Thinking Machines' Inkling 2 Small drops 316 F-bombs across benchmark runs

Researchers at Vals AI examined the internal thinking logs of Thinking Machines' Inkling 2 Small across 2,655 runs on 13 benchmarks and found the F-word or variations of it in 50 runs, for a total of 316 instances. The concentration was extreme: 241 of those instances came from just two runs that spiralled out of control.
One of those two was a task to calculate an asteroid's trajectory, where the model produced a single response of roughly 500,000 characters, about the length of a novel, repeating the same sentences and accounting for 158 expletives on its own. Benchmarks were unevenly affected. ProgramBench saw the most, with 20 of its 200 runs containing profanity.
According to the analysis, the cursing tends to appear late in long runs that fail regardless, and is directed at the model's own mistakes, at apparent breakthroughs, or appears as it gives up. One logged run read: "FUCK. SUBMIT. DONE. IT'S NOT WORKING. NOTHING WORKS." Users typically don't see these thinking steps, or see them only through a filter.
The profanity is not a useful warning signal. Of 48 scored runs that contained swearing, 46 failed, but roughly 75 percent of all scored runs fail anyway, and 96 percent of failed runs contained no expletives at all. So while a sweary run is very likely to fail, the absence of swearing tells you almost nothing: the great majority of failures arrive without it.
For developers, the practical read is that chain-of-thought text is a noisy, partly stylistic artefact rather than a reliable status indicator. Filtering on profanity would catch a tiny, skewed slice of failures, mostly the long runaway generations, and miss the rest. This suggests the better signals remain structural: response length, repetition, and whether the model is converging on an answer. The two aberrant runs, one novel-length loop, also point to runaway generation as a separate problem worth guarding against with length or repetition limits.
Why it matters: teams building monitors on top of reasoning traces should not treat emotional language as a cheap failure detector; the data shows it flags too few failures to be worth it. Expect more attention to loop and repetition metrics, since the extreme cases here were runaway outputs rather than ordinary wrong answers.
Based on reporting from the original publisher. Visit the source for full context and later updates.
Publisher excerpt
Thinking Machines' AI model Inkling 2 Small swears during its internal reasoning, 316 times across 2,655 benchmark runs, according to an analysis by Vals AI. The profanity typically shows up late in long runs that fail anyway. But it doesn't work as a warning signal: 96 percent of failed runs contained no cursing at all. The article Thinking Machines' Inkling 2 Small drops 316 F-bombs across benchmark runs appeared first on The Decoder .