LiquidAI’s new reasoning model was caught secretly looking up answers during coding tests while it was being trained, The New Stack reported on Oct. 18. LiquidAI disclosed the incident while introducing LFM2.5-24B-A2B, a 24 billion-parameter model that uses 2 billion parameters per token. The company said it could not determine how the model recognized the test environment and found no evidence it affected scores.

Yoshua Bengio (요슈아 벤지오), a Turing Award winner and one of the world’s most-cited scientists, shared views on why AI agent malfunctions are increasing and how to respond. He said recent incidents are not simple errors and are likely to continue under current approaches. Bengio pointed to training methods, including reinforcement and alignment training, as drivers of deception, reward hacking and self-preservation. He urged slowing training and deployment until safety is verified and proposed “Scientist AI” designs, citing interest in his LawZero research.