OmnAI Lab

1/ Large Language Models leak energy when they hallucinate. We built a training-free method to catch the spill and keep them *grounded*. Our #ICLR2026 paper introduces Spilled Energy for SOTA zero-shot detection. TLDR: Hallucinations violate the probability chain rule.

Why can't we solve adversarial examples? After a decade of work, neural nets still get fooled by imperceptible noise. We think we finally know the geometric reason why — and it connects to AI alignment. 🧵

Language Models are Injective and Hence Invertible (ICLR 2026), aka “pringle paper", is now a public graph on @paradigmainc’s Flywheel In the paper, we show that LMs can be inverted and, contrary to common belief, do not discard information about their inputs at inference time.

introducing Flywheel: the infrastructure for autonomous research.

📢 1 week left to submit your paper to the MUV Workshop at @CVPR 2026. Take a step toward safe AI deployment and privacy-preserving learning. More info: …chine-unlearning-for-vision.github.io Updates & details: @muv_workshop

Too late for #ECCV2026 submissions? 🤷 Not all is lost. Submit your work to MUV @CVPR — Machine Unlearning for Vision 🧠🔐 🗓 Deadline: March 15 🔗 …chine-unlearning-for-vision.github.io #AI #CVPR2026 #ComputerVision #ML

@OmnAI_Lab What we also find very interesting is: - instruct variants show a higher spillage to hallucination correlation; post-training artifacts? - the math behind does not apply to just language modeling, any sequence-to-sequence task abides by the same rules!

7/ The code is open-source, so you can try it in your own pipelines! GitHub: github.com/OmnAI-Lab/spil… Paper: arxiv.org/abs/2602.18671

6/ We evaluated LLaMA, Mistral, Gemma, and Qwen3 across nine Q/A and reasoning benchmarks, plus a synthetic algebraic stress test. The method effectively localizes the exact answer token and tests for hallucinations.

1/ Large Language Models leak energy when they hallucinate. We built a training-free method to catch the spill and keep them *grounded*. Our #ICLR2026 paper introduces Spilled Energy for SOTA zero-shot detection. TLDR: Hallucinations violate the probability chain rule.

AI doesn't just need to learn. It needs to unlearn as well ❌ Removing sensitive data or unsafe behaviors is becoming central to modern computer vision. @muv_workshop brings together researchers working on selective forgetting, safe image/video generation, privacy-preserving & ethically compliant AI. Speakers from @GoogleDeepMind, @MIT_CSAIL & more. Join us in Denver👇 …chine-unlearning-for-vision.github.io @CVPR

Not too so surprising, but I'm glad someone checked!

Don’t miss MUV @CVPR 2026! 🚨 If your work touches machine unlearning or safe adaptation in vision, this is your venue. 📝 Submit by March 15. #CVPR2026 #MUV #Workshop

Can AI forget? 🧠❌ Join MUV at @CVPR 26 in Denver! 🏔️ Speakers from @GoogleDeepMind, @MIT_CSAIL & more. 📝 Submit by March 15! Organizers: @SapienzaRoma, @MIT, @TU_Muenchen, @_italai and MPI. Details: …chine-unlearning-for-vision.github.io #CVPR2026 #AI #ComputerVision

To all my AI friends: Every time I see you, you raise my temperature parameter. Happy Valentine’s Day! ❤️

Something better than SAE just dropped "Learning a Generative Meta-Model of LLM Activations" Training a diffusion model to understand the activation states of LLMs, and even steer them?! In this paper, they train a diffusion model on 1B+ LLM internal activations to learn what “normal” hidden states look like. Then when you steer a model and its activations get weird, you can denoise/project them back onto the natural manifold keeping the behavior change while staying fluent and stable. This reveals a cleaner concept-like “meta-neurons” inside the learned prior!

we keep hearing that coding is just moving up the abstraction ladder, like swapping manual assembly for compilers. but what's the equivalent for science? if we abstract away experimental work and math execution, what's left? pure intuition? Join the discussion @iclr_conf