Counterfactual token generation paper cited in stateless-LLM debate: near-zero-cost what-if reasoning for any LLM

gerardsans · x · 2026-09-05

Extending the "models have no self" debate, gerardsans cited an arXiv paper noting LLMs are stateless and can't reason about counterfactual alternatives to past tokens (e.g., rewriting a story if a different protagonist name had been sampled).

The paper's fix: a causal model of token generation built on the Gumbel-Max structural causal model lets any LLM perform counterfactual token generation at almost no extra cost — trivial to implement, no fine-tuning or prompt engineering — demonstrated on Llama 3 8B-Instruct and Ministral-8B-Instruct.

Original post →

More from Research

Research channel →