Idan Schwartz (@idansc) profile picture

Assistant Prof. BIU; Postdoc Tel-Aviv University; PhD Technion/UIUC; Chief Scientist https://nitter.cf/t.co/ecKXykKnYv; Interests: Deep Learning, Multi-modal

Tel Aviv, Israel
Joined June 2009
Photo by Idan Schwartz (@idansc)Photo by Idan Schwartz (@idansc)Photo by Idan Schwartz (@idansc)Photo by Idan Schwartz (@idansc)Photo by Idan Schwartz (@idansc)Photo by Idan Schwartz (@idansc)Photo by Idan Schwartz (@idansc)Photo by Idan Schwartz (@idansc)Photo by Idan Schwartz (@idansc)Photo by Idan Schwartz (@idansc)
One of the coolest features of word embeddings is to perform arithmetic. For example, woman + (king - man) = queen. Is it be possible to do the same visually? In our new #CVPR2022 paper, we show that in CLIP's embedding space, it is. Code: github.com/YoadTew/zero-shot… [1/3]
Zero-Shot Image-to-Text Generation for Visual-Semantic Arithmetic abs: arxiv.org/abs/2111.14447
7
47
2
284
Idan Schwartz retweeted
even more interesting work on understanding CoT controllability, hyped to see it!
New paper: We find reasoning models' CoT controllability to be much higher than current evals suggest. Models rarely follow user instructions about their CoT, but follow them far more often when written as self-instructions inside the CoT, which can be used to evade monitors 🧵
3
5
52
2,493
Idan Schwartz retweeted
Replying to @AdamJaber248
Thanks! Bootstrap works almost always: 98% for claude models and similar results for the rest of the models unless they confuse the CoT and the output. This was quite surprising to us as well, considering the close-to-zero success of user requests for most models. More details in the paper :)
1
3
186
We found an interesting phenomenon: starting a CoT with a self-instruction makes it significantly more controllable than giving the same instruction in the prompt. Interestingly, it also helps evade CoT monitors.
New paper: We find reasoning models' CoT controllability to be much higher than current evals suggest. Models rarely follow user instructions about their CoT, but follow them far more often when written as self-instructions inside the CoT, which can be used to evade monitors 🧵
3
12
428
The AI physical Olympics will be held in Tel Aviv 🇮🇱!! Details and registration below! Tel Aviv, May 18–19, 2027. 10,000 people in the arena. Forget scripted, pre-recorded demos where the robot succeeds on take 47. This one will be live! The P.AI Games—organized by the Israeli National AI Directorate | Prime Minister's Office and led by Guy Altagar—is putting embodied AI to the ultimate test: 10 real-world arenas. $1.35M in prize money. Live scoring in front of an arena crowd. Here are the events: → Pick up oily crystal glasses and pack a dishwasher in under 8 minutes without cracking one. → High-speed dynamic baton handoff between two moving robots across a 20-meter track. → Face off in real over-the-board chess against another robot, moving pieces and hitting the clock. → Balance a raw egg and a full tray of water up an incline ramp with zero spills. → Grab a 14-lb bowling ball and take down a full rack of pins. Humanoids, quadrupeds, wheeled bases, custom manipulators—every morphology is welcome. Whether you want to dominate a single niche challenge or take on all ten, this is where the bar gets set. No entry fee. Plus, participating teams receive dedicated financial support to make the trip happen. If you run a robotics startup, lead an OEM team, or build cutting-edge embodied models in academia, this is your arena. Drop a comment or DM me if your team wants to throw their hat in the ring—I'd love to see you on the floor. Details and registration: pai-games.com/
24
33
9
400
26,710
Idan Schwartz retweeted
Interesting finding about reasoning models and CoT. They find that reasoning models can control their chain of thought much better than earlier tests suggested. Earlier evaluations put the constraint in the user prompt and found very low controllability. Researchers put the same constraint inside the chain of thought as a self-instruction, and models followed it far more often. A generic reminder to follow the user's constraint also helped a lot. The self-instruction can be triggered from the user prompt alone, so no access to the reasoning channel is needed. Paper: arxiv.org/abs/2610.23257 Chat with Paper: academy.dair.ai/papers/reaso…
14
22
1
97
9,008
Idan Schwartz retweeted
1/4 We trained an 8B safety classifier on just 1,275 carefully annotated prompts. It beat GPT-5.4, Claude Sonnet 4.6, and leading open-weight guardrails trained on far more data, on average across 5 safety benchmarks. Quality > quantity.
2
13
2
27
1,789
Reasoning Models Can Control Their Chains of Thought When They Tell Themselves To Uriel Dolev, Idan Schwartz, Yoav Goldberg arxiv.org/abs/2610.23257 [𝚌𝚜.𝙰𝙸 𝚌𝚜.𝙲𝙻 𝚌𝚜.𝙻𝙶] 💬Code: github.com/udolev/Self-Instr…
4
3
251
Idan Schwartz retweeted
Introducing Generative Commerce. @daletlabs – at New York, NY
53
49
35
255
208,155
Idan Schwartz retweeted
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
Idan Schwartz retweeted
Robot policies can move but can't think. LLMs can think but can't move. So we connected them. Real robot: 16.7% → 97.3% Sim (LIBERO-PRO): 12.8% → 53.3%
23
53
9
303
96,714
Idan Schwartz retweeted
Nobody is reviewing your pull requests. AI made code cheap to write, and review the bottleneck. I re-imagined codebases as cities, where pending changes are activity, and contributors are the inhabitants. (see it below on openclaw's repo)
3
2
1
8
749
Idan Schwartz retweeted
GLM-5.2 can now be run locally!🔥 The 2-bit model retains ~82% accuracy after we shrunk it from 1.51TB to 238GB (-84% size). Run on a 256GB Mac or RAM/VRAM setups. GLM-5.2 is the strongest open model to date. Guide: unsloth.ai/docs/models/glm-5… GGUF: huggingface.co/unsloth/GLM-5…
Introducing GLM-5.2: Frontier Intelligence, Open Weights - Significant improvements in coding and agentic tasks - Strong long-horizon capabilities with a 1M context window - Two levels of reasoning effort: GLM-5.2 (max) pushes the limits, while GLM-5.2 (high) strikes a strong balance between performance and token efficiency - MIT-licensed open weights - Same API pricing as GLM-5.1 Tech Blog: z.ai/blog/glm-5.2 Weights: huggingface.co/zai-org/GLM-5… API: docs.z.ai/guides/llm/glm-5.2 Coding Plan: z.ai/subscribe Chat: chat.z.ai
266
844
275
7,251
1,904,522
מעכשיו אימרו: "המחלקה למדעי המחשב ובינה מלאכותית". חדשות כאן 11 בכתבה על בחירות הסטודנטים והחששות בעידן הAI. תחום מדעי המחשב וגם תחום ההנדסה רלוונטים מתמיד. @ComscienceBiu @ittaishick
1
1
7
541
Idan Schwartz retweeted
Our paper: "LaMI: Augmenting Large Language Models via Late Multi-Image Fusion" has been selected for an Oral Presentation at #ACL2026! LaMI boosts LLM visual commonsense by generating complementary images from a text prompt and late-fusing their evidence into the prediction 🧵
6
17
1
65
4,926
Had a fascinating deep conversation with @ylecun on Geometry and World Models at @iclr_conf 🤩
10
5
210
12,682
Idan Schwartz retweeted
3D editing has long relied on workarounds: per-asset optimization, 2D view propagation, or hacking frozen priors. The bitter lesson is the same one image editing already learned. Train a native model, end-to-end. Introducing ShapeUP, accepted to SIGGRAPH 2026 💫
5
39
4
244
19,194
Idan Schwartz retweeted
Funding opportunity for PhD students for 4 month visits in Israeli universities. Contact if you're interested in an internship with me. Focus areas: Interpretability and controllability of LLMs, AI safety, multi-agent communication, AI for Science. azrielifoundation.org/fellow…
17
1
64
7,808
Idan Schwartz retweeted
Are you interested in interning with me and my lab? A unique opportunity for a 4-month research stay, with generous funding as an Azrieli visiting PhD fellow! DM me if you're interested. azrielifoundation.org/fellow…
9
49
2
270
29,100
Idan Schwartz retweeted
Hiring: fully funded PhD on Trustworthy & Controllable LLMs at the University of Groningen. Focus: alignment/reasoning, making LLMs more factual, fair, safe and reliable. Strong ML/Python skills required. Top-tier publications are a major advantage. Apply: werkenbij.rug.nl/vacature/ph…
2
1
8
756
Idan Schwartz retweeted
New research from @bfl_ml 🥳 Meet Self-Flow: our self-supervised framework for image, audio, video & world models 🤖 bfl.ai/research/self-flow Do generative models really need DINO to learn strong representations? We propose teaching them directly via a joint framework instead 🧵
13
61
6
281
69,336