Max Torop

EE PhD Candidate @ Northeastern University. Prior Research Intern @ Apple.

[Max Torop @toropmax]

@brendanh0gan Had a great time working with you this summer, Brendan! Best of luck with your next chapter!

[Tooba Imtiaz @tooba502]

Excited to present "LVT: Large-Scale Scene Reconstruction via Local View Transformers" at @SIGGRAPHAsia 2025, Hong Kong 🏙️

💡Linear-complexity, large-scale scene reconstruction in a single inference pass!

⚡Fast-forward: Mon 9am HKT (Hall 3F)
🗣️ Technical Presentation: Thu 1:10pm HKT (S426)

🔗 Project Page: https://t.co/x0mk6IMMoy
📄 Paper: https://t.co/RUnnKuHIRd

Work done during my internship at @Google Beam, with my amazing collaborators John Flynn, @lucyrchai, Kathryn Heal, @XuanLuo14, and Jungyeon Park ✨

#ACMSIGGRAPH #SIGGRAPHAsia2025 #3DGS

[Max Torop @toropmax]

@NeurIPSConf Thanks again to my great co-authors Aria Masoomi Masih Eskandar (@meskandars) and Jennifer Dy.

#NeurIPS2025 #SteeringVector #ActivationSteering #RepresentationEngineering #LLM

[Max Torop @toropmax]

Come stop by our steering vector poster 🪩 DISCO: Disentangled Communication Steering for Large Language Models 🪩 at @NeurIPSConf 2025 today from 4:30–7:30PM in Exhibit Hall C,D,E (Poster #4001)!

📜 Paper: https://t.co/wgkOVb9ccW
🤖 Code: https://t.co/6wIP8UV1nC

[Max Torop @toropmax]

@brendanh0gan Ahh got it, I meant model loosely in the sense of the original model vs the post-GRPO model for getting the activations. I haven't played much with deriving steering vectors from models that differ in weights, but it's really interesting!

[Brendan Hogan @brendanh0gan]

🎄 Advent of Small ML: Day 4:🎄Steering Vectors for Reasoning (Can we "inject" reasoning skills?)

most work on steering vectors focuses on style or alignment, making a model more "honest," "refusal-prone," or making it speak like a pirate - the idea was originally introduced to me by @MaxTee1996 ( I linked his post about his paper below - go check it out!)

but I was curious if we could use steering vectors to transfer reasoning capabilities?

can we extract the "reasoning boost" from a fine-tuned model and apply it to a base model without actually changing the weights?

so - I trained qwen 2.5 7B on the MATH500 dataset using GRPO (similar to previous days). it went from 33.75% → 56% pass@1.

to build the steering vectors I ran data through both the base and finetuned models and captured the difference in their activations

and calculated the steering vector for each layer as mean(finetuned) - mean(base).

then I injected these vectors back into the frozen base model to see if it got smarter - and it seemed to work (results below)

it didn't match the full fine-tuned model (which is expected), but it proves that "reasoning" isn't just locked in the weights, it produces a distinct activation signature that we can capture and transplant.

I think this is cool because steering vectors are incredibly lightweight (one vector per layer vs. massive checkpoints).

if this scales, you could theoretically have one base model and just "inject" different skill vectors (math, coding, creative writing) on the fly, rather than swapping full LoRAs or checkpoints.

Repo + Plots below. 👇

[Max Torop @toropmax]

Getting selected as a "Top Reviewer" for #NeurIPS2025 was a pleasant surprise to start my day!

[Max Torop @toropmax]

@sarthmit This is a super neat idea!

[Max Torop @toropmax]

@meskandars @mllabneu @NeurIPSConf #Steering #LLM

[Max Torop @toropmax]

🎉 Very happy to share that our work "DISCO: Disentangled Communication Steering for Large Language Models" has been accepted to NeurIPS 2025!

🪩 DISCO controls LLM behavior using steering vectors on query and value representations.

🔗https://t.co/VHL5KTrB3T

#NeurIPS2025

[Max Torop @toropmax]

@meskandars @mllabneu I'm also looking for research internships next Summer. If you think that I might be a good fit for your group, I would love to connect and chat.

[Sophie Xhonneux @SophieXhon11060]

📣Call for Blog Posts at #ICLR2026 @iclr_conf

Following the success of the past iterations, we are opening the Call for Blog Posts 2026!

https://t.co/6X1pKeS79i

Please retweet!

[Max Torop @toropmax]

@npparikh Thanks!

[Max Torop @toropmax]

@npparikh This is a really nice articulation of something many intuitively feel. Any other essays of a similar nature (i.e. light read, insightful, work/productivity related) that you recommend?

[Max Torop @toropmax]

I had a really pleasant experience today presenting our NeurIPS 2023 paper SmoothHess (https://t.co/lT0EnfB835) to the DtAK lab in Harvard. Thanks @FinaleDoshi for hosting and @siddharthswar for facilitating!

[Max Torop @toropmax]

I worked with Yu during my Masters when he was a PhD student. He is a great researcher and a really nice guy, this is a great opportunity!

[Yu Sun @YuSunMark]

I am looking to hire several Ph.D. students starting Fall 2025 @JHUECE. Anyone who is passionate for computational imaging/machine learning/statistical inference is welcome to apply. MS/Postdoc positions are also available! Here is the link: https://t.co/8bs1asw7mK #PhDposition

[Yu Sun @YuSunMark]

I am looking to hire several Ph.D. students starting Fall 2025 @JHUECE. Anyone who is passionate for computational imaging/machine learning/statistical inference is welcome to apply. MS/Postdoc positions are also available! Here is the link: https://t.co/8bs1asw7mK #PhDposition