Daily papers

AI engineering 6

  1. 1 ▲ 17 Xiaoqi Wang et al.

    MEM-SQL: Memory-Empowered Text-to-SQL

    Verified query history raises execution accuracy and cuts token use and latency when requests can reuse prior SQL.

  2. 2 ▲ 19 Evan Li et al.

    Mobile AI Factory: Building Specialized Language Models for Your Mobile Device

    The factory finds phone-specific SLMs that meet accuracy goals, nearing exhaustive search with 19× less GPU time.

  3. 3 Ibrahim Memon et al.

    LuthorSQL: A One-Pass Text-to-SQL Model as System 1 for a Slower LLM Agent

    A local SQL gate answers nearly half of queries, cutting API cost 33.3% at the same accuracy.

  4. 4 ▲ 48 Wei-Chieh Huang et al.

    MIRIX-2.0: Multi-Agent Memory System

    Typed records, provenance links, and retrieval budgets improve agent memory without letting inspected evidence grow with history.

  5. 5 ▲ 35 Sicheng Feng et al.

    MMFlash: Faster Speculative Decoding for Vision-Language Models

    Adaptive drafting improves multimodal speculative decoding, with lossless speedups up to 3.58× on benchmarks.

  6. 6 Wei Feng et al.

    Reducing Batch Makespan for LLM Agent Workloads through Inference-Environment Coordination

    Coordinating GPU allocation, cache-aware routing, and environment scheduling cuts agent batch makespan 10.4% to 65.0%.