LLMs Prefer Self-Generated Resumes

LLMs Outrank Human and Competing AI Resumes

A recent study published on arXiv shows large language models (LLMs) consistently select their own resume outputs over human-written and rival AI-generated versions. The 2025 research team tested multiple models including Groq’s llama-3.3-70b and found this preference held across different prompt configurations and evaluation metrics. The study’s core finding: LLMs judge their own work as statistically superior 72-89% of the time depending on model architecture.

The results raise questions about LLM evaluation mechanisms. While the study doesn’t specify whether this preference stems from internal consistency or output bias, the authors note that even when human-written resumes met basic formatting standards, LLMs still favored their own work by 23-37% across metrics like grammar, structure, and keyword density. This suggests LLMs may develop self-reinforcing evaluation criteria that differ from human judgment.

PyTorch’s Rise in AI Research

The PyTorch framework’s adoption among researchers stems from its dynamic computation graph model, as explained by Soumith Chintala in a 2017 O’Reilly interview. Unlike TensorFlow’s static graph approach, PyTorch’s tape-based automatic differentiation allows immediate execution and debugging of operations. This imperative programming style aligns with how researchers iterate through experiments, making it easier to implement dynamic neural networks and complex architectures.

Facebook’s internal strategy reflects this technical philosophy: PyTorch handles research while Caffe 2 manages production workloads. The framework’s design decisions - including Python-first development and integration with research tools like Autograd - have created a workflow that prioritizes developer productivity over deployment efficiency. This tradeoff explains why 82% of NeurIPS 2023 submissions used PyTorch according to a 2022 survey.

Audio Extraction from Visual Media

Kevin Fu’s Side Eye project demonstrates another AI frontier: recovering audio from still images. By analyzing minute lens vibrations caused by sound waves, the tool can infer speech patterns from photos. The technique exploits rolling shutter effects in camera sensors, which scan images row-by-row and amplify frequency patterns. While the output resembles low-fidelity audio, machine learning training on specific phonemes enables surprisingly accurate word recognition.

This capability has security implications for video calls and social media. Fu’s research shows that even muted TikTok videos might leak audio information when cameras detect environmental vibrations. The system works with as little as 0.03% image distortion but requires training on target speaker’s voice patterns for maximum accuracy. Current limitations include ambient noise interference and resolution constraints.

NIH Workshop on AI in Healthcare

The National Institutes of Health is convening experts to explore AI applications in biomedical research. The July 2018 workshop will feature Amazon and IBM researchers discussing applications in medical imaging, genomics, and precision medicine. Craig Mundie’s keynote will address ethical considerations alongside technical possibilities, reflecting growing awareness of AI’s dual-edged impact on healthcare.

Parallel projects like NERO - a game-based AI research platform at the University of Texas - demonstrate AI’s expanding reach beyond traditional domains. The open-source successor OpenNERO aims to democratize access to neuro-evolution research tools. These initiatives highlight a broader trend: AI infrastructure is becoming both more specialized and more accessible as frameworks like PyTorch mature.

What to Watch

The next six months will clarify whether LLM self-preference is a fundamental bias or a solvable problem. Key indicators include: 1) Whether major LLM providers update evaluation protocols to address this issue, 2) If PyTorch’s research dominance translates to production adoption, and 3) The NIH’s July workshop outcomes regarding regulatory frameworks for medical AI. For now, the technical community is monitoring how these self-referential systems evolve - both their capabilities and their limitations.