Tandem RLVR is introduced, bringing the recently proposed tandem-training paradigm beyond proof-of-concept settings and into modern RLVR, and initial evidence that tandem rollout structure is a promising direction for keeping RLVR reasoning gains accessible to weaker partners is provided.
Di-Fan Jiao, Raghav Singhal, Robert West et al.· arXiv.org· 0 citations
This work investigates how open-source guard models represent two types of trajectory-level agentic harms and finds that they are linearly readable inside the model, even though guard models predict no better than chance on pairs that differ only in the called tool's schema.
SafeGEO is constructed, an evaluation suite with 22 GEO attack variants across 600 recommendation cases that empirically show that GEO attacks can promote flawed target products and whether agent-side design choices can mitigate this risk.
Qian-Feng Wen, Y. Liu, Xin Liu et al.· arXiv.org· 5 citations· ⚡1
Evaluating six matched base and post-trained models on the Pew American Trends Panel, it is found that base models are the stronger emulators: they produce response distributions closer to human ground truth and better preserve demographic structure.
Seth Grief-Albert, Jessica Y. Bo, Di-Fan Jiao et al.· 0 citations
We use cookies to run the site and, with your consent, for analytics and to show ads.
See our Cookie Policy.
