arxiv arXiv cs.CL · 3d ago · src: 6d ago · research

KuaFu compresses user behavior into understanding at billion scale

from English

Tencent introduces KuaFu, a unified behavior-compression layer that condenses raw user history into compact representations for conversational agents and recommenders. The system uses a two-axis projector to compress each behavior item into 2-4 tokens of width 128-256, addressing bottlenecks in processing long sequences for a billion users.

KuaFu matches or exceeds uncompressed single-task production models across four profiling tasks while raising per-GPU throughput by 37%-350% and saving 190 GPUs. On public benchmarks, it outperforms prior compressors at the same ratio, with a 4B model surpassing an 8B counterpart on RecBench.

After running on Tencent's advertising and recommendation platform for ten months, KuaFu lifted overall GMV by 1.37%.

Importance 2/3 arXiv cs.CL Tencent Inference efficiency Retrieval & RAG

Read original

media Hugging Face Forums · 12d ago · 13 views

Semantic routing reduces Qwen2.5 input tokens by 41% and latency by 45% in 14K context stress test

A GPU stress test using frozen 4-bit Qwen2.5-7B-Instruct on HotpotQA bridge questions demonstrates that task-aware semantic routing significantly improves efficiency and accuracy under high-distractor conditions. The study compared full capped context against a query-aware semantic selector retaining 60% of the budget, revealing substantial gains in speed and precision.

lab Tencent Hunyuan (HF) · 17d ago · 21 views

Tencent releases Simple Attention Sparsification checkpoints for Qwen3

Tencent has released Simple Attention Sparsification (SAS) checkpoints that learn to rank and select KV blocks for each query in Qwen3 models. Unlike methods distilling dense attention scores, SAS adds continuous gates to selected blocks, enabling end-to-end optimization of context ranking with the language-modeling loss.

media Hugging Face Forums · 17d ago · 16 views

Bridge-aware routing beats BM25 and random promotion on Qwen2.5-7B

A 600-case evaluation on the MuSiQue dataset using frozen Qwen2.5-7B-Instruct and MiniLM models tests whether bridge-aware context selection can outperform full context and other baselines.

lab Tencent Hunyuan (HF) · 24d ago · 37 views

Tencent releases EVIE-4.5B with dynamic Prefix-MRL and training-free token compression

Tencent has released EVIE-4.5B, a high-performance multilingual visual document retrieval model featuring dynamic Prefix-Multivector Retrieval Latency (Prefix-MRL) and training-free Hierarchical Agglomerative Clustering (HAC) token compression.

lab Tencent Hunyuan (HF) · 24d ago · 32 views

Tencent releases EVIE-8B, a 4096D visual document retrieval model

Tencent has released EVIE-8B, a high-capacity visual document retrieval model featuring 4096-dimensional token representations and bidirectional full attention. The model serves as a teacher foundation for the lightweight EVIE-4.5B Prefix-MRL variant and achieves state-of-the-art performance on visual document retrieval benchmarks.