Tag
该论文研究英文构建的 LLM 管道能否跨语言复用，通过支票提取、Mintaka 问答和 FinanceBench 三个任务评估准确性、推理努力与行为稳定性，结论是可有条件复用但仍需按语言按领域验证。
Stanford University's CS329Z course on Engineering AI Agents has started, with lecture slides now public, covering topics from models to autonomous systems and featuring instructors like Diyi Yang and Michael Ryan.
Adding a self-correction loop to an LLM pipeline for structured data extraction reduced consistency from 85% to 62%, due to compounding noise and regeneration drift. The article discusses potential solutions like granular diff mechanisms or deterministic gates.
This paper presents a question-type-specific LLM framework for the BioASQ 14b Task B biomedical QA challenge, using different inference strategies (snippet shuffling, chain-of-thought, multi-agent collaboration) for yes/no, factoid, and list questions, achieving competitive results including first place in the factoid subtask of Batch 4.
FAPO is a framework for fully autonomous prompt optimization of multi-step LLM pipelines, combining prompt editing and structural changes. It outperforms the GEPA baseline in 15 of 18 comparisons, with gains up to +33.8 pp on security tasks.
QUIVER introduces a formal framework for quantifying how perturbations propagate through compound AI systems structured as computation graphs, defining sensitivity matrices, trajectory divergence, bifurcation thresholds, and distribution faithfulness, with validation on production and public pipelines.
This paper formalizes workflow learning in multi-agent LLM pipelines as an interface-constrained semi-Markov decision process (IC-SMDP) and proposes IC-ICQQ, an asynchronous decentralized Q-learning algorithm with a finite-sample bound that decomposes error sources, providing the first finite-sample guarantee for neural Q-learning under decentralized partial observability.
This paper presents a microservice architecture for production document AI pipelines that combine classification, OCR, and LLM extraction, sharing design decisions and batch profiling insights that reveal OCR, not LLM parsing, dominates latency.