Farnoush Rezaei-Jafari

Research Intern @MSFTResearch | ML Ph.D. Candidate @TUBerlin and @bifoldberlin | Explainable AI, Interpretability, Efficient Machine Learning

[Farnoush Rezaei-Jafari @FarnoushRJ]

New Preprint! 🎉
w/ @EberleOliver @ashk__on and @NeelNanda5

Can we speed up finding important components in LLMs without losing faithfulness?
❌ Attribution patching: fast but noisy
❌ Integrated Gradients: more faithful but costly
✅ RelP: combines the best of both worlds 🧵

[Farnoush Rezaei-Jafari @FarnoushRJ]

@shpotes I am also at NeurIPS in Mexico City and I’d be happy to chat and connect.

[Farnoush Rezaei-Jafari @FarnoushRJ]

I am at #NeurIPS2025 in Mexico City!

I’ll present “RelP: Faithful & Efficient Circuit Discovery in Language Models via Relevance Patching” today at the ResponsibleFM workshop!

📌 Location: Don Alberto 1
⏰ Time: 5 p.m. – 7 p.m. CST

Don’t forget to come by our poster!

[Explainable AI Berlin @xai_berlin]

🚀 Visit our #NeurIPS posters @NeurIPSConf!
Meet and interact with our authors at all locations — San Diego, Mexico City, and Copenhagen.
Details in the thread.
👇👇👇

[Mohsen Fayyaz @MohsenFyz]

Do MLLMs really understand temporal dynamics in a video?
In our #NeuRIPS2025 paper we show that current Video-LLMs have critical limitations in temporal understanding.
To this end we proposed Stacked Temporal Attention in Vision Encoders: STAVE
https://t.co/yUvQXIrrvA 🧵👇1/8

[Mohsen Fayyaz @MohsenFyz]

📢 Call for Papers – @LSHVU @NeurIPSConf 2025
Submissions are now open for the 7th Workshop on Holistic Video Understanding: Toward Video Foundation Models
Join us to explore multimodal reasoning, scaling laws, and evaluation for generalist video AI.

https://t.co/i2KSW82HyR

[Farnoush Rezaei-Jafari @FarnoushRJ]

Excited to be selected as a Top Reviewer for NeurIPS 2025! 🎉

#neurips2025

[Farnoush Rezaei-Jafari @FarnoushRJ]

I’m excited to share that our paper has been accepted as a spotlight at the NeurIPS 2025 Mechanistic Interpretability Workshop! 🎉

[Farnoush Rezaei-Jafari @FarnoushRJ]

New Preprint! 🎉
w/ @EberleOliver @ashk__on and @NeelNanda5

Can we speed up finding important components in LLMs without losing faithfulness?
❌ Attribution patching: fast but noisy
❌ Integrated Gradients: more faithful but costly
✅ RelP: combines the best of both worlds 🧵

[Farnoush Rezaei-Jafari @FarnoushRJ]

@thebasepoint I appreciate you raising this. AttnLRP paper (https://t.co/ucU2vSme5w) proposes specific rules for handling attention layers. Our current RelP implementation supports only a subset of existing LRP rules, but it should be extended to include more within TransformerLens.

[Neel Nanda @NeelNanda5]

This was a lovely paper! Mech interp often uses attribution patching, gradient based approximations to principled casual interventions. But it's often inaccurate. Farnoush found that LRP, an XAI technique for better gradient based approximations, massively improves accuracy!

[Farnoush Rezaei-Jafari @FarnoushRJ]

Paper: https://t.co/k4W6paePXM
Code: https://t.co/gK7ncmwN4j

[Farnoush Rezaei-Jafari @FarnoushRJ]

New Preprint! 🎉
w/ @EberleOliver @ashk__on and @NeelNanda5

Can we speed up finding important components in LLMs without losing faithfulness?
❌ Attribution patching: fast but noisy
❌ Integrated Gradients: more faithful but costly
✅ RelP: combines the best of both worlds 🧵

[Farnoush Rezaei-Jafari @FarnoushRJ]

✔ ️ achieves comparable faithfulness to Integrated Gradients (IG) for sparse feature circuit discovery in Pythia-70M without IG’s extra computational cost.

[Farnoush Rezaei-Jafari @FarnoushRJ]

@siavashiar @NeurIPSConf Thanks! 🙂

[Farnoush Rezaei-Jafari @FarnoushRJ]

What an unforgettable experience at #NeurIPS2024 in beautiful Vancouver! 🎉 Grateful for the chance to present my paper, attend amazing talks, and connect with brilliant minds. The ideas, energy, and city vibes made it truly special. Until next time, @NeurIPSConf !

[Farnoush Rezaei-Jafari @FarnoushRJ]

Goodbye Vancouver! 👋

[Farnoush Rezaei-Jafari @FarnoushRJ]

🔗 Paper: https://t.co/PJCmojME5I
🔗 Code: https://t.co/bNZmBHJP6j
🔗 Poster: https://t.co/W8bP1arvGG

#XAI #ExplainableAI #explainability #Interpretability

[Farnoush Rezaei-Jafari @FarnoushRJ]

Join us tomorrow at our poster! 🌟

I’ll present our work, "MambaLRP: Explaining Selective State Space Sequence Models," at #NeurIPS2024.

📅 Thursday, December 12 🕟 4:30 PM PST 📍 East Exhibit Hall A-C, Poster #2804

Looking forward to seeing you there!