Unverified50% confidenceBenchmarkExact time
Compared with TorchAO's 4-bit AdamW, ZIP-SR and ZE-EDEN reduce the average validation loss gap relative to 32-bit AdamW at every evaluated scale, with a maximum gap reduction of 70%.
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
10/13/2026
First Seen
Valid until: 1/11/2027
Sources
Related Entities
Related Claims
UnverifiedAn EMNLP study found that 4-bit models suffer an average retrieval loss of 23% at 128,000 tokens, with 70B LLaMA losing one-third of its score while QWEN of the same scale loses almost none.69% similarUnverifiedAcross eleven datasets spanning images, tabular data, and text, Phase-HDC reduces memory by 16–23× compared to standard float32 Adam and by 4–6× compared to 8-bit Adam.66% similarUnverifiedint8量化将参数从16位压缩至8位，理论上减少约50%显存占用；int4将显存占用降低至原始的25%左右64% similarVerified4-bit quantization reduces model size to 1/8 of the original at the cost of approximately 1–3% accuracy loss.63% similarVerifiedINT8 quantization loses approximately 0.5% accuracy while halving VRAM usage; INT4 quantization compresses VRAM to 1/4 of the original with an accuracy loss of approximately 1–3%.61% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1013234API
curl https://kongchang.com/api/v1/knowledge/claims/1013234MCP
get_claim(id=1013234)