Beyond Pattern Recognition: Probing Mental Representations of LMs
Moritz Miller, Kumar Shridhar · 2025
Read Beyond Pattern Recognition: Probing Mental Representations of LMs by Moritz Miller, Kumar Shridhar (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2502-16717.txt.
Description
Moritz Miller and Kumar Shridhar probe whether language models truly build mental models or only mimic them, in a 2025 paper. Models ace reasoning tasks when handed a whole problem at once, with polished step-by-step answers. Human thinking instead updates a picture piece by piece as facts arrive.
The authors feed problems gradually, letting each new detail refine the model's internal representation, and compare against standard full-prompt solving. They test text-only and vision-plus-text models across sizes and problem complexities on a math word problem dataset. Both families struggled to maintain and update a coherent mental representation under the gradual regime. Success with full prompts, the results imply, can hide brittle inner bookkeeping. The authors frame the test as a window into how these systems actually think, and find the view less flattering than benchmark scores suggest.
Catalog record
| CREATOR | Moritz Miller, Kumar Shridhar |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 38 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2502-16717