Dongqi Fu
@DongqiFu_UIUC
Followers
1K
Following
433
Media
7
Statuses
55
Research Scientist at @AIatMeta | Ph.D. from @IllinoisCDS | Working on #GeometricDeepLearning #SequenceModeling #ProbabilisticGraphicalModel
Joined January 2017
Haystack Engineering: Context Engineering for Heterogeneous and Agentic Long-Context Evaluation.
arxiv.org
Modern long-context large language models (LLMs) perform well on synthetic "needle-in-a-haystack" (NIAH) benchmarks, but such tests overlook how noisy contexts arise from biased retrieval and...
0
1
2
๐จReleasing ๐ฅ๐๐ ๐ผ๐๐ฒ๐ฟ ๐ง๐ฎ๐ฏ๐น๐ฒ๐: ๐๐ถ๐ฒ๐ฟ๐ฎ๐ฟ๐ฐ๐ต๐ถ๐ฐ๐ฎ๐น ๐ ๐ฒ๐บ๐ผ๐ฟ๐ ๐๐ป๐ฑ๐ฒ๐
, ๐ ๐๐น๐๐ถ-๐ฆ๐๐ฎ๐ด๐ฒ ๐ฅ๐ฒ๐๐ฟ๐ถ๐ฒ๐๐ฎ๐น, ๐ฎ๐ป๐ฑ ๐๐ฒ๐ป๐ฐ๐ต๐บ๐ฎ๐ฟ๐ธ๐ถ๐ป๐ด. ๐ We push RAG to Multi-tables! ๐Code: https://t.co/kppOT6uXBC ๐Paper: https://t.co/KclevgAf6z
1
4
8
RAG over Tables ๐ซ
๐จReleasing ๐ฅ๐๐ ๐ผ๐๐ฒ๐ฟ ๐ง๐ฎ๐ฏ๐น๐ฒ๐: ๐๐ถ๐ฒ๐ฟ๐ฎ๐ฟ๐ฐ๐ต๐ถ๐ฐ๐ฎ๐น ๐ ๐ฒ๐บ๐ผ๐ฟ๐ ๐๐ป๐ฑ๐ฒ๐
, ๐ ๐๐น๐๐ถ-๐ฆ๐๐ฎ๐ด๐ฒ ๐ฅ๐ฒ๐๐ฟ๐ถ๐ฒ๐๐ฎ๐น, ๐ฎ๐ป๐ฑ ๐๐ฒ๐ป๐ฐ๐ต๐บ๐ฎ๐ฟ๐ธ๐ถ๐ป๐ด. ๐ We push RAG to Multi-tables! ๐Code: https://t.co/kppOT6uXBC ๐Paper: https://t.co/KclevgAf6z
0
0
3
Haystack Engineering: Context Engineering for Heterogeneous and Agentic Long-Context Evaluation @Mufei_Li et al. present a benchmark built on Wikipedia's hyperlink network. ๐ https://t.co/l0fpCNAbx6 ๐จ๐ฝโ๐ป https://t.co/iJZWY1gQcC
github.com
Haystack Engineering: Context Engineering for Heterogeneous and Agentic Long-Context Evaluation - Graph-COM/HaystackCraft
1
4
10
Very inspiring work! Congrats!
๐ Flow Matching Meets Biology and Life Science: A Survey Flow matching is emerging as a powerful generative paradigm. We comprehensively review its foundations and applications across biology & life science๐งฌ ๐Paper: https://t.co/ynsegKOgXz ๐ปResource:
0
0
0
โผ๏ธWe propose to learn temporal positional encodings for spatial-temporal graphs, L-STEP, in our #ICML2025 paper. Even simple MLPs can achieve leading performance in various temporal link prediction settings! ๐ Paper:ย https://t.co/ioapcELCD5 ๐ปย Code:ย https://t.co/sQWMlBZ1N6
0
2
11
Interesting to see the analysis between Inference scaling and LLM safety!
๐ฒ Not only reasoning?! Inference scaling can now boost LLM safety! ๐ Introducing Saffron-1: - Reduces attack success rate from 66% to 17.5% - Uses only 59.7 TFLOP compute - Counters latest jailbreak attacks - No model finetuning On the AI2 Refusals benchmark. ๐ Paper:
0
0
0
Identify Invariant Subgraphs for Spatial-Temporal OOD Problems ๐๐๐
Solving Out-of-Distribution problem of Spatial-Temporal Graphsโ Check out our #AISTATS2025 paper, our method distinguishes invariant components during training and makes link predictions robust to distribution shifts Paper: https://t.co/6ZPxK4TZHI
0
0
10
We'll present 4 papers and 1 keynote talk at #ICLR2025. Prof. Jingrui He and Prof. Hanghang Tong will be at the conference. Let's connect! โ๏ธ
0
3
8
๐ฅvarious graph types and various evaluation metrics
๐ฌGraph Self-Supervised Learning Toolkit ๐ฅWe release PyG-SSL, offering a unified framework of 10+ self-supervised choices to pretrain your graph foundation models. ๐Paper: https://t.co/fwlWTsmquK ๐ปCode: https://t.co/oNaz18zCht Have fun!
0
0
3
Beyond numerical-only time series models, letโs see how language can help time series forecasting and imputation๐คน
๐ Your time-series-paired texts are secretly a time series! ๐ Real-world time series (stock prices) and texts (financial reports) share similar periodicity and spectrum, unlocking seamless multimodal learning using existing TS models. ๐ฌ Read more: https://t.co/YYAH4J1Wne
0
0
1
๐ก How can we describe a graph to LLMs ? ๐ง For example, G(n, p) uses number of nodes and connection probability to describe a graph. ๐ Please check out our survey, What Do LLMs Need to Understand Graphs: A Survey of Parametric Representation of Graphs https://t.co/6bNIAwH5nG
0
6
12
Excited to know our #ICLR2025 paper is selected as Spotlight ๐๐๐ We studied how to generate large-scale temporal heterogeneous graphs balancing privacy, utility, and efficiency We also found diffusion model may not be the solution to answer them all https://t.co/yiAaG71pvp
openreview.net
Nowadays, temporal heterogeneous graphs attract much research and industrial attention for building the next-generation Relational Deep Learning models and applications, due to their informative...
0
2
13
This thursday, Feb 13th, 11am ET, at the Reading group, @kthrn_uiuc will present: Temporal Graph Neural Tangent Kernel with Graphon-Guaranteed (NeurIPS 2024). Looking forward to seeing you there! ๐๐ Zoom link on website https://t.co/oKBW6mUhNj
openreview.net
_Graph Neural Tangent Kernel_ (GNTK) fuses graph neural networks and graph kernels, simplifies the process of graph representation learning, interprets the training dynamics of graph neural...
Can we obtain temporal graph neural representation without neural networkโ Check out our #NeurIPS2024 paper, Temporal Graph Neural Tangent Kernel with Graphon-Guaranteed ๐ Paper: https://t.co/3u0bobtWJ9 Code: https://t.co/JYrCtiUjxm
0
3
8
2 papers accepted by #ICLR2025, 1 paper by #AISTATS2025, and 1 paper by #KDD2025, congrats all collaborators ๐ TL;DR 1๏ธโฃ How to generate temporal heterogeneous graphs 2๏ธโฃ How to tokenize a graph 3๏ธโฃ How to infer invariant subgraphs 4๏ธโฃ How to compress an evolving knowledge graph
0
3
13
๐ง When deepening is unavoidable, how to avoid oversmoothing of deeper GNNs? ๐ก Topology-guided contrastive learning can help ๐ข Check our new #TMLR paper, DrGNN: Deep Residual Graph Neural Network with Contrastive Learning Link:
0
0
1
๐ง How deep a GNN should be? ๐ก Number of layers is related the diameter of the input graph ๐ง When deepening is unavoidable, how to avoid dimensional collapse of deeper GNNs? ๐ก Topology-guided residual connection of neural layers can help
1
3
8
Can we obtain temporal graph neural representation without neural networkโ Check out our #NeurIPS2024 paper, Temporal Graph Neural Tangent Kernel with Graphon-Guaranteed ๐ Paper: https://t.co/3u0bobtWJ9 Code: https://t.co/JYrCtiUjxm
1
2
10
We will present 6 papers at #NeurIPS2024 covering active learning, graph learning, time series and trustworthy ML. Come and have a chat during the poster sessions!๐
0
4
6