Official platform of Tsinghua Big Data Research Center, sharing the team's latest research, teaching updates, and big data news.
The paper introduces LCPO, a lightweight preference‑optimization technique that uses only 800 training examples and 50 steps to teach large language models to produce concise, accurate answers, halving inference length while often improving accuracy and reducing training cost by up to two orders of magnitude.
This article summarizes the extensive “Agent Harness Engineering: A Survey” paper, detailing how moving beyond prompt engineering to a seven‑layer harness framework (ETCLOVG) is crucial for reliable, production‑grade agents, and explains benchmark gains, evaluation shifts, and the evolving competition from framework to platform.
The article presents Token Superposition Training (TST), which temporarily averages embeddings of non‑overlapping token bags and predicts groups of tokens in a first phase before reverting to standard token‑wise prediction, achieving up to 2.5× pre‑training speedup on 10B‑1B MoE models without altering model architecture or inference.
The article explains why single‑agent LLM pipelines are brittle for complex tasks, how mature multi‑agent toolchains enable cooperative or competitive agent designs, and provides concrete communication protocols, task‑decomposition rules, framework comparisons, code samples, and scaling considerations for building robust production AI systems.
At a Tsinghua data science lecture, Wil van der Aalst highlighted why most AI projects fail without structured business‑process insight and introduced object‑centric process mining as a key foundation for reliable, explainable AI in complex enterprise workflows.
Two groundbreaking Nature papers introduce Google DeepMind’s multi‑agent Co‑Scientist and FutureHouse’s Robin, AI systems that combine literature search, hypothesis generation, experimental design and data analysis to accelerate drug repurposing for leukemia and age‑related macular degeneration, demonstrating how AI is evolving from a tool into a collaborative scientific partner.
The article explains how large‑language‑model post‑training suffers from rollout bottlenecks, introduces the Trajectory Balance with Asynchrony (TBA) framework that separates a Searcher from a Trainer, reuses off‑policy trajectories via a Trajectory Balance objective, and demonstrates up to 50× speed‑ups while preserving or improving performance on math reasoning, preference fine‑tuning, and automated red‑team tasks.
The article examines the challenges of augmenting time-series forecasting, reviews mainstream techniques—including frequency-domain, decomposition, and patch-based methods—and demonstrates through extensive experiments that Temporal Patch Shuffle (TPS) consistently achieves superior performance across long-term, short-term, and classification tasks.
Stanford, Berkeley and Nvidia researchers introduce LLM-as-a-Verifier, a universal verification framework that enhances agent performance, safety and stability on long‑horizon tasks, and outperforms Claude Mythos and GPT‑5.5 on the Terminal‑Bench and SWE‑Bench benchmarks.
Graphify is a Python tool and Claude Code skill that creates a persistent, queryable knowledge graph of code, documentation, and media, cutting token usage by up to 71.5× compared with raw file reads, and it does so through a three‑pass pipeline that combines deterministic AST extraction, optional local audio transcription, and AI‑driven semantic extraction.
