THUIR_lab

152 posts

THUIR_lab banner
THUIR_lab

THUIR_lab

@thuir_lab

Information Retrieval Lab, Tsinghua University.

Beijing, China Katılım Eylül 2021
25 Takip Edilen356 Takipçiler
THUIR_lab
THUIR_lab@thuir_lab·
We introduce an equity-oriented ranking framework that models heterogeneous provider preferences, moving beyond traditional equality-based fairness. Building on this, we propose EquityRank, a gradient-based algorithm that jointly optimizes user relevance and provider-side equity.
English
0
0
0
11
THUIR_lab
THUIR_lab@thuir_lab·
🔍 Agentic Retrieval: Dynamic planning for multi-step fact analysis & proactive evidence collection. ⚖️ Legal-specific RL: Combining SFT & Reinforcement Learning to optimize legal correctness, logic, and structure. (2/3)
English
1
0
0
22
THUIR_lab
THUIR_lab@thuir_lab·
🚀 Introducing Judge-R1: Enhancing Automated Legal Judgment Generation! 👨‍⚖️ Tackling information omission & reasoning flaws in legal LLMs through a dual-optimization approach. (1/3) #LLM #LegalTech #NLP #Agent #SIGIR2026
English
1
0
0
47
THUIR_lab
THUIR_lab@thuir_lab·
💡 Insights: Agentic methods outperform RAG in structure/fluency, but citation accuracy remains a massive challenge for current LLMs. (3/4)
English
1
0
0
21
THUIR_lab
THUIR_lab@thuir_lab·
🚀 Can LLMs generate high-quality scientific surveys? Current evaluations lack unified standards. Meet SurGE 📊: A new benchmark & evaluation framework for Scientific Survey Generation! (1/4) #SIGIR2026 #LLM #NLP #MachineLearning #AI #Research
English
1
0
0
81
THUIR_lab
THUIR_lab@thuir_lab·
🎉🎉🎉
Qingyao Ai@QingyaoAi

Excited to share that our paper "Analytical Search" has been accepted at #SIGIR2026! 🎉 Traditional search paradigms only emphasize naive information finding and struggle with complex analytical information need (like trend analysis and causal impact assessment). Analytical Search reframes search as an evidence-governed, process-oriented analytical workflow rather than just finding an isolated puzzle piece. Our unified framework integrates: 1. Query Understanding: Extracting implicit intent and decomposing tasks. 2. Evidence Retrieval: A recall-oriented approach across heterogeneous sources. 3. Reasoning-Aware Fusion: Multi-step inference and quantitative/logical aggregation. 4. Adaptive Verification: Generating verifiable, traceable conclusions. It’s time to shift from simple document ranking to end-to-end problem solving! 📄 Read the full pre-print here: arxiv.org/abs/2602.11581 #SIGIR2026 #InformationRetrieval #SearchEngines #LLMs #DataAnalytics #AI

ART
0
0
1
69
THUIR_lab retweetledi
Changyue Wang
Changyue Wang@imBeBr2·
📢Our work on detecting hallucinations in Large Reasoning Models (LRMs) has been accepted at #AAAI2026 as an Oral Presentation! 📷 We introduce RACE, the first hallucination detection framework specifically designed for opaque-box Large Reasoning Models (LRMs). (1/3)
Changyue Wang tweet media
English
1
3
3
406
THUIR_lab
THUIR_lab@thuir_lab·
We present PROSIM, a framework to model prosocial behavior in LLM agents across diverse social settings. LLMs show human-like prosociality, adapt to policy norms, punish unfairness, and reveal how inequities suppress cooperation and erode social norms through networks. (2/2)
English
0
0
0
49
THUIR_lab
THUIR_lab@thuir_lab·
We present PROSIM, a framework to model prosocial behavior in LLM agents across diverse social settings. LLMs show human-like prosociality, adapt to policy norms, punish unfairness, and reveal how inequities suppress cooperation and erode social norms through networks. (2/2)
English
0
0
0
28
THUIR_lab
THUIR_lab@thuir_lab·
In the paper, we propose SDE, a protocol that transfers hidden state changes in addition to language, enabling agents to preserve reasoning signals and achieve SOTA on complex tasks.
English
0
0
0
49
THUIR_lab retweetledi
Changyue Wang
Changyue Wang@imBeBr2·
🎉 Our paper on "Knowledge Editing through Chain-of-Thought" has been accepted at #EMNLP2025 for main conference! We introduce EditCoT, a novel framework that efficiently updates LLMs by editing their Chain-of-thought.@thuir_lab @emnlpmeeting 👉Preprint: arxiv.org/pdf/2412.17727
Changyue Wang tweet media
English
1
1
3
141
THUIR_lab
THUIR_lab@thuir_lab·
It generates backstories to model individual value systems, improving accuracy by over 10%🚀.
English
0
0
0
44
THUIR_lab
THUIR_lab@thuir_lab·
What if an LLM could learn your life story to truly understand your values? Our paper "SimVBG: Simulating Individual Values by Backstory Generation" has been accepted at #EMNLP2025 main conference! arxiv.org/abs/2505.23827
THUIR_lab tweet media
English
1
1
4
464

Keşfet

@JunjieChen826 @emnlpmeeting @elonmusk @BarackObama @taylorswift13 @cristiano @BillGates @NASA