Tencent's KuaFu compresses billion-scale user behavior into 2-4 cacheable tokens per item
_reachsumit · x · 2026-09-28
Tencent's arXiv paper KuaFu presents an LLM-based user profiling system that compresses each behavior item into a few cacheable tokens, with RL training to curb hallucinated user interests.
Industrial bottlenecks addressed
- Task-specific pipelines still read hundreds of behavior items per user (tens of thousands of prompt tokens) for tasks like interest summarization
- Profiles refresh weekly for a billion users (100K QPM aggregate), setting a hard throughput floor under fixed GPU budgets
- Truncation or coarse compression introduces four hallucination types (fabrication, omission, date misattribution, broken logic) that only surface as diffuse downstream degradation
Method
- A two-axis projector compresses each item to 2-4 tokens of width 128-256 (10x token-axis, 20x width compression; per-item cache 10KB → 0.5KB)
- Fidelity-oriented four-stage training with layered intermediate evals
More from Research
- KAN author Ziming Liu: LLMs learn alchemy, not the science of AI — ZimingLiu11 · 2026-10-24
- Training-Physics Intuition Compressed Into a Decision Tree Hits 78.2% Out-of-Fold — ZimingLiu11 · 2026-10-24
- LLMs Rank Optimizers at 85% but Are Data-Blind on Architectures, New Benchmark Shows — ZimingLiu11 · 2026-10-24
- NVIDIA releases 43K instruction-hierarchy rubrics for judging system vs user prompts — kalyan_kpl · 2026-10-24
- Architecture IQ benchmark shows LLMs still lag humans in training intuition, ResNet-style — ZimingLiu11 · 2026-10-24
- François Chollet declares "Symbolic learning works" in terse but weighty statement — fchollet · 2026-10-24