Bar Alon (@Bar_A_94) profile picture

AI researcher @ stealth PhD candidate @ Tel Aviv University

Joined October 2024
Photo by Bar Alon (@Bar_A_94)Photo by Bar Alon (@Bar_A_94)Photo by Bar Alon (@Bar_A_94)Photo by Bar Alon (@Bar_A_94)Photo by Bar Alon (@Bar_A_94)Photo by Bar Alon (@Bar_A_94)
Working on generative models, I'm often asked about the risks of not being able to distinguish real from generated images. In our new work, we show how to watermark images just by choosing which samples to release, without modifying the model. And we *prove* it's distortion-free!
Can you watermark any generative model just by choosing what to release? Our black-box method is provably distortion-free and robust to content-preserving edits. With @OPatashnik and @OREAXEAX Paper → arxiv.org/abs/2611.23208 Project page → avivyossef.github.io/distort…
6
45
1,446
OpenReview status: Delivered. Read: Never.
24
📄📄📄📄 Accepted to EACL - Main First-to-Repeat Sampling (F2R): Adaptive Majority Voting at a Fraction of the Cost More details & paper - Soon 🏛️🇬🇷
3
81
Wow. I know this feeling! When u have a new idea and it works great, but the baselines are not honest/reproducible 🥵
Most fun public comment in ICLR2027. Bro reproductivity is important☠️☠️☠️ #ICLR2027
1
116
1/7 🥥 New speculative decoding method! Introducing DCOCO: Contextual Correction of Parallel Drafts. Parallel drafters are fast, but lose context between predicted tokens. DCOCO restores it without giving up parallelism. Up to 6.6× faster than vanilla decoding. 🧵
30
70
7
708
56,154
Which model should you trust to keep building your app on @Base44, @Lovable, or @Replit as it gets more complex? Our latest paper benchmarks how coding models handle evolving apps with interdependent features. Check it out 👇
Even frontier models struggle when coding features depend on each other! Introducing EvoWeb, a new long-horizon coding benchmark: 32 apps × 10 interdependent requests. After every coding-agent update, a browser agent QA-tests the live app. 📄 arxiv.org/pdf/2610.16402
2
1
10
3,169
SwitchSD improves LLM speculative decoding by using a lightweight signal from the LLM’s internal representations to predict when to copy from context vs. when to use a drafter 🚀 A fun intersection of interpretability and acceleration! Congrats to @roy_eisenstadt and the team! 🎊
1/10 Accepted to #NeurIPS2026! 🎉 Can an LLM tell us when to copy instead of running a neural drafter? We introduce SwitchSD: speculative decoding that uses the target model’s hidden states to switch between neural drafting and context copying. Here’s the story.
2
1
21
1,141
1/10 Accepted to #NeurIPS2026! 🎉 Can an LLM tell us when to copy instead of running a neural drafter? We introduce SwitchSD: speculative decoding that uses the target model’s hidden states to switch between neural drafting and context copying. Here’s the story.
3
8
2
43
5,183
Infinite keepy-uppies ⚽️ More cool stuff in the replies 👇 Don’t ask me about sim-to-real transfer :)
9
4
3
62
8,725
1/ Text-to-video models generate beautiful videos, but they still struggle to follow complex prompts. Relations like left/right, towards/away, on top of, behind, or even multi-stage instructions are often generated incorrectly. In our new paper, accepted to SIGGRAPH Asia 2026, we introduce CVG, an inference-time guidance method that improves compositional video generation without retraining or architectural changes.
8
8
31
76,918
Looking forward to presenting this work at ACL 2026 as an oral presentation! 🗓️ Sunday, July 5 🎙️ Oral Session D: Special Theme – Explainability of NLP Models 1 🕘 9:30 AM 📍 Harbor G Come by if you’re attending ACL! happy to chat about the paper!
📄 🚨 You ask an LLM why it generated an answer. It gives you a clean, confident explanation. The problem is it’s often not the real reason. We measure this gap, then close it with Faithfulness Serum. Accepted to the #ACL2026 main conference 📄 arxiv.org/abs/2604.14325 1/6 🧵
1
2
10
422
Bar Alon retweeted
We hope everyone is excited for ACL2026@San Diego. Feel free to take a look at the published papers below! Main: 2026.aclweb.org/program/acce… Findings: 2026.aclweb.org/program/find… Demos: 2026.aclweb.org/program/demo… Industry track: 2026.aclweb.org/program/indu… #NLProc #ACL2026
10
2
52
8,914
Bar Alon retweeted
🎉 Happy to share our latest work: Bootstrap Your Generator: Unpaired Visual Editing with Flow Matching (accepted to #ICML2026)! TL;DR: We train image and video editing models without any paired data. No ground-truth edit data, no external reward models.
7
14
2
40
6,896
Do LLM explanations truly reflect their internal rationale? In our recent ACL paper, we introduce a framework to measure explanation reliability using counterfactuals, and propose Faithfulness Serum💉- a method that enables LLMs to produce faithful explanations! Details in 🧵👇
📄 🚨 You ask an LLM why it generated an answer. It gives you a clean, confident explanation. The problem is it’s often not the real reason. We measure this gap, then close it with Faithfulness Serum. Accepted to the #ACL2026 main conference 📄 arxiv.org/abs/2604.14325 1/6 🧵
1
7
478
📄 🚨 You ask an LLM why it generated an answer. It gives you a clean, confident explanation. The problem is it’s often not the real reason. We measure this gap, then close it with Faithfulness Serum. Accepted to the #ACL2026 main conference 📄 arxiv.org/abs/2604.14325 1/6 🧵
7
2
2
14
1,285
5/6 How faithful are closed-source models? How should the strength of the intervention be selected? Which layers should be injected? For answers to these questions, along with many more insightful ablations and details, see the paper 📄
1
52
4/6 🪛We fix this with Faithfulness Serum💉, an attribution-guided generation method based on attention intervention. No training. We compute token attributions and inject them into attention at the layers where the model commits to an answer. Now the explanation follows the hint
1
56