Search papers, labs, and topics across Lattice.
The Hong Kong University of Science and Technology
9
0
14
LLMs can be easily misled by fabricated evidence, with even top-tier models failing to fully mitigate this vulnerability.
Fictitious knowledge entries can safeguard RAG databases without compromising query integrity, achieving high detection rates with minimal interference.
Masking just 5% of attention heads in vision-language models tanks performance on long-context tasks, revealing a surprisingly sparse and critical set of "multimodal retrieval heads" that attend to both text and images.
Legal LLMs can be taught to respect the arrow of time, outperforming state-of-the-art models by up to 30% on legal reasoning tasks simply by enforcing temporal consistency via reinforcement learning.
Jointly ranking answers to complex knowledge graph queries with multiple free variables is now tractable, thanks to a neural-symbolic search method that avoids full enumeration.
A dedicated guard agent, trained via reasoning-intensive methods, can effectively neutralize prompt injection attacks in web-navigating agents without sacrificing performance.
ErrorLLM tackles the challenge of refining LLM-generated SQL by explicitly modeling and detecting implicit semantic errors, leading to substantial improvements in text-to-SQL performance.
Current memory systems like RAG and long-context LLMs stumble in AMemGym's interactive long-horizon conversations, revealing critical performance gaps in maintaining consistent user state.
NGDB-Zoo unlocks up to 6.8x faster training for Neural Graph Databases by decoupling logical operators and integrating semantic priors from pre-trained text encoders, all while maintaining high GPU utilization.
