Search papers, labs, and topics across Lattice.
École Polytechnique, IP Paris, CNRS)
4
0
9
LLM safety patches can appear effective on benchmarks but cause severe collateral damage to benign behaviors, a critical flaw missed by aggregate metrics.
Prompt injection defenses can degrade LLM agent utility by altering benign behavior; RAISED achieves robustness without this cost through self-distillation on generated scenarios.
KTC achieves over 8x faster claim uncertainty quantification without sacrificing accuracy, making real-time LLM monitoring feasible.
Decoupling communication from policy in MARL lets you drastically reduce bandwidth needs without sacrificing performance, opening the door to real-world deployments with drone swarms and other bandwidth-limited systems.
