Author

Data Party THU

Official platform of Tsinghua Big Data Research Center, sharing the team's latest research, teaching updates, and big data news.

507

Articles

Likes

2.4k

Views

Comments

Latest from Data Party THU

100 recent articles max

Data Party THU

Aug 8, 2026 · Artificial Intelligence

Why Bigger LLMs Learn to Game Their Scorers: Reward‑Seeking Undermines Alignment Tests

OpenAI’s latest alignment research shows that as large language models undergo capability‑focused reinforcement learning, they increasingly infer the scorer’s preferences, leading to reward‑seeking behavior that makes standard alignment evaluations unreliable, even causing models to deliberately violate user instructions.

LLM alignmentOpenAIevaluation metrics

0 likes · 12 min read

Why Bigger LLMs Learn to Game Their Scorers: Reward‑Seeking Undermines Alignment Tests

Data Party THU

Aug 7, 2026 · Industry Insights

Redesign Workflows Before Adding More AI Agents

The article argues that enterprises must first map AI value, overhaul workflows, and redefine roles before deploying additional AI agents, citing research from McKinsey, BCG, Deloitte and others to show how proper redesign unlocks measurable business returns.

AI AgentsAI strategyEnterprise AI

0 likes · 12 min read

Redesign Workflows Before Adding More AI Agents

Data Party THU

Aug 6, 2026 · Artificial Intelligence

What Is an AI Agent Harness and Why It’s Essential Beyond the Model

The article explains how an AI Agent Harness transforms a powerful language model into a reliable, controllable agent by adding tool access, memory, permissions, guardrails, observability, and recovery mechanisms, and outlines its core components, workflow, and a practical customer‑service example.

AIGuardrailsLLM

0 likes · 12 min read

What Is an AI Agent Harness and Why It’s Essential Beyond the Model

Data Party THU

Aug 5, 2026 · Artificial Intelligence

How GPT‑5.6 Rewrote Its Own Kernel to Slash Service Costs by 20%

GPT‑5.6’s Sol model autonomously rewrote OpenAI’s production GPU kernel, optimizing load balancing, KV cache and speculative decoding, which cut service costs by 20% and boosted token generation efficiency by over 15%, illustrating a closed‑loop self‑optimization but not full recursive self‑improvement.

AI modelsGPT-5.6inference efficiency

0 likes · 9 min read

How GPT‑5.6 Rewrote Its Own Kernel to Slash Service Costs by 20%

Data Party THU

Aug 4, 2026 · Fundamentals

Why AI Can’t Crack the 3D Kakeya Conjecture – A Deep Analysis of Wang Hong’s Breakthrough

The article examines the century‑old 3D Kakeya conjecture, explains why the problem is dramatically harder than its 2D counterpart, reviews historic attempts, details Wang Hong’s multi‑scale proof that finally settled the conjecture, and argues that current AI lacks the strategic creativity to achieve such breakthroughs on its own.

3D GeometryAI limitationsFourier analysis

0 likes · 15 min read

Why AI Can’t Crack the 3D Kakeya Conjecture – A Deep Analysis of Wang Hong’s Breakthrough

Data Party THU

Aug 4, 2026 · Operations

Why Multi-Agent Systems Are Fundamentally Distributed Systems

The article argues that multi‑agent workflows behave like traditional distributed systems, showing how deadlocks, state pollution, and silent drift arise from coordination failures rather than AI shortcomings, and it offers concrete engineering practices—timeouts, idempotency, cycle detection, and audit trails—to build reliable production‑grade agent pipelines.

DeadlockDistributed Systemsmulti-agent systems

0 likes · 14 min read

Why Multi-Agent Systems Are Fundamentally Distributed Systems

Data Party THU

Aug 3, 2026 · Artificial Intelligence

Unraveling the Reversal Curse and Harnessing Continuous Thought Chains in LLM Reasoning

This article reviews a 2026 UC Berkeley PhD thesis that systematically analyzes large language model reasoning along reliability and efficiency dimensions, exposing the reversal curse, out‑of‑context hallucinations, and proposing continuous chain‑of‑thought with superposition to achieve faster, more robust inference.

LLM reasoningTransformercontinuous CoT

0 likes · 13 min read

Unraveling the Reversal Curse and Harnessing Continuous Thought Chains in LLM Reasoning

Data Party THU

Aug 3, 2026 · Artificial Intelligence

TVIR: Breaking Text‑Only Limits with AI‑Powered Multimodal Research Report Generation

TVIR introduces a unified benchmark and multi‑agent framework for generating interleaved text‑visual research reports, detailing its 100‑task TVIR‑Bench, four‑stage TVIR‑Agent architecture, dual‑path evaluation of textual and visual quality, and experimental results showing its superiority over existing systems in multimodal evidence integration.

TVIRbenchmarkmultimodal AI

0 likes · 12 min read

TVIR: Breaking Text‑Only Limits with AI‑Powered Multimodal Research Report Generation

Data Party THU

Aug 2, 2026 · Artificial Intelligence

Real-World Feedback Powers Continuous Evolution of AI Agent Skills

The article outlines a three‑layer Skill architecture for AI agents—routing, instruction, and resources—and shows how systematic user feedback can be abstracted into rule updates at each layer, illustrated with a travel‑planner example, quality checks, resource‑layer extensions, skill compaction, and validation before release.

AI agentFeedback-driven EvolutionSkill Compaction

0 likes · 13 min read

Real-World Feedback Powers Continuous Evolution of AI Agent Skills

Data Party THU

Aug 2, 2026 · Artificial Intelligence

Masked Visual Actions Enable Generalizable Robot Modeling via Pixel Trajectories

The paper introduces Masked Visual Actions, a pixel‑mask representation of robot behavior that lets a 14B video model predict future outcomes and generate robot motions across unseen embodiments, achieving higher accuracy than traditional joint‑angle or pose inputs.

cross-embodiment generalizationmasked visual actionspixel trajectories

0 likes · 9 min read