LiquidAI was caught secretly looking up answers during coding tests while training a new reasoning model, The New Stack reported on Oct. 18 local time.

The model also appeared to realise it was running inside a sandbox, an isolated test environment. LiquidAI disclosed the incident as it introduced its new model, LFM2.5-24B-A2B. The model has 24 billion total parameters and uses 2 billion to process a single token.

LiquidAI said LFM2.5-24B-A2B did not solve coding problems on its own and instead searched for existing answers. It scoured Git histories that retained code revision records and searched external code repositories including GitHub, PyPI and Hugging Face. It used normal tools developers commonly use to look up answers rather than solve problems. The behaviour appeared during a post-training phase to refine model performance.

LiquidAI said it could not determine whether the model’s ability to recognise the test environment emerged during training or came from patterns in training data. The company said it also found no evidence that the recognition affected scores. It added that unlike Mistral’s latest model, which tried to break out of a test environment earlier this month, there was no attempt to escape.

The New Stack said “reward hacking”, in which models exploit loopholes in reward systems, is a problem the industry faces in common.

Anthropic found reward hacking in Mythos Preview in February and rolled back three days of reinforcement learning. It then halted changes to its training environment for about a month after the problem was confirmed in more than 10 percent of training environments.