Guy Bar-Shalom

PhD student @TechnionLive
Trying to better understand LLMs
Ex Research Intern @Verily | Ex Research Intern @Meta
Working with @HaggaiMaron and @el_yaniv

[Guy Bar-Shalom @GuyBarSh]

New blogpost out 📃

"Detecting LLM Misbehaviors from the Inside Out with Deep Learning on Structured Data" (https://t.co/tdcTr3waPZ) [1/8]

[Yam Eitan @ytn_ym]

1/

How much can you compress an LLM’s KV cache?

tl;dr it depends on how you train your model.

Many strong context compaction methods, such as Cartridges and attention matching, operate post-hoc: given a fixed model and a context, they try to compress the resulting KV cache.

@yoav_gelberg and I ask the complementary question:
can we train the model to produce KV representations that are easier to compress?

In other words: keep the compression method fixed, and change the representations it sees.

[Guy Bar-Shalom @GuyBarSh]

- "Beyond Next Token Probabilities: Learnable, Fast Detection of Hallucinations and Data Contamination on LLM Output Distributions", AAAI 2026 (https://t.co/ixdfiyqNqO) [8/8]

[Guy Bar-Shalom @GuyBarSh]

New blogpost out 📃

"Detecting LLM Misbehaviors from the Inside Out with Deep Learning on Structured Data" (https://t.co/tdcTr3waPZ) [1/8]

[Guy Bar-Shalom @GuyBarSh]

- "Neural Message-Passing on Attention Graphs for Hallucination Detection", ICLR 2026 (https://t.co/NotDP8R8bg) [7/8]

[Guy Bar-Shalom @GuyBarSh]

Check out our new ICLR 2026 paper - we explore hallucination detection through graph learning. Take a look!

[Fabrizio Frasca @ffabffrasca]

🧵"Neural Message Passing on Attention Graphs for Hallucination Detection" at #ICLR2026 !

🕸️We apply GNNs on the structured data LLMs produce as they generate text (e.g. attentions) to predict their errors.

📄 https://t.co/IQEyA7zaht
🤝 @GuyBarSh (co-1st) @YftahZ @HaggaiMaron

[Guy Bar-Shalom @GuyBarSh]

These works were joint efforts with a group of amazing collaborators: @ffabffrasca @HaggaiMaron @YftahZ @ytn_ym @yaniv_galron @yoav_gelberg @itayevron @mayabechlerspei @ido_guy Ami Tavory Moshe Eliasof Ran Elbaz

[Guy Bar-Shalom @GuyBarSh]

Happy to share my new #ICLR2026 papers !

[Guy Bar-Shalom @GuyBarSh]

📌 [4/4] On the Expressive Power of GNN Derivatives

We study how using gradients of GNNs can increase their expressive power, providing a principled way to go beyond standard message passing.
https://t.co/XwyjvdhVNw

[Haggai Maron @HaggaiMaron]

📄 Beyond Token Probes: Hallucination Detection via Activation Tensors with ACT-ViT
w/ @GuyBarSh , @ffabffrasca, Yaniv Galron, @YftahZ
https://t.co/D9M59DcbzV

[Fabrizio Frasca @ffabffrasca]

@GuyBarSh and I will be presenting the poster today, stop by 🤗

📍 Fri, Dec 5 • 4:30–7:30 PM PST • Exhibit Hall C,D,E # 4000

[Omer Belhasin @omerbelhasin]

🤔 Can discrete diffusion models actually outperform standard classifiers?

We show that it can!
📄 https://t.co/TwQx7iP17o
💻 https://t.co/TjPeGWcIED
🌐 https://t.co/ga3YOazPog

[Guy Bar-Shalom @GuyBarSh]

[7/7] Code:

https://t.co/pLAWJRwWfL

[Guy Bar-Shalom @GuyBarSh]

[1/7] New paper: "Beyond Token Probes: Hallucination Detection via Activation Tensors with ACT-ViT" #NeurIPS2025 [https://t.co/8aD1n8ejyp]

Joint work with: @ffabffrasca (co-first), @yaniv_galron, @YftahZ , @HaggaiMaron

[Guy Bar-Shalom @GuyBarSh]

[6/7] Results (over 15 LLM/dataset combinations):

• Consistently outperforms classic probes
• Zero-shot generalization to new datasets
• Fast adaptation to unseen LLMs by tuning only their new corresponding adapter