Search papers, labs, and topics across Lattice.
Affiliation:
7
0
10
3
Semantic convergence in language models may be an inherent trait from pretraining, not just a byproduct of alignment, challenging conventional beliefs about output diversity.
SeKV achieves a remarkable 5.9% performance boost in long-context LLM inference while slashing GPU memory usage by over half.
Language models undergo a crystallization-like process during alignment, transitioning from high entropy to a concentrated distribution that reveals fundamental limits of alignment.
Detection rates for harmful ASCII art plummet beyond certain resolution thresholds, exposing a critical vulnerability in VLM moderation systems.
Topic metadata can dramatically enhance retrieval efficiency in RAG systems, achieving over 8 times faster performance without sacrificing evidence quality.
SproutRAG achieves a 6.1% boost in information efficiency by intelligently structuring document chunks without relying on external LLMs or lossy summarization.
Reducing visual token usage by 46% while improving performance shows that CUAs can leverage more historical data effectively without overwhelming compute budgets.
