AI agent teams that check each other's work before a bad answer snowballs into a wrong one.
A new paper describes HASSUM, a framework that estimates how reliable an AI agent's answer is by measuring semantic entropy and semantic density - essentially, how much an answer's meaning wavers across multiple tries, rather than just how confident the model's raw output probabilities look. When those signals suggest shaky ground, the system can trigger extra steps: verifying the output, reprompting the agent, adding another round of deliberation, or picking the most consistent answer from several attempts. The researchers built it into a hierarchical multi-agent setup and tested it on three benchmarks - StrategyQA, JailbreakBench, and TruthfulQA - covering complex reasoning, adversarial prompts, and factual accuracy. Using semantic entropy and semantic density together beat using either one alone, and the choice of confidence threshold and model size both changed how well it worked.
The real problem this targets is one multi-agent systems inherit and amplify: when one agent hands a bad answer to the next, hallucinations do not just persist, they compound. Most orchestration setups today assume every step is trustworthy and move on. Because HASSUM works off the meaning of an answer rather than a specific model's internals, it can in theory bolt onto different agent architectures instead of being a one-off fix.
That portability is the pitch, but it is untested outside three benchmarks and says nothing about the extra latency or compute cost of all that reprompting and verification - the kind of overhead that tends to get quietly left out of academic papers.