Reposted by Tal Linzen
Can LLMs introspect?
Anthropic said yes.
However, CDS PhD student Shashwat Singh, CDS Associate Professor Tal Linzen (@tallinzen.bsky.social) & CDS Faculty Fellow Shauli Ravfogel (@shauli.bsky.social) found the evidence falls short.
nyudatascience.medium.com/cds-research...
nyudatascience.medium.com
CDS Researchers Challenge Anthropic’s Evidence That Language Models Can Introspect
In 2025, Anthropic reported that its Claude models could detect when researchers injected a concept directly into their neural activity…
![Overview of the paper.
Title: The Emergent Symbolic Structure of Artificial Neural Networks
Authors: Tom McCoy, Paul Soulos, Tal Linzen, Paul Smolensky
Left: Neural networks encode information in vectors (there is then an image of a vector), yet they excel at tasks long thought to require symbolic structure (there is then an image of a symbolic representation, specifically a syntax tree). How do LLMs do it?
Right: We find that LLM representations can be closely approximated with symbolic structures. This approximation lets us edit the structure of an LLM’s output by editing the structure of its internal representations, as shown. There is then an image of two edits to LLMs. In the first one, the original input is 3 + 6 * 8, with an answer of 51. But if we swap the positions of the 3 and the 6, the output becomes 30. In the second one, the original input is a Python command repeating the list [Z, U] three times, producing [Z, U, Z, U, Z, U]. But if we edit the input in a way that adds a Q at the end of the input, the output becomes [Z, U, Q, Z, U, Q, Z, U, Q].](https://cdn.bsky.app/img/feed_thumbnail/plain/did:plc:bcsvlv7ubeagj5virdvatadd/bafkreic3zm2rpc4vtqrz2t3y4vcl3t2owp2sxay42vuh5zm5zkg4jhr64e)
