The final Fate joins COINSE literature, as a paper titled Atropos: Improving Cost-Benefit Trade-off of LLM-Based Agents under Self-Consistency with Early Termination and Model Hotswap is accepted into ISSTA 2026. Congratulations, Naryeong!

Atropos extends our existing work, Lachesis, which predicts whether a set of LLM agent inferences would achieve correct output based on majority voting. Atropos moves the prediction points earlier in inference, and aims to predict the correctness of the output mid-inference. Surprisingly, the early termination prediction accuracy is 85\% when performed in the middle of inference.

Further, Atropos introduces the concept of hot-swapping: if we start a task on a small local language model and predict the task to fail, Atropos can lift up the context up to that point and migrate it to a stronger model, improving the chance of successful output generation. Based on evaluation of three state-of-the-art agents, Atropos can achieve 74.35\% of proprietary LLM performance while spending only 23.9\% of monetary cost, by only migrating tasks that cannot be successfully completed on the local machine.

This paper is also the final piece of trilogy of three Fates:

• Clotho: Measuring Task-Specific Pre-Generation Test Adequacy for LLM Inputs, Yoon, J., Kim, S., Feldt, R. and Yoo, S., Proceedings of the ACM on Software Engineering (PACMSE), 3(FSE)
• Lachesis: Predicting LLM Inference Accuracy using Structural Properties of Reasoning Paths, Kim, N., Kang, S., An, G. and Yoo, S., Proceedings of the 6th International Workshop on Deep Learning for Testing and Testing for Deep Learning
• Atropos: Improving Cost-Benefit Trade-off of LLM-Based Agents under Self-Consistency with Early Termination and Model Hotswap, Kim, N. and Yoo, S., Proceedings of the ACM on Software Engineering (PACMSE), 3 (ISSTA)