Ali Modarressi

PhD student, NLP Researcher @ CIS LMU | Prev. Intern @AdobeResearch

[Z.Taghavi @zeinabTaghavi]

Excited to be at #ICML2026 in Seoul! 🇰🇷

I’ll be presenting our work With Argus Eyes: Assessing Retrieval Gaps via Uncertainty Scoring to Detect and Remedy Retrieval Blind Spots.

How can we know when a retriever is likely to fail — before it misses the information we need?

In this work, we explore how to uncover retrieval blind spots and use uncertainty signals to guide targeted improvements in retrieval systems.

If you are attending ICML, I would be happy to meet, discuss retrieval, reasoning, our work, and hear your thoughts. Come by our poster! ✨

📄 Paper: With Argus Eyes: Assessing Retrieval Gaps via Uncertainty Scoring to Detect and Remedy Retrieval Blind Spots

📍 Location: HALL A – No. 1502
🗓️ Time: Thursday, July 9 • 10:30 AM – 12:15 PM KST

🔗 Paper & Code:

arXiv: https://t.co/whVByWCfQI
GitHub: https://t.co/dOrvyDddBI

#ICML #InformationRetrieval #MachineLearning #NLP #LLMs

[Ali Modarressi @AModarressi]

Check out the original paper and code here:
🔗 arXiv: https://t.co/ae7HWZdIr9
💻 GitHub: https://t.co/Rq0XPm2pD3
[3/3]

[Ali Modarressi @AModarressi]

If you're at #ICML2026 🇰🇷 next week, don't miss our latest work on remedying retriever blind spots! 🔎
@zeinabTaghavi is presenting our paper at the conference. I unfortunately won't be there to join, but you should definitely go check out the poster, and chat with her! 👇 [1/3]

[Ali Modarressi @AModarressi]

📄 With Argus Eyes: Assessing Retrieval Gaps via Uncertainty Scoring to Detect and Remedy Retrieval Blind Spots
📍 Location: HALL A – No. 1502
🗓️ Time: Thu, Jul 9 • 10:30 AM – 12:15 PM KST
[2/3]
https://t.co/fsgcLkCodR

[Z.Taghavi @zeinabTaghavi]

🎉 Excited to share our new paper accepted at ICML 2026!

🔍 Did you know retrievers can have their own blind spots-where certain documents are consistently hard to retrieve?

📉 Retrieval failures are not random; some documents are consistently hard to retrieve.

📊 We introduce RPS to quantify retrievability and show it can be predicted from embeddings before indexing.

🛠️ ARGUS detects these blind spots and mitigates them with targeted augmentation.

📈 We observe consistent gains in retrieval performance (+3.4 nDCG@5, +4.5 nDCG@10 on average).

🤝 Grateful to work with an amazing team: @AModarressi , @HinrichSchuetze , and @amarfurt, at @CisLmu and @hslu.

💻 Code, supplementary materials, and reimplementations: https://t.co/knvjqY1uaO
📄 Paper: https://t.co/6Y68KIe3Dn

#ICML2026
#ICML
#Trustworthy
#RAG
#InformationRetrieval

[Ali Modarressi @AModarressi]

Check out our new collab with @ZeinabTaghavi, accepted at #ICML2026!🎉
Dense retrievers have blind spots (see: Collapse of Dense Retrievers & ImpliRet; both worth a look 👀). ARGUS detects them from embeddings before indexing and mitigates them with targeted augmentation.

[Z.Taghavi @zeinabTaghavi]

🎉 Excited to share our new paper accepted at ICML 2026!

🔍 Did you know retrievers can have their own blind spots-where certain documents are consistently hard to retrieve?

📉 Retrieval failures are not random; some documents are consistently hard to retrieve.

📊 We introduce RPS to quantify retrievability and show it can be predicted from embeddings before indexing.

🛠️ ARGUS detects these blind spots and mitigates them with targeted augmentation.

📈 We observe consistent gains in retrieval performance (+3.4 nDCG@5, +4.5 nDCG@10 on average).

🤝 Grateful to work with an amazing team: @AModarressi , @HinrichSchuetze , and @amarfurt, at @CisLmu and @hslu.

💻 Code, supplementary materials, and reimplementations: https://t.co/knvjqY1uaO
📄 Paper: https://t.co/6Y68KIe3Dn

#ICML2026
#ICML
#Trustworthy
#RAG
#InformationRetrieval

[Ali Modarressi @AModarressi]

Couldn't make it to ICLR in person this year, but our work is there! 🎉 Stop by Poster #603 tomorrow if you're around 🙌

Really proud of this collab w. @mohsen_fayyaz : steering MoE model behavior by flipping expert activations at test time.

#ICLR2026

[Mohsen Fayyaz @mohsen_fayyaz]

📌 SteerMoE has been accepted to ICLR 2026 and will be presented at Poster #603, Poster Session 2, Pavilion 3, Thu 23 Apr, 3:15–5:45 p.m.
🙏 I’m unable to attend in person, so many thanks to @LucasBandarkar for presenting on our behalf. Stop by if you’re around.

[Yuki @y_m_asano]

And a nice interlude that had our audience participate and think critically about AI for AI. And finished with work from @AModarressi et al. on continual learning for LLMs by having it read and write to a memory.

[Ali Modarressi @AModarressi]

While I’m attending #EACL2026 virtually, be sure to catch Pedro's presentation at the 11:30–13:00 session this Wednesday in S. Le Riad! He’ll be discussing our recent collaboration:
📄 "Persistent Personas? Role-Playing, Instruction Following, and Safety in Extended Interactions"

[CIS, LMU Munich @CisLmu]

Persistent Personas? Role-Playing, Instruction Following, and Safety in Extended Interactions
[Oral]: Wed. 25 Mar - S. Le Riad @ 11:30-13:00
Pedro Henrique Luz de Araujo, @MicHedderich, @AModarressi, @HinrichSchuetze, Benjamin Roth
https://t.co/4rvLZJVzaT

[CIS, LMU Munich @CisLmu]

Going to Rabat for #EACL2026? So are we! 🇲🇦
We are bringing a packed schedule of papers, talks, and workshops.
Check out our lineup below and come say hi! 👋 🧵
#NLProc @EACL2026

[NetBlocks @netblocks]

⚠️ Update: #Iran has now been offline for 96 hours, limiting reporting and accountability over civilian deaths as Iranians protest and demand change; fixed-line internet, mobile data and calls are disabled, while other communication means are also increasingly being targeted ⌛️

[Elham @ellay_ve]

Iran is offline. No internet, no calls, no witnesses. Don’t let silence win. Be their voice.

#DigitalBlackoutIran

[Leonie Weissweiler @LAWeissweiler]

🧑‍🔬I’m recruiting PhD students in Natural Language Processing @UniLeipzig Computer Science, together with @Sca_DS!

Topics include, but aren’t limited to:

🔎Linguistic Interpretability
🌍Multilingual Evaluation
📖Computational Typology

Please share!

#NLProc #NLP

[Leonie Weissweiler @LAWeissweiler]

🥳Life Update!

I’m thrilled to share that I’ll be starting as assistant professor for Natural Language Processing @UniLeipzig in April! I’m deeply grateful to everyone who supported me on this journey.

I will be recruiting PhD students with @Sca_DS, stay tuned for the details!

[Ali Modarressi @AModarressi]

Happening now @ Hall C!
No. 148

[Ali Modarressi @AModarressi]

Excited to be here in Suzhou for #EMNLP2025!
I’ll be presenting “ImpliRet”, check out our poster on Friday Nov. 7th at 14:00.
If you’re into long-context, IR, or just want to chat, come *Pay Ali* a visit 😁

[Ali Modarressi @AModarressi]

Excited to be here in Suzhou for #EMNLP2025!
I’ll be presenting “ImpliRet”, check out our poster on Friday Nov. 7th at 14:00.
If you’re into long-context, IR, or just want to chat, come *Pay Ali* a visit 😁

[Z.Taghavi @zeinabTaghavi]

🚨Introducing ImpliRet 🚨
Most reasoning-heavy IR benchmarks focus on complex queries. But what if the reasoning is needed on the document side?🤔 ​​Our new benchmark tests retrieval when facts are implicitly stated and it exposes a tough challenge—best nDCG@10 is only 14.91%

[Ali Modarressi @AModarressi]

Thanks for the invite and hosting me! 🙏🏻

[Hosein Mohebbi @hmohebbi75]

Was great having @AModarressi visit us at Tilburg University, to share his research on evaluating LLMs in long-context settings!

[Ali Modarressi @AModarressi]

💥New paper 💥
MoE experts shape not only domains but also behaviors like safety and faithfulness. Toggling them at inference lets us steer outputs toward safer or more faithful behavior, or bypass guardrails.
Excited to share my collab on @mohsen_fayyaz’s work at @AdobeResearch.

[Mohsen Fayyaz @mohsen_fayyaz]

🚨 You can bypass ALL safety guardrails of GPT-OSS-120B 🚨❗🤯

How? By detecting behavior-associated experts and switching them on/off.

📄 Steering MoE LLMs via Expert (De)Activation
🔗 https://t.co/U2YRyXon4H
🧵👇

[Z.Taghavi @zeinabTaghavi]

🎉 Excited to share that ImpliRet is accepted to the Main Conference of #EMNLP2025!

We introduce a benchmark for reasoning-intensive IR, where retrievers must surface implicit facts hidden in documents across 3 reasoning categories.

📄 https://t.co/iPoW0Tn9Oe

[Ali Modarressi @AModarressi]

@yanaiela Congrats Yanai! 🎉

[Ali Modarressi @AModarressi]

📢 Paper alert: ImpliRet!
In ImpliRet, we reveal how hard it is for retrievers and even LLMs in RAG/long-context setups to reason over implicit facts hidden in documents.

Proud to share this work with @zeinabTaghavi, alongside Yunpu Ma and @HinrichSchuetze.

[Z.Taghavi @zeinabTaghavi]

🚨Introducing ImpliRet 🚨
Most reasoning-heavy IR benchmarks focus on complex queries. But what if the reasoning is needed on the document side?🤔 ​​Our new benchmark tests retrieval when facts are implicitly stated and it exposes a tough challenge—best nDCG@10 is only 14.91%