In 2025, Kambhampati’s lab replaced correct reasoning traces with incorrect ones without reducing formal reasoning performance. Correct-trace training still produced invalid records.
In 2025, Kambhampati’s lab replaced correct reasoning traces with incorrect ones without reducing formal reasoning performance. Correct-trace training still produced invalid records.
More from this article
See all →Human-readable chains of thought are unnecessary: meaningless filler tokens, including dots, can…
A model’s hidden reasoning tokens can be arbitrary embeddings, foreign-language fragments,…
OpenAI has not publicly revealed raw chains of thought since 2024. Its released traces are…
A reasoning trace can look like an auditable record of model thought while having no reliable…
A correct answer followed by plausible thoughts creates the seductive illusion that those thoughts…
Calling a program “UNDERSTAND” or “GOAL” can smuggle the missing capability into its name and fool…
A reasoning model may be simply an LLM with specialized training, not a fundamentally new…
A model can imitate enough reasoning examples to produce a verifiable answer without learning a…
Don't lose this one
A free account saves cards like this to your Collection, and Korva resurfaces them so you actually remember.