Lingwei Gu

CS PhD @UWaterloo | Co-organizer of TREC RAG 2026 | Interests: NLP/IR
Previously @ Doordash

[Nour Jedidi @nour_jedidi]

The official version of our paper “Don’t ‘Overthink’ Pointwise Reranking: Is Reasoning Truly Necessary?” is out! 🎉

📄 https://t.co/wPioDrMACN

See the 🧵 below for details! 👇

[Lingwei Gu @Lingwei_Gu]

https://t.co/m0dBWrAcyR

[Rikiya Takehi @rikiyatakehi]

We asked GPT-6 Astra to edit Qwen3.5-0.8B weights to improve math.
It directly changed only 0.016% of weights, fixing 0.8B's reasoning loops & improving math up to 6x and even instruction following 3x, without gradients or training.

We call this Weight-Space Programming: frontier LLMs can now natively read/write neural network weights! 🧵
Blog: https://t.co/bOWCcFhCVq

[Lingwei Gu @Lingwei_Gu]

Qwen3.5-0.8B simply with thinking off can achieve 44.9%!

tried reproducing this on a single H100 with their own eval code. the edit checks out: 51.7% on MATH-500 (they got 48.1%)

but… the same base model with thinking just turned off, zero edits, already gets 44.9% 👀

[Lingwei Gu @Lingwei_Gu]

still cool that an LLM found one direction in 2 MLP layers that fixes the looping, and it’s +7 pts over thinking-off. just not “6x better at math” yet

[Lingwei Gu @Lingwei_Gu]

and thinking off isn’t some trick, the model card literally says the 0.8B runs in non-thinking mode by default

[Sahel Sharifymoghaddam @Sahel_Sharify]

Love BrowseComp-Plus? Meet 𝐂𝐌𝐀𝐒𝐒! 🚀
https://t.co/AJY898CbwZ

🔥Your next challenge for agentic search: 423 question–answer pairs spanning BrowseComp-Plus, DeepSearchQA, Total Recall QA, and HotpotQA—all projected onto ClimbMix, a realistic corpus of over 550M documents.

[Nandan Thakur @beirmug]

Recently revamped the FreshStack website and leaderboard! 🚀

Added newer models, including recent releases: Cohere Embed v5, Google's EmbeddingGemma2 & Perplexity late-interaction models.

Use the benchmark to pick robust retrieval models and findings! 👀

[Jimmy Lin @lintool]

tl;dr - the team @Sahel_Sharify @Lingwei_Gu @nour_jedidi speedran nanochat by @karpathy using ClimbMix by @shizhediao et al. and then SFT'ed on relevance data from @beirmug @crystina_z to arrive at a competitive pointwise reranker: https://t.co/QQgZ2rk4my

[Jimmy Lin @lintool]

What are we up to? 🏡 Project Greenhouse: Progress Toward Fully Open and Sovereign Agentic Search w/ @Sahel_Sharify @Lingwei_Gu @nour_jedidi https://t.co/QQgZ2rjwx0

[Nour Jedidi @nour_jedidi]

Excited to share that our paper was accepted to EMNLP Findings 🎉

We ask a simple question: Does reasoning improve pointwise rerankers?

The answer? It can actually hurt... even for Rank1, a reranker trained to reason-then-answer 😱

📄 https://t.co/ZyL7fYghDL

[Nour Jedidi @nour_jedidi]

To make an LLM a strong dense retriever, you often need training…

But what if you only have a few-examples 🥩🥦 and no training budget?

Just add RICE 🍚, a simple way to turn your LLM into an effective dense retriever that requires zero training 👨‍🍳

📄 https://t.co/A1dQGj4ShG

[Sahel Sharifymoghaddam @Sahel_Sharify]

1/ 🧵
BrowseComp-Plus (ACL 2026) set the standard for disentangled agentic search eval...

🙌 𝐁𝐫𝐨𝐰𝐬𝐞𝐂𝐨𝐦𝐩-𝐏𝐥𝐮𝐬_𝐂𝐌 makes it even better!

🚀 Same BCP questions grounded on 𝐂𝐥𝐢𝐦𝐛𝐌𝐢𝐱-𝟒𝟎𝟎𝐁 with 𝟓𝟓𝟑𝐌 𝐝𝐨𝐜𝐬.
💻 https://t.co/vf0Q5VcfGl

[TREC RAG @ 2026 @TREC_RAG]

And with that, all TREC RAG 2026 submissions are in 🎉

We had an incredible turnout this year, with 30 teams submitting more than 80 RAG runs and 110 retrieval runs!

A huge thank you to everyone who participated! We can’t wait to start evaluating your submissions 🚀

[Lingwei Gu @Lingwei_Gu]

At SIGIR this week. If anyone wants to grab a coffee or chat about RAG and LLMs let me know!

[Jimmy Lin @lintool]

It's now officially 👨‍🎓 DR @xueguang_ma - who yesterday successfully defended his Ph.D. 🎉 Congrats!

[Pouya Sadeghi @Ipouyall]

🏆 Thrilled that our ACL 2026 paper “To Lie or Not to Lie?” was selected as an ACL 2026 SAC Highlight! 🎉

The work asks a simple but urgent question: are LLMs equally safe around the world? 🧵👇

@aclmeeting #ACL2026NLP #NLProc

[TREC RAG @ 2026 @TREC_RAG]

🚨Official test topics for TREC RAG 2026 have been released! 🚨

🔎 Find them here: https://t.co/WBQ66K73Ex

Time to see what your agents can do 👀

[Nour Jedidi @nour_jedidi]

Nuggets, rubrics, side-by-side comparisons, citation support… all useful ways to evaluate RAG, but these pieces are generally scattered across different tools and workflows

Introducing RAGDoll 🐈, an end-to-end toolkit for evaluating long-form RAG answers 🧵

[TREC RAG @ 2026 @TREC_RAG]

Disentangling retrieval’s role in RAG requires tighter NLP ↔ IR alignment - that’s why we’re using @nvidia's ClimbMix 📚

But we don’t know which questions LLMs trained on ClimbMix "know"...and which they don’t

Now we have NanoKnow projections for NQ + SQuAD on ClimbMix 🔥

[Manveer Singh Tamber @ManveerTamber]

Agents can use tools to gather information, remember context, and act on your behalf.

That makes them useful. It also makes them dangerous. Agents can leak information they shouldn’t.

Introducing PrivacyAlign! PrivacyAlign uses human-annotation-grounded training and evaluation to cut privacy leaks by up to half and make automated privacy evaluation for agents more reliable.

Project Page: https://t.co/4gzPESUeO9

🧵(1/10)