Researchers introduce GateRQ, a method for compressing conditional-memory tables—large stores of text patterns attached to a language model. Instead of assigning extra precision to frequently accessed rows, GateRQ uses a lookup’s gate value to decide whether to read additional 1-bit codes; the paper says the gate can be computed from a 1-bit code.
On DeepSeek-V4.1-Flash, the researchers report 7.8 times less GPU table memory and 11%–31% higher vLLM throughput than with FP8 tables. They report losing about 1 PopQA point at an average 1.50 bits read per lookup, or 0.2 points at 2.10 bits, across two models; the method does not require retraining.
Why it matters
For teams serving models with conditional-memory tables, GateRQ could reduce GPU memory use while increasing throughput, though the reported results do not establish how it performs outside the tested models.
