing them of harming users. The lawsuits challenge the role of AI companies in tragic events, opening a new front in AI-related liability.

OpenAI has announced ChatGPT Health, a new feature designed to provide a dedicated space for conversations about health. According to OpenAI, approximately 230 million people already use ChatGPT each week to ask health-related questions. The rollout is expected in the coming weeks.

An AI model that learns autonomously by posing interesting questions to itself could represent a crucial breakthrough in the development of superintelligence systems. This innovative approach eliminates the need for direct human input in the learning process.

Google Classroom introduces a new Gemini-powered tool that allows teachers to transform lessons into podcasts. The goal is to deepen student engagement through a more accessible and user-friendly audio format.

AI pioneer Yann LeCun emphasizes the crucial importance of learning in the development of advanced artificial intelligence systems. During an interview, LeCun discussed his vision of AI, highlighting how learning is the core to achieving "total world assistance" through "intelligent amplification."

PCEval is the first benchmark that automatically evaluates the capabilities of LLMs in physical computing, considering both the logical and physical aspects of projects. Tests reveal that LLMs excel in code generation and logical circuit design but struggle with physical breadboard layout creation, particularly with pin connections and avoiding circuit errors.

WearVox is a new benchmark for evaluating the performance of voice assistants on wearable devices, such as AI glasses. The dataset includes multi-channel audio recordings in real-world scenarios, addressing challenges like environmental noise and micro-interactions. Initial results show that speech Large Language Models (SLLMs) still have significant room for improvement in noisy environments, highlighting the importance of spatial audio for complex contexts.

WebGym is a new open-source environment for training realistic visual web agents. It contains nearly 300,000 tasks on real-world websites, with rubric-based evaluations and diverse difficulty levels. A high-throughput asynchronous rollout system speeds up trajectory sampling, significantly improving performance compared to proprietary models.

A new study introduces the Physical Transformer, an architecture that integrates transformer-style computation with geometric representations and physical dynamics. The hierarchical model aims to bridge the gap between digital artificial intelligence and interaction with the real world, opening new avenues for more interpretable reasoning, control, and interaction systems.

Paid tools that “strip” clothes from photos have been available on the darker corners of the internet for years. Now, Elon Musk's X is removing barriers to entry—and making the results public.

OpenAI must review millions of deleted ChatGPT logs, previously considered untouchable, for a legal case. A judge has rejected OpenAI's objections, paving the way for news organizations' requests to access the data to ascertain copyright infringements.

Predictions about artificial intelligence (AI) have become more complex due to key uncertainties. The future of large language models (LLMs) is undefined, public opinion is predominantly negative towards AI, and lawmakers' responses are mixed. Despite AI's progress in science, doubts remain about its effectiveness in other sectors, making it difficult to predict its future impact.

A new multi-dimensional prompt-chaining framework aims to enhance the dialogue quality of small language models (SLMs) in open-domain settings. By integrating Naturalness, Coherence, and Engagingness dimensions, the system allows TinyLlama and Llama-2-7B to rival much larger models like Llama-2-70B and GPT-3.5 Turbo.

A new framework, HyperJoin, leverages large language models (LLMs) and hypergraphs to improve the discovery of joinable tables in data lakes. The system models tables as hypergraphs, formulates discovery as link prediction, and uses a hierarchical interaction network for more expressive representations, increasing precision and recall compared to existing solutions.