541 Followers
418 Following
893 Posts

Professor at Indiana University Bloomington. Develops network, data, and machine learning methods and applies them to (mostly) social problems.

@yy@twitter is NOT my account.

homepagehttps://yongyeol.com
blueskyhttps://bsky.app/profile/yyahn.bsky.social
Threads.nethttps://www.threads.net/@yyahn

A cool new paper about detecting "hallucinations" (or "bullshitting") of LLMs.

The idea is deceptively simple. If we can cluster potential answers based on the semantic entailment, and then calculate the entropy of the potential answers, then we can see how "unsure" the model is. The more "unsure" it feels, the likelier that the generated answer is a confabulation.

"Detecting hallucinations in large language models using semantic entropy"

https://www.nature.com/articles/s41586-024-07421-0

Detecting hallucinations in large language models using semantic entropy - Nature

Hallucinations (confabulations) in large language model systems can be tackled by measuring uncertainty about the meanings of generated responses rather than the text itself to improve question-answering accuracy.

Nature

Hey I made a visualization of
the proposed Elon Musk pay package.

https://yyahn.com/elon_paypackage/

Let me know what you think! πŸ™ˆ You can also suggest improvements: https://github.com/yy/elon_paypackage

Just how big is the proposed pay package for Elon Musk?

When we build two sets of models that rely on these two categories of features, they perform very differently (as you may expect). The visual-model can only marginally predicts the price whereas the social-model approaches the predictive power of auction house professionals (who set the expectation for the price range). In other words, pretty much all predictive power comes from social signals such as β€œwhat was the price of the previous piece by this artist?” 🧡
We divided the features of each art piece into two categories: visual πŸ‘οΈ and social πŸ‘€. Visual features include traditional computer vision features as well as high-level features from CNN. Social features are whatever non-content information we have for the artist, market, etc. In a way, social features are what we can know when an artist says "I have created a 4x3 painting but I can't show it to you." 🧡
Glad that my son now gets the coolness of infinite recursion (and 🌈)

The model gives us an intuitive way to think about simple v. complex contagion as well as vulnerability to mis/disinfo.

When the existing belief system is "primed" to accept a certain belief, it can spread virally; when it conflicts with existing beliefs, social reinforcement is critical.

🧡

This model can also reproduce the optimal modularity phenomenon. When the new belief is going against the existing belief system, the clustering matters more and there is an optimal amount of clustering that can spread the information best.

🧡8/

It is also possible to reproduce the famous experiment by Damon Centola.

When we have the "simple" configuration, it spreads better in a random network than a clustered "large world"; when we have the "complex" setting, it spreads better in the clustered large-world than a small-world!

🧡8/

But, if new belief would disrupt a stable belief network, it will be difficult to be adopted.

So, with belief weights, depending on the relationship between the new belief and the existing beliefs, the model exhibits two distinct dynamics: simple & complex contagion!

🧡7/

We (w Nathaniel Rodriguez & Johan Bollen) introduced a simple model of belief interaction network in 2016: https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0165910

In this model, beliefs are signed edges connecting concepts, which lead to the belief interaction via social balance condition. Another point was that ..

🧡3/

Collective Dynamics of Belief Evolution under Cognitive Coherence and Social Conformity

Human history has been marked by social instability and conflict, often driven by the irreconcilability of opposing sets of beliefs, ideologies, and religious dogmas. The dynamics of belief systems has been studied mainly from two distinct perspectives, namely how cognitive biases lead to individual belief rigidity and how social influence leads to social conformity. Here we propose a unifying framework that connects cognitive and social forces together in order to study the dynamics of societal belief evolution. Each individual is endowed with a network of interacting beliefs that evolves through interaction with other individuals in a social network. The adoption of beliefs is affected by both internal coherence and social conformity. Our framework may offer explanations for how social transitions can arise in otherwise homogeneous populations, how small numbers of zealots with highly coherent beliefs can overturn societal consensus, and how belief rigidity protects fringe groups and cults against invasion from mainstream beliefs, allowing them to persist and even thrive in larger societies. Our results suggest that strong consensus may be insufficient to guarantee social stability, that the cognitive coherence of belief-systems is vital in determining their ability to spread, and that coherent belief-systems may pose a serious problem for resolving social polarization, due to their ability to prevent consensus even under high levels of social exposure. We argue that the inclusion of cognitive factors into a social model could provide a more complete picture of collective human dynamics.