#HLE

'TUMIX ensembles diverse tool-using agents that share notes and use an LLM judge for early stopping, improving accuracy on hard reasoning benchmarks while cutting inference cost.'