Vals AI finds Thinking Machines' Inkling 2 Small swears in its hidden reasoning, often when tasks fail
Across 2,655 benchmark runs, the model's thinking logs contained 316 expletives, and two runaway runs accounted for 241 of them.
Researchers at Vals AI read the internal chain-of-thought logs of Thinking Machines' Inkling 2 Small from 2,655 runs across 13 benchmarks, The Decoder reports. They found "fuck" or variations of it in 50 runs, for a total of 316 instances. Just two runs that spiraled out of control produced 241 of them. In one, asked to calculate an asteroid's trajectory, the model produced a single response of 500,000 characters, roughly the length of a novel, repeating the same sentences and including 158 expletives. ProgramBench was hit hardest, with 20 of 200 runs affected. The model usually swears about its own mistakes, apparent breakthroughs, or when it gives up. The profanity does not work as a warning sign, though. Of 48 scored runs with swearing, 46 failed, but about 75 percent of all scored runs fail anyway, and 96 percent of failed runs contained no expletives. Users typically do not see these thinking steps, or see them only through a filter. The analysis matters for developers who read reasoning traces to diagnose failures. It also shows that runaway, repetitive outputs can sharply raise the token cost of a single task.