Track papers across labs, topics, and trends.
Search papers, labs, and topics across Lattice.
National Institute of Electronics and Information Technology (NIELIT) Indian Institute of Information Technology
3
1
7
KV-cache compression methods can fail catastrophically by evicting dormant but critical tokens, but a simple "sponsorship" mechanism can prevent this, boosting credential retrieval from 0% to 100% even with aggressive compression.
LLM explanation faithfulness varies wildly depending on how you test it, and might even be *anti*-faithful, so stop relying on single-intervention benchmarks.
LLMs are far more susceptible to authority and framing biases than the field's obsession with demographic bias suggests.