Joining #EMNLP2026 remotely due to visa issues, but Iโd love to connect!
Iโm presenting OpenToolsโbecause reliable AI agents need reliable tools.
๐
Oct 26, 12:30โ14:00 | Virtual Presentations 4
Excited to receive the Tinker Research Grant from @thinkymachines ๐! Looking forward to applying the funding toward advancing my research on challenging and impactful problems.
Today weโre announcing research and teaching grants for Tinker: credits for scholars and students to fine-tune and experiment with open-weight LLMs.
Read more and apply at: thinkingmachines.ai/blog/tinโฆ
โ๏ธ Heading to Suzhou ๐จ๐ณ for #EMNLP2025!
Iโll be presenting our paper (Main Conference), and also supporting the event as a Student Volunteer ๐.
Looking forward to connecting with everyone!
๐ Poster Section: Hall C | Wednesday, Nov 5 | 14:30โ16:00
@emnlpmeeting
Thrilled to share that โImproving Large Language Models Function Calling and Interpretability via Guided-Structured Templatesโ paper has been accepted to EMNLP 2025 (Main Conference)!๐
๐ Check it out on
arXiv: arxiv.org/abs/2509.18076
project page: hygiadang.com/publication/emโฆ
1/3
2/3
๐ What we studied:
We explored how well-structured templates/curricula can guide LLMs to improve function-calling capabilities, and proposed a data construction pipeline to generate fine-tuning datasets that further enhance performance in function-calling domains.
3/3
โจ Gratitude:
Work done during my internship at @amazon. Huge thanks to amazing mentors (Tianyi Liu, @cserxy ) and collaborators (@JingfengY , @jiang_haoming , @TaoYang2020 , Pei Chen, Helen Wang, Huasheng Li, Bing Yin) and to my advisor @Meng_CS for continuous support!
Hy Dang retweeted
Professor Edward H.Sargent a Canadian Scientist, who published many papers in top journals i.e., #science and #Nature. He wrote 10 tips on how to write papers. @PhDvibe @PhDVoice @Labiofy #PhDposition #phdlife #PhD #postdoc #chemtwitter #Science #AcademicChatter
Hy Dang retweeted
Nature's guide on how to write effective abstracts
ALT Nature's guide on the sections and structure of an effective scientific abstract.
Hy Dang retweeted
If you embed an entire document, you'll lose retrieval precision.
If you chunk a document, you'll lose contextual information between chunks.
These are some concerns when you're building long-context RAG applications.
But "Late chunking" may just be the sweet spot in the middle.
Late chunking:
โข a new approach introduced by @JinaAI_
โข helps preserve context across large documents
โข applies chunking later in the embedding process
Read more: weaviate.io/blog/late-chunkiโฆ
Hy Dang retweeted
RAG vs. Long-Context LLMs
I have yet to see a convincing paper or technical blog showing that long-context LLMs can or will replace RAG.
So far I've seen specific long-context applications where long-context LLMs thrive and current retrieval benchmarks are not convincing.
This new paper reports that longer-context LLMs suffer from diminished focus on relevant information, which is one of the primary issues that a RAG system addresses (i.e., uses more relevant information).
They propose an order-preserving RAG mechanism that improves performance on long-context question answering.
It's not perfect and in fact as retrieved chunks increases the quality of responses go up and then declines. But there is a sweet spot where it can achieve better quality with a lot fewer tokens than long-context LLMs.
I am getting better results for RAG plus the added benefit of efficiency as well. Things could change rapidly but still not counting out RAG for now.
Hy Dang retweeted
Here's why ColBERT embeddings are all the rage right now (at an intuitive level):
You probably already know that vector search is pretty cool.
โข It allows you to search for things semantically.
โข It's robust to synonyms.
But do you know what sucks about vector search?
It kind of lacks a certain degree of explainability.
The kind of explainability you'd, for example, have in keyword search:
There, you can see which keywords match between the query and the result.
ColBERT embeddings enable some explainability in vector search.
This is because
โข Conventional dense embeddings: pool the token-level embeddings to a single representation.
โข ColBERT embeddings: Keep the token-level embedding representation
We'll get to the technical details of ColBERT and the late interaction mechanism in another post, so stay tuned.
Hy Dang retweeted
Top RAG Papers of the Week
[1] Meta Knowledge for Retrieval Augmented Large Language Models
[2] RAGLAB: A Modular and Research-Oriented Unified Framework for Retrieval-Augmented Generation
[3] Graph Retrieval-Augmented Generation: A Survey
[4] CommunityKG-RAG: Leveraging Community Structures in Knowledge Graphs for Advanced Retrieval-Augmented Generation in Fact-Checking
[5] W-RAG: Weakly Supervised Dense Retrieval in RAG for Open-domain Question Answering
For paper summaries and links, check - linkedin.com/pulse/top-rag-pโฆ
For brief video tutorial overview of the papers - youtube.com/watch?v=jZ4n6l9hโฆ
#rag #toppapers #resarch #llms #generativeai
Hy Dang retweeted
This Python tool looks super useful to crawl websites and convert data into LLM-ready markdown or structured data.
I find myself doing this a lot and most of the time it is a tedious effort. Great to see a service that does data extraction catered for LLM-based pipelines.
github.com/mendableai/firecrโฆ
Hy Dang retweeted
Graph RAG Works Better Than Standard RAG
GraphRAG leverages structural information across entities to enable more precise and comprehensive retrieval, capturing relational knowledge and facilitating more accurate, context-aware responses.
This improves the accuracy of standard RAG systems. This survey paper (link in alt) is a great read.
ALT https://arxiv.org/pdf/2408.08921
Hy Dang retweeted
A useful collection of Prompt Engineering resource released from @AnthropicAI for building on Anthropic API.
Hy Dang retweeted
Your LLM is only as good as how robust your prompting method is.
Seems you can enhance the robustness of LLMs by "prompting out" irrelevant information from context. Think of it as a self-mitigation process that first identifies the irrelevant information and then filters it out.
Hy Dang retweeted
OpenAI provides a comprehensive guide on enhancing the accuracy of Large Language Models (LLMs), emphasizing methods to improve response correctness and consistency.
Rather than approaching LLM accuracy optimization as a straightforward, linear process that progresses from prompt engineering to retrieval-augmented generation (RAG) to fine-tuning, it's more beneficial to view it as a complex, multifaceted challenge requiring various strategies.
This approach can be conceptualized as a two-dimensional matrix:
1. The first dimension focuses on what information the LLM requires (context).
2. The second dimension addresses how the LLM should perform or behave (behavior).
This framework suggests a more nuanced and holistic approach to optimizing LLM performance, considering both the knowledge base and the operational aspects of the model.
๐ ๐๐ก๐๐ญ ๐ญ๐ก๐ ๐๐๐ ๐ง๐๐๐๐ฌ ๐ญ๐จ ๐ค๐ง๐จ๐ฐ: Optimize for context when
(i) the model lacks contextual knowledge because the information was not in the training dataset,
(ii) its knowledge is out of date, or
(iii) it requires knowledge of proprietary information.
This dimension maximizes ๐ซ๐๐ฌ๐ฉ๐จ๐ง๐ฌ๐ ๐๐๐๐ฎ๐ซ๐๐๐ฒ.
๐ ๐๐จ๐ฐ ๐ญ๐ก๐ ๐๐๐ ๐ง๐๐๐๐ฌ ๐ญ๐จ ๐๐๐ญ: Optimize for behavior when
(i) the model is producing inconsistent results with incorrect formatting,
(ii) tone or style of speech is not correct, or
(iii) reasoning is not being followed consistently.
This dimension maximizes ๐๐๐ก๐๐ฏ๐ข๐จ๐ซ ๐๐จ๐ง๐ฌ๐ข๐ฌ๐ญ๐๐ง๐๐ฒ.
With this mental model, we can decide whether to enrich the context and/or improve the consistency of model behavior.
The guide also includes numerous valuable practical tips, such as best practices for continual evaluation, stacking fine-tuning with RAG, balancing cost against accuracy, and many more.
