Amit Roth

CS PhD Candidate @TAU

[Amit Roth @Amit_Roth]

We'll be presenting "Hack-Verifiable Environments" at NeurIPS 2026 in Sydney! 🎉
Come talk to us about reward-hacking evaluation and our recent work: "Hack-Verifiable Terminal Bench".

[Amit Roth @Amit_Roth]

🚀New Paper🚀 Hack-Verifiable Environments: Towards Evaluating Reward Hacking at Scale

AI agents are increasingly deployed across critical environments. How do we know they're solving tasks the intended way, and not just appearing successful by hacking the reward?

Details👇🏽

[Amit Roth @Amit_Roth]

🚀New Paper🚀 Hack-Verifiable Environments: Towards Evaluating Reward Hacking at Scale

AI agents are increasingly deployed across critical environments. How do we know they're solving tasks the intended way, and not just appearing successful by hacking the reward?

Details👇🏽

[Amit Roth @Amit_Roth]

@longphann_ Love your work on Cheatbench! It tackles a lot of the issues we ran into :)

[Amit Roth @Amit_Roth]

@ValsAI We made an analysis for such behaviors and published the environments and traces, you can see our blog here -
https://t.co/lIYXkMYRq0

[Amit Roth @Amit_Roth]

Thanks for @neversupervised and @EfroniYonathan for the joint work!

Read our blog for all details, resources and traces: https://t.co/lIYXkMYRq0

[Amit Roth @Amit_Roth]

🖥️ Hack Verifiable Terminal Bench is out! 🎉

We planted honeypots inside real terminal tasks, so every hack by frontier agents is caught the moment it happens. A reliable way to measure reward hacking.
Fully open source: 89 environments + 2,225 traces.
https://t.co/lIYXkMYRq0
👇

[Amit Roth @Amit_Roth]

We can now evaluate reward hacking in a reliable way, without LLM judges. The story an agent tells can be spotless while the filesystem tells the truth, the honeypots catch every hack deterministically, the moment it happens.

[Talia Sternberg @TaliaSter]

Do speech language models secretly "think" in text?

In our new paper, we show that interleaved Speech LMs enter an implicit transcription stage: spoken words become text-decodable in middle layers, without ASR supervision.

Feature, shortcut, or limitation?

[Amit Roth @Amit_Roth]

@GallilMaimon @adiyossLC Congratulations, Gallil!
It was great working with you and learning from you.

[Hadar Davidson @HadarDavidson]

Excited to share Colored Noise Sampling (CNS)!🎉
Instead of injecting white noise, our SDE sampler exploits the inherent spectral bias of diffusion models. We dynamically color the injected noise to focus on frequencies where details are missing, substantially improving FID.🧵1/9

[Amit Roth @Amit_Roth]

📄 Paper - https://t.co/WmZR0CwyKX
🌐 Project Page - https://t.co/2MFkUyzG3b
�� Code - https://t.co/n8mbg355gR

With @Ankur_Samanta_ @MatanHalevy @YoavLevine @EfroniYonathan

[Amit Roth @Amit_Roth]

🚀New Paper🚀 Hack-Verifiable Environments: Towards Evaluating Reward Hacking at Scale

AI agents are increasingly deployed across critical environments. How do we know they're solving tasks the intended way, and not just appearing successful by hacking the reward?

Details👇🏽

[Amit Roth @Amit_Roth]

Using this framework, we built Hack-Verifiable TextArena: a suite of text-based environments with built-in reward hacking opportunities, designed to evaluate agent behavior reliably and at scale. 🛠️🤖💻

[Iddo Yosha @iddoyosha]

StressTest is heading to #ACL2026! ✌️

Updates include:
☕️ StressPresso, new benchmark for stress tasks generalization.
🧠 More models and more complex reasoning tasks!

Thanks to the GREAT collaborators @GallilMaimon @adiyossLC !

Project and paper: 🔗 https://t.co/pA7FAtaHOJ

[Gallil Maimon @GallilMaimon]

🚨New paper🚨 Self-Execution Simulation Improves Coding LLMs

Current reasoning LLMs <think> before giving a code solution.

We post-train them to explicitly simulate test execution, in order to verify and fix their own code leading to additional gains!

Details 👇🏽🧵

[Talia Sternberg @TaliaSter]

1/12 What if monkeys could have a language model too? 🐒💬

Our EMNLP 2025 paper introduces GmSLM — a Generative Marmoset Spoken Language Model.

A pipeline for modeling, generating, and evaluating vocal behavior, trained on marmoset vocalizations. 👇

[Guy Yariv @guy_yariv]

We present DyPE, a framework for ultra high resolution image generation.
DyPE adjusts positional embeddings to evolve dynamically with the spectral progression of diffusion.
This lets pre-trained DiTs create images with 16M+ pixels without retraining or extra inference cost.
🧵👇