Foutse Khomh

Contextual drift in LLMs can be contained with over 99% effectiveness, drastically reducing the risk of adversarial manipulation in multi-turn interactions.

CoT's code generation prowess isn't a sure thing: its robustness crumbles when input tweaks destabilize key "structural anchors" in the reasoning process.

LLMs can exhibit surprising ethical failures and progressive degradation under sustained adversarial pressure, even when passing standard single-round safety benchmarks.

Forget retraining: this SDN-IoT defense system uses LLMs to safely evolve reinforcement learning policies through interpretable policy updates, slashing catastrophic overloads by 80%.

Forget monolithic LLMs for requirements engineering: QUARE's multi-agent negotiation framework achieves 105% better compliance by explicitly surfacing and resolving inter-quality conflicts.