Publications

You can also find my articles on my Google Scholar profile.

Preprints


AfterDeleteBench: Auditing Post-Deletion Information Persistence in LLM Agents Permalink

arXiv preprint, 2026

Deleting a record from an LLM agent’s memory doesn’t remove the information: deleted secrets are recovered in about two thirds of cases and rebuilt from stored fragments in 62–99% of cases across Qwen, Mistral, and Phi-3.

Recommended citation: Ehsan Saleh, Roozbeh Nahavandi, Sama Salarian, Raef Bassily, Andrew Perrault. "AfterDeleteBench: Auditing Post-Deletion Information Persistence in LLM Agents." arXiv preprint, 2026.
Download Paper

Many Preferences, Few Policies: Compact Portfolios for Multi-Objective LLM Alignment Permalink

Under review at ICLR 2027

PALM, a provably near-optimal small portfolio of aligned LLMs that serves users with heterogeneous preferences across math reasoning, safety, and helpfulness.

Recommended citation: Cheol Woo Kim*, Jai Moondra*, Roozbeh Nahavandi, Andrew Perrault, Milind Tambe, Swati Gupta. "Many Preferences, Few Policies: Compact Portfolios for Multi-Objective LLM Alignment." Under review at ICLR 2027. (* equal contribution)
Download Paper