High symbolic prediction accuracy in subjective embodied contexts precedes true agent intelligence.

🇺🇸 Happy July 4th 🇺🇸

Read our paper here: https://t.co/RlJfFkdpHX and the dataset can be found here:

Looking ahead to 2025, the AI community faces key challenges in value alignment of AI systems, as the adoption of these systems grows. Our work aims to provide concrete tools that support researchers to better understand & improve how human values are embedded into AI systems.

The fact that this underrepresentation of prosocial and civic human values cuts across all three audited RLHF datasets shows a systemic issue that needs to be tackled within the LLM development pipeline and process.

Key findings from our work revealed that all the RLHF datasets were predominantly oriented toward information-utility values, while values around human rights were significantly underrepresented within them. This insight shows why LLMs sometimes struggle with moral reasoning.

To investigate this hypothesis, we developed a framework and conducted a comprehensive audit of three prominent open source RLHF datasets used in training open and closed source language models to foreground the human values embedded in them.

Our work addresses a critical question: What kinds of human values are embedded in RLHF datasets? We hypothesized that human values embedded within RLHF datasets significantly influence the development of LLMs and, thus, require careful attention from researchers & engineers.

We've made both the ground truth & classification datasets publicly available here: https://t.co/22uMQqNUvh For colleagues unable to join during the conference: We're pleased to introduce our work on Value Imprint, which received a spotlight award at the NeurIPS 2024 D&B track.

New paper Alert 🚨 📢 #lastXpostof2024. Grateful once again to the researchers and practitioners who stopped by our presentation on the Value Imprint framework at NeurIPS 2024. For those reaching out and interested in improving/enhancing this work. Thread.. 🧵

The benefit of human values for computational systems is that it provides a framework for dealing with embodied uncertainty in ways that are generative and non-restrictive yet meaningful.

Recently, I've become quite fascinated by the intricacies of my log data and the value of consistent checks. It's amazing how much you can learn from just observing.

Spent some time over the weekend reflecting on my journey thus far.

currently reading some early works in social computing and online communities and seeing some interesting terms like twittering, twitterers, and twittizens.

The implication is that what is true (in our socially constructed world) today might become untrue and disfavored tomorrow, and vice versa.

Most of the things we consider intelligence in current AI Systems are actually a collection of socially constructed human preferences and norms that are situated in time. That is to say that intelligence in those types of contexts does not equal constant and definitive truth.

There is a need to design a new specialization/discipline that focuses on studying how human values and norms are operationalized in technological and computational systems.

Human values are metaphors for societal ordering.

Link to paper here:

Another great presentation by @shivanikapania on the importance of "Annotator Diversity in Data Practices" for AI/ML systems. Happy to see that more attention is being paid to the ethics of data annotation in AI/ML research. Nice team work with @alxndrt & @justaturbo #CHI2023