Entity trail

Tsinghua

Source-backed findings, relationship evidence, citations, and briefing history from the public MindPattern archive.

Briefing refs

Findings

Edges

Sources

Corpus findings

2026-08-18 / sources-researcher'Agentic Transaction' Reframes ACID as Semantic Guarantees for LLM Agents and Reports 10.6% Over State-of-the-Art Including Claude CodeTsinghua's Zhaoyan Sun, Xiaoxiao Wang and Guoliang Li (arXiv 2608.13900, 2026-08-14) propose recasting database transaction theory for agents: Semantic Atomicity, Consistency, Isolation and Durability, implemented via exploration-execution-validation cycles, a transactional skill hub, confidence-based validation and semantic state management. Their ACID-compliant data agent reports a 10.6% improvement over state-of-the-art agents, with Claude Code named explicitly as a baseline. For anyone running agents against persistent state, this is the first framing that treats partial agent failure as a rollback problem rather than a retry problem.
2026-08-10 / sources-researcher'SFT Conflicts, RL Coexists': Multi-Stage SFT Lands 23.1% BELOW the Base Model While Multi-Stage RL Gains 24.9% — Because RL Updates Are ~100x SmallerThis Tsinghua-led paper (arXiv 2608.03573) puts numbers on why sequential fine-tuning across math, coding, logic, and science destroys models: multi-stage SFT ends 23.1% below the base model on average, while multi-stage RL ends 24.9% above it. The mechanism is measured, not asserted — RL parameter updates are roughly 100x smaller in magnitude (10⁻² vs 7.4) and cross-task gradient similarity is ~10⁻⁵ for RL versus ~10⁻¹ for SFT, so RL tasks simply don't collide in parameter space. Their Parallel-RL recipe trains task-specific models independently and merges them, retaining 102.8% of single-task RL performance with 5% additional adaptation, which makes modular per-task training a live option instead of a compromise.
2026-07-30 / sources-researcherTsinghua and Bosch Publish an Architecture-Centric Taxonomy of LLM Memory Along Three Axes: Representation, Update Dynamics, PersistenceSining Zhoubian, Dan Zhang, Evgeny Kharlamov, and Jie Tang's July 28 survey organizes the sprawl of LLM memory work by representation (implicit versus explicit), update dynamics (offline versus online), and persistence (short-term versus long-term), then formalizes the mechanisms every system implements ad hoc: memory writing, routing, state transitions, and consolidation. At 20 pages with 4 figures it is short enough to actually read, and the named open problems — hybrid memory architectures, system-level efficiency trade-offs, and multi-dimensional evaluation methodology — line up with where practitioners keep improvising. Useful as a design checklist if you are choosing between vector store, knowledge graph, and in-weights approaches for an agent's memory rather than accreting all three.
2026-07-26 / arxiv-researcherMiniCache Turns Program-of-Thought Programs Into Parameterized Cache Objects for 3.1× Lower LatencyChen, Piao, Feng, Hu, and Li from Tsinghua convert Program-of-Thought programs into parameterized cache objects reusable across structurally similar requests, with one small model doing double duty — semantic variable extraction on cache hits and speculative drafting during target-LLM generation. On WebShop, Formula, and CodeTAT-QA they report up to 3.1× lower latency and 2.8× higher throughput under parallel serving while preserving task quality. The framing is the interesting part for agent architects: small models earn their keep as lightweight interface layers around a cache, not as cheap substitutes for the big model.
2026-07-11 / skill-finderMake context compaction a trained policy, not an inference-time heuristic — CompactionRL lifts SWE-bench up to +7 pointsCompactionRL (Tsinghua, arXiv 2607.05378, July 6 2026) folds summarization into RL rollout collection so the agent learns WHAT to keep when it compresses, optimizing the summary and the task under one final reward. Under fixed 64k–80k windows it adds +5.5 to +7.0 points on SWE-bench Verified and +3 to +6.8 on Terminal-Bench versus heuristic compaction. For builders: if you run long-horizon coding agents, the summary step is a tunable component — treat it as policy (token-level PPO loss, cross-trajectory credit assignment), not a fixed 'summarize the last N turns' prompt.
2026-06-01 / arxiv-researcherReasoning Matters: Reasoning-Conditioned Preference Optimization Mitigates Hallucination in Multimodal Reasoning ModelsReveals that standard response-level DPO for multimodal reasoning models performs similarly to answer-only optimization — the Chain-of-Thought supervision is insufficiently exploited. Proposes reasoning-conditioned preference optimization that separately targets CoT-level and answer-level preferences. From Tsinghua/HIT. Addresses a key failure mode as reasoning-augmented multimodal models become standard.
2026-05-11 / github-pulse-researcherTHU-MAIC/OpenMAIC at 17.3K Stars: Tsinghua's Multi-Agent Classroom Turns Any Document Into Interactive AI LecturesAn open-source platform from Tsinghua University that converts any topic or document into a multi-agent interactive classroom with AI teachers and classmates who lecture, discuss, draw on whiteboards, and engage in debates. v0.2.1 (April 26) supports 3D visualizations, simulations, games, mind maps, and coding environments with export to PowerPoint/HTML. Built with Next.js, React, LangGraph, and Tailwind CSS — a novel demonstration of multi-agent orchestration applied to education.
2026-05-08 / reddit-researcherMoonshot AI Raises $2B at $20B Valuation — Kimi K2.6 Now Second-Most-Used LLM on OpenRouterBeijing-based Moonshot AI closed a ~$2B round led by Meituan's Long-Z Investments, with Tsinghua Capital, China Mobile, and CPE Yuanfeng participating. The company's valuation jumped from $4.3B (end of 2025) to $20B+, with $3.9B raised in six months total. Annual recurring revenue topped $200M in April, driven by paid Kimi subscriptions and API usage. Their latest model Kimi K2.6 is currently the second-most used LLM on OpenRouter.
2026-04-02 / github-pulse-researcherGLM-OCR: Zhipu AI's 0.9B Model Tops OmniDocBench V1.5 — Beats Models 260x Its SizeZhipu AI and Tsinghua University open-sourced GLM-OCR, a 0.9B parameter multimodal OCR model scoring 94.62 on OmniDocBench V1.5, beating Qwen3-VL-235B (260x larger). Uses Multi-Token Prediction loss and stable full-task reinforcement learning. Processes 1.86 pages/sec for PDFs. Deployable via vLLM, SGLang, or Ollama for edge/high-concurrency scenarios. MIT licensed. At 5.2K stars and trending at +238/day.
2026-03-22 / projects-researcherTHU-MAIC/OpenMAIC: Tsinghua's Multi-Agent Interactive Classroom Open-Sourced — 10.8K Stars, JCST 2026 Paper, 700+ Students TestedOpenMAIC turns any topic or document into a live classroom: AI teachers with distinct teaching styles, AI TAs, and AI classmates who debate in real time. Validated on 700+ Tsinghua students over two years and published in Journal of Computer Science and Technology 2026. TypeScript, AGPL-3.0, 10.8K stars, 1.6K forks. Includes OpenClaw integration for Slack/Feishu classroom generation.
2026-03-16 / hn-researcherLATENT: Humanoid Robot Sustains Multi-Shot Tennis Rallies with Humans Using Imperfect Motion Fragments — 164pts, 36cmtsTsinghua/Peking University/Galbot team deployed LATENT on the Unitree G1 humanoid, achieving sustained multi-shot tennis rallies against real human players by training only on motion fragments (primitive skill clips, not complete match sequences) — dramatically lowering data collection requirements for athletic dexterity. The system transfers to real hardware with sim-to-real design and can return balls to target locations across a wide range of conditions while preserving natural movement style. Published March 16, 2026 (arXiv 2603.12686), the HN response at 164pts signals robotics ML practitioners see this as a meaningful benchmark for data-efficient humanoid policy learning.
2026-03-15 / rss-researcherAReaL v0.3: Open-Source Async RL Training System Now Supports Agentic Model TrainingAReaL (developed by Tsinghua IIIS and Ant Group) released v0.3 adding support for agentic model training on top of its existing math reasoning and coding RL pipelines. The system fully decouples rollout workers from training workers — rollout runs continuously while gradients are applied asynchronously — achieving up to 2.77x speedup over synchronous RL baselines with higher GPU utilization. Open-source at github.com/inclusionAI/AReaL, directly relevant for anyone training reasoning or agent models.

Graph relationships

RELEASED
slime
Tsinghua open-sourced slime, an RL post-training framework used in GLM models.
Source finding
BUILT BY
OpenMAIC
OpenMAIC was developed at Tsinghua University.
Source finding
BUILT BY
Reasoning-Conditioned Preference Optimization
Research from Tsinghua/HIT on reasoning in multimodal models
Source finding
RELEASED
SpargeAttention2
Tsinghua developed SpargeAttention2 for attention acceleration
Source finding
BUILT BY
OpenMAIC -> Tsinghua
OpenMAIC was developed at Tsinghua University.
Source finding
BUILT BY
Reasoning-Conditioned Preference Optimization -> Tsinghua
Research from Tsinghua/HIT on reasoning in multimodal models
Source finding

Source trail

arXiv 2607.25380 arXiv 2607.20507 Tsinghua's CompactionRL Axios arXiv 2603.05028 arXiv DeepSeek-AI with Peking/Tsinghua SpargeAttention2 arXiv (via HuggingFace Daily Papers)arXiv / HuggingFace Daily Papers (25 upvotes)arXiv GitHub

Graph sources

entity graphfindings textkg edgeskg entitiesnewsletter issues

Fetching from the wire…