Dream2Discern predicts multi-agent failures from hard variants
Original title: Dream2Discern: Learning from Hard Latent Variants for Runtime Failure Prediction in Multi-Agent Systems
It reports over 90% trajectory-level weighted accuracy on AFTraj-2K and StepShield without collecting new trajectories.
Dream2Discern is a runtime auditor for LLM multi-agent systems that predicts terminal outcomes from execution prefixes. During Dream, it perturbs representation operators to raise prediction loss and create hard latent variants. During Discern, it trains on those variants using the original labels, regenerating them as learning proceeds. The authors say it outperformed competitive failure auditors on AFTraj-2K and StepShield while balancing accuracy and timeliness.
- No additional trajectory collection or outcome verification is required.
- Weighted trajectory-level accuracy exceeds 90% on both reported benchmarks.
- Benchmarks named are AFTraj-2K and StepShield.
- Variants are regenerated as the auditor's predictions change.
- Products
- Dream2DiscernAFTraj-2KStepShield
From the original article
1 min read at huggingface.co
Reliable deployment of LLM-based multi-agent systems calls for runtime auditors that predict terminal outcomes from observed execution prefixes. These auditors typically learn from a limited corpus of trajectories with verified outcomes.
By Boxuan Zhang, jianing Zhu, Xiaotian Han · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.