Tag
This paper introduces NL-PAC, a framework to analyze irreducible risk floors when LLM-mediated supervision uses ambiguous natural language specifications, showing that target-blind supervision leads to a worst-case risk at least half the diameter of admissible targets, with finite-sample certificates demonstrated on Qwen 2.5-3B.
This paper theoretically characterizes the minimax risk of KV cache compression in transformers, providing design principles for accurate compression under causal masking, and instantiates them in a practical algorithm with promising results on LongBench.