论文库

The key-value (KV) cache dominates memory capacity and bandwidth in long-context LLM serving. While 4-bit microscaling formats (NVFP4, MXFP4) shrink the cache and run attention on FP4 Tensor Cores, a fundamental gap remains between algorithmic quantization and system-aware acceleration: existing FP4 KV systems use vanilla quantization that causes severe accuracy drops, whereas accuracy-oriented 4-bit algorithms rely on vector codebooks whose lookup overhead degrades decoding throughput. We present BitDecoding-FP4, an algorithm--system co-design for accurate and efficient native FP4 KV decoding. We mitigate three error sources: Dual-Residual Query Quantization splits queries into two FP4 tensors, cutting query logit error from 60% to 1%; Underflow-Aware Probability Scaling prevents E4M3 scales from rounding long-context probabilities to zero; and Outlier-Aware Key Quantization combines static smoothing, Hadamard rotation, and scale search to suppress key outliers under token-wise scaling. At the system level, we remove the softmax bottleneck via transposed softmax with fused probability quantization and stacked query MMA on paged KV caches. Across dense and MoE models with GQA and MLA, BitDecoding-FP4 stays within 0.5 points of BF16 KV accuracy with NVFP4 (1 point with MXFP4), and tracks BF16 KV in RL training rollouts. In end-to-end serving on NVIDIA B300 with Qwen3-4B and Qwen3-30B-A3B, it delivers 2.4-3.0$\times$ higher peak serving throughput than BF16 KV. The implementation is available at https://github.com/OpenBitSys/BitDecoding-FP4.

Learning dexterous manipulation from limited visual demonstrations is challenging because policies must acquire both perceptual grounding and precise control, while reinforcement learning (RL) in simulation can acquire rich control behaviors from large-scale interaction with privileged information. We introduce \textbf{Latent RL Prior}, a two-stage framework that reuses this control knowledge for data-efficient visual policy learning. During privileged pretraining, we distill large-scale RL rollouts into a flow-matching action decoder conditioned on proprioception and latent task context. During visual adaptation, we freeze the decoder and learn a visual interface that maps images to its latent context using action supervision alone, without privileged labels or latent targets. The same decoder can further be shared across arms by learning coordinated visual contexts for bimanual manipulation. Across three simulated tasks, our method outperforms full-parameter fine-tuning of $π_{0.5}$ and GR00T N1.7 in the low-data regime; on door lifting and bimanual assembly, it surpasses full-data baselines using $10\times$ and $5\times$ fewer demonstrations, respectively. Further analysis shows that direct visual-to-latent conditioning improves adaptation performance and training stability over routing visual predictions through the pretrained state encoder. Robustness to unseen mass variation and greater linear accessibility of contact-force information in decoder representations further provide evidence of retained RL-acquired control knowledge. Project website: https://yixiaowang7.github.io/latent-rl-prior-page/.

Recent Hypergraph-based retrieval-augmented generation (HyperRAG) methods use hyper edges to connect multiple entities simultaneously, enabling more efficient multi-entity evidence organization than pairwise graph structures. However, existing HyperRAG methods often rely on LLM-generated summaries to construct hyperedges, which can introduce hallucinations while also incurring high indexing costs. In addition, during retrieval, existing methods typically rely on either one-hop neighbor expansion or PageRank diffusion. The former may miss useful multi-hop evidence, while the latter can suffer from uncontrolled propagation over excessive hub nodes, leading to semantic drift and noisy reasoning chains. To address these challenges, we propose HyperSU, a novel hypergraph-based RAG framework featuring semantic-unit hyperedges and clue guided bidirectional retrieval. During construction, HyperSU formulates hyperedge construction as an entity-aware minimum-description length (MDL) optimization problem, inducing source-grounded semantic-unit hyperedges that balance sentence-level semantic coherence and entity compactness. It then constructs a hypergraph by modeling each semantic unit as a hyperedge over its co-mentioned entities. During retrieval, HyperSU performs clue-guided bidirectional expansion over the semantic-unit hypergraph, enabling both multi-hop evidence discovery and evidence-anchored noise reduction. Experiments show that HyperSU consistently improves answer accuracy over standard, graph-based, and hypergraph-based RAG baselines, raising average accuracy on GraphRAG Bench by 5.5 points over the strongest baseline (68.7 vs. 63.2), with larger gains on reasoning intensive tasks, where the relative improvement reaches 14.1%.