Zihao (Gavin) Yang

1/ (New paper!) If swapping the gender in an input prompt makes the AI model give a different answer it means that it has to have a gender bias, right? Wrong. 🧵on counterfactual prompting for LLM evals: Paper: arxiv.org/abs/2605.01048

14/ Joint work with Mosh Levy [@mosh_levy] and Yoav Goldberg [@yoavgo] at Bar-Ilan / AI2, and my advisor Byron Wallace [@byron_c_wallace] at Northeastern. Huge thanks to everyone who gave feedback along the way. Comments, pushback, and replications all very welcome 🙏

13/ We packaged the framework so you can drop it into your eval & study: pip install cfprompt Paper: arxiv.org/abs/2605.01048 Experiments: github.com/redagavin/coun… Package: github.com/redagavin/cfpr…

1/ (New paper!) If swapping the gender in an input prompt makes the AI model give a different answer it means that it has to have a gender bias, right? Wrong. 🧵on counterfactual prompting for LLM evals: Paper: arxiv.org/abs/2605.01048

Behind Cambrian-S are the passionate researchers that drive it. This video is a presentation, but more so representation. I shot the short as an ode to the very humans behind, and these unique, surprising spaces and memories that are we. Please enjoy! May the experiment go on--

@sainingxie told us to ONLY work on "crazy ideas." Almost a year ago, we started Cambrian-S because "Supersensing" sounded super crazy. This crazy idea kept me awake and caffeinated for months. Today, all that work is live: Cambrian-S is here. So grateful to have built this alongside this incredible team. Please take a look here. Hope you find this idea crazy as well! Website: cambrian-mllm.github.io/cambrian-s/ Github: github.com/cambrian-mllm/… arXiv: @sainingxie" target="_blank" rel="nofollow noopener">arxiv.org/abs/2511.04670…

[Videos are entanglements of space and time.] Around one year ago, we released VSI-Bench, in which we studied visual spatial intelligence: a fundamental but missing pillar of current MLLMs. Today, we are excited to introduce Cambrian-S, our further step that goes beyond visual spatial intelligence to spatial supersensing. The core idea behind our work is that: we believe real supersensing intelligence requires the ability to not only see, but also actively anticipate, select, and organize its sensory input by constructing an internal world model. 👇Scroll down to delve deeper into our position, analysis, explorations, and findings along this supersensing journey. 🧵[1/n]

Since I started working on MLLMs, one question has been nagging at me... MLLMs are evolving incredibly fast. The progress from 2023 to 2025 has been wild. But... are we truly innovating, or just adapting? Think about the recent hot topics: visual reasoning, long video, agents, sparse/linear attention... It feels like we're just downstream, adapting SOTA LLM architectures and breakthroughs. (Huge kudos to the LLM researchers!🙏 ) It leaves me wondering: Is the future of MLLM research just waiting for the next big LLM wave and then figuring out how to "bolt on" vision? Or are there unique, fundamental multimodal intelligence problems that simply scaling the current paradigm can't solve? (1/n)

Introducing Cambrian-S it’s a position, a dataset, a benchmark, and a model but above all, it represents our first steps toward exploring spatial supersensing in video. 🧶