Why LLMs Think in English and How It Affects Their Performance

• ​Models default to English as their internal reasoning medium, even when the user provides inputs in other languages.

• ​This internal switch improves correctness on many tasks but introduces translation-induced distortions that propagate through the reasoning chain.

• ​In tightly constrained tasks (logic puzzles, grid constraints, multi-step combinatorics), the same models produce confident yet structurally invalid outputs.

• ​The session dissects why English becomes the computation layer, what information is lost in the cross-language transitions, and where reasoning collapses despite surface fluency.

• ​Includes observations from controlled experiments showing brittle behavior, unstable intermediate steps, and mode collapse under puzzle constraints.

​Speaker Bio:
Deepon Halder - works on data curation, benchmarking, and unconventional ideas at AI4Bharat, IITM.
Website - deeponh.github.io

Pre-read:
https://arxiv.org/abs/2510.20647v1
https://arxiv.org/abs/2510.24932