Large Models

POPS: Recovering Unlearned Multi-Modality Knowledge in MLLMs with Prompt-Optimized Parameter Shaking

An adversarial attack that recovers supposedly unlearned multi-modality knowledge from MLLMs via prompt-suffix optimization and fine-tuning, exposing vulnerabilities in machine unlearning defenses.

CATNIP: LLM Unlearning via Calibrated and Tokenized Negative Preference Alignment

A token-level confidence-calibrated negative preference alignment method for LLM unlearning that removes undesirable knowledge without requiring retention data or contrastive pairs.

LLMs Can Get "Brain Rot"!

We find that LLMs can get Brain Rot just like human after browsing enormous brainless social media.

More is Less: The Pitfalls of Multi-Model Synthetic Preference Data in DPO Safety Alignment

A study revealing safety-specific pitfalls of multi-model synthetic preference data in DPO alignment.

GuardAgent: Safeguard LLM Agents by a Guard Agent via Knowledge-Enabled Reasoning

The first automated guardrail for agents.

SEAL: Steerable Reasoning Calibration of Large Language Models for Free

A training-free approach that calibrates chain-of-thought reasoning in LLMs, improving accuracy while reducing computational overhead.

MedHallu: A Comprehensive Benchmark for Detecting Medical Hallucinations in Large Language Models

Benchmark for medical hallucination by LLMs.

Extracting and Understanding the Superficial Knowledge in Alignment

We examined how superficial LLM alignments are thru a linear distillation method.

GuideLLM: Exploring LLM-Guided Conversation with Applications in Autobiography Interviewing

We develop a chatbot for reminiscence therapy

LLM-PBE: Assessing Data Privacy in Large Language Models

A comprehensive privacy assessment of LLMs.