Laurence Liang

ML & Robotics Research

"Open Questions: October 2026"

2026-10-01

Questions

Cliff at Grainval (1882) by Claude Monet.

  • Is it coincidence that looped transformers and diffusion models both show promise? From my point of view, looped transformers and diffusion models both involve at least two forward passses of the inputs through the model weights. (Technically a "forward pass" in the reverse process for diffusion models, to be precise.) Is the common architectural choice, iterative forward passes, a key component to more capable and efficient learning architectures? What would other representations of more efficient and capable models look like?

  • How can a model make adquate predictions on out-of-distribution inputs? If a model is given inputs that it has never seen before (out of distribution), how will the model behave? What are the guarantees? How does this relate to "emergent abilities"? Are emergent abilities only a result of model size (so any sufficiently large model will exhibit signs of OOD generalization), or are emergent abilities still dependent on traces of 'said' capability in the initial data mix?

  • What are creative photography techniques? Currently interested in shutter drag, tri-chromatic compositing, frame composition, double exposures (or more), push-/pull-development, 'winging it' in manual mode (next step), flash photography (lead/lag flash), night photography only using ambient light, prisms, and photography with a high f-stop. If anyone has any further suggestions on how to be creative with photography, would love to learn more!

  • What are the true capabilities of a model? Inspired by the s1 paper and other related work, I find it super interesting that one could "augment" the capabilities of a model beyond conventional scaling laws (compute, data, params). In the case of s1, the model simply had longer "thinking" traces, though there's also other work that demonstrated that activation steering can also affect (and in some cases slightly improve) model capabilities. Thus, what is the true extent of model capabilities, and how can model capabilities be augmented and safeguarded beyond conventional scaling laws?

  • Why is sphere packing so sought after in math/CS research? I also found out that sphere packing can have some interesting practical applications, so quite curious to learn more about the initial motivation on why the math/CS community has been interested in sphere packing and its subsequent applications.

  • How do Kalman Filter varieties work? (At the equation level) This is something that always bugged me in undergrad. I understand that it comes down to two equations, an internal ("physics") model and a statistical ("observation") model, and then the state estimate is maximized using a max-likelihood method. However, how does it all connect at the equation (variable-) level? Also, how does one derive a state-model of an IMU? Intuitively it shouldn't be too hard (assuming only acceleration as observations in xyz and rpy combined with a gravity projection vector), though what is the ground truth of a correct state IMU model? During undergrad, it was hard to find a definite IMU reference, and this is something that seemed out of grasp.

  • How does SLAM work? (At the equation level) I understand the intuition of SLAM (Simultaneous Location and Mapping). Given a robot that is inserted in an unknown environment, the robot must simultaneously create a map of its environment and also determine its location. I also understand that the "trick" to solve SLAM is loop closure, where the robot must return to it's origin which enables all the observations to be "stretched" along a closed loop. What I struggle with is understanding, how does the math work at the equation level which enables SLAM?

  • How can you build a graphics driver (i.e. Vulkan) from scratch? It turns out graphic card compability (or lack of) is a pretty big challenge in robotics (in my experience when working on university research projects). Of course there is also dependency management challenge that gets tied into setting up simulation envs, though I consider dependency management a solvable problem assuming that conventional pacakging tools (i.e. Conda, uv, or others) are used. It's graphics driver compability that really seems to be a challenge to me.

  • What is the equivalency between attention and SSMs? This is something that I have been introduced to during a research project during undergrad where a team of students and I looked into cross-architectural distillation from an attention model to a SSM. However, the exact derivation of the SSM lineage and its equivalency to attnetion at the variable/equation level still remains elusive.

  • Why does GRPO work? A couple of friends introduced me to GRPO and how the reward signal is calculated relative to a 'sample' of predictions as opposed to an absolute score. However, not fully understanding why GRPO works (and even aspects of the "how") still bugs me a bit, and I feel like fully understanding GRPO at the equation level (and related methods, going as far back as PPO/Q-Learning) will clarify a lot more.

Thanks for reading! If you might happen to have any suggestions or recommendations, I'd love to get in touch!