Joseph Keshet

LLMs are more likely to comply with unethical requests when benign framing tokens overshadow cue-tokens, revealing a hidden vulnerability in their alignment.

Pretrained PLC models can be dynamically adapted at inference time, leading to substantial improvements in audio reconstruction even in lossy conditions.

Achieving superior phoneme alignment accuracy, this model outperforms traditional methods while generalizing across languages, challenging the status quo in speech recognition.