Parth Asawa
@pgasawa
Followers
651
Following
512
Media
12
Statuses
75
CS PhD student @Berkeley_EECS
Cupertino, CA
Joined July 2020
Training our advisors was too hard, so we tried to train black-box models like GPT-5 instead. Check out our work: Advisor Models, a training framework that adapts frontier models behind an API to your specific environment, users, or tasks using a smaller, advisor model (1/n)!
14
39
236
🚀 Excited to release our new paper: “Barbarians at the Gate: How AI is Upending Systems Research” We show how AI-Driven Research for Systems (ADRS) can rediscover or outperform human-designed algorithms across cloud scheduling, MoE expert load balancing, LLM-SQL optimization,
8
31
125
My students thought it would be funny to teach a model to be a better advisor. One interesting finding -- if the advisor model isn't helpful, it eventually learns to stop giving advice. Ouch. More importantly, we can now use RL to train weak open-weights advisor models to
Training our advisors was too hard, so we tried to train black-box models like GPT-5 instead. Check out our work: Advisor Models, a training framework that adapts frontier models behind an API to your specific environment, users, or tasks using a smaller, advisor model (1/n)!
1
7
34
“We challenge the assumption that advisor models must be stronger than the advisee but instead view advisors as having the ability to learn through experience and then transfer those lessons to a more powerful student model – that is unable (or unwilling) to learn directly.” 🤣
Training our advisors was too hard, so we tried to train black-box models like GPT-5 instead. Check out our work: Advisor Models, a training framework that adapts frontier models behind an API to your specific environment, users, or tasks using a smaller, advisor model (1/n)!
2
3
19
Advisor models seem v. cool to me in terms of guiding larger, closed models. The advisor seems like a prompt optimizer, but can be much smaller/cheaper to work with than optimizing prompts directly on the big closed model. RL on the advisor makes a ton of sense too. Cool work!
Training our advisors was too hard, so we tried to train black-box models like GPT-5 instead. Check out our work: Advisor Models, a training framework that adapts frontier models behind an API to your specific environment, users, or tasks using a smaller, advisor model (1/n)!
0
2
24
Excited to see this work train "Advisor Models" using SkyRL! Great work by @pgasawa and @aczhu1326.
Training our advisors was too hard, so we tried to train black-box models like GPT-5 instead. Check out our work: Advisor Models, a training framework that adapts frontier models behind an API to your specific environment, users, or tasks using a smaller, advisor model (1/n)!
3
4
14
Very cool work! We cannot easily customize black box LLMs but we can actually dynamically steer them using smaller advice models!
Training our advisors was too hard, so we tried to train black-box models like GPT-5 instead. Check out our work: Advisor Models, a training framework that adapts frontier models behind an API to your specific environment, users, or tasks using a smaller, advisor model (1/n)!
1
6
56
@pgasawa Asked the students: How can we make learnable dynamic prompting ? They ended up writing a paper about training us.
1
2
15
I'm very excited about Advisor models: How can we personalize GPT5, when it’s behind an API? Sure, we can write prompts, but something learnable? We propose Advisor models which are small models that can be RL trained to give advice to a black-box model like GPT5. We show how to
Training our advisors was too hard, so we tried to train black-box models like GPT-5 instead. Check out our work: Advisor Models, a training framework that adapts frontier models behind an API to your specific environment, users, or tasks using a smaller, advisor model (1/n)!
5
38
220
In case you didn't get the joke reference like myself, they are referring to: https://t.co/FDXXql2r9K
Training our advisors was too hard, so we tried to train black-box models like GPT-5 instead. Check out our work: Advisor Models, a training framework that adapts frontier models behind an API to your specific environment, users, or tasks using a smaller, advisor model (1/n)!
1
3
28
Training advisors (human) is hard, but training advisors (AI) is easier. Introducing Advisor Models: a novel approach for "training" frontier black-box models for your application. Check out the thread for more details, and reach out to @pgasawa and myself if you want to chat!
Training our advisors was too hard, so we tried to train black-box models like GPT-5 instead. Check out our work: Advisor Models, a training framework that adapts frontier models behind an API to your specific environment, users, or tasks using a smaller, advisor model (1/n)!
1
2
12
📜 Paper: https://t.co/NU2mNypAzz 💻 Code: https://t.co/SZYOxcd9M4 This project was co-led with @aczhu1326 and advised by @matei_zaharia, @AlexGDimakis, and @profjoeyg. Reach out to @aczhu1326 and me if you want to chat about interesting applications! (8/n)
github.com
How to Train Your Advisor: Steering Black-Box LLMs with Advisor Models - az1326/advisor-models
0
2
22
Training advisor models also gave us some fun anecdotes. For example, one advisor learned to just stop talking. Fortunately (unfortunately?), our advisors haven’t learned to do that with us yet. (7/n)
1
0
11
We also conduct an analysis of the advisor and student responses and find the advisor is able to perfectly learn and articulate unstated environment latents. Here’s an example of what advice might look like before and after training: (6/n)
1
0
6
The modular design has key benefits unlike typical FT/RL tradeoffs: • Robustness: Specialize an advisor for one task (style) and the system won't forget how to do another (math). • Transfer: Train an advisor with a cheap model, then deploy it with a powerful one. (5/n)
1
0
8
The training framework excels at a variety of tasks related to personalization, reasoning, and more, demonstrating its ability to adapt advisors to latents in an environment and improve the black-box model’s performance over no training or static prompt baselines. (4/n)
1
0
9
Enter: Advisor Models. We use RL to train a small, often weaker, 'advisor' model to give dynamic advice to a powerful black-box 'student'. The advisor learns from experience and transfers that specialized knowledge to the generalist student in-context. (3/n)
1
0
12
Powerful black-box models are the go-to for LLM applications. But customizing black-box models is difficult, often limited to static prompts that can't adapt to different inputs. So how do you leverage frontier capabilities for a specialized environment? (2/n)
1
0
8
Super exciting to see how much the @rox_ai team has accomplished!! Engineering, acting, revenue ops -- they seem to do it all 😄
6 months, 25 million revenue agents & 3 trillion tokens later... Rox is now globally available 🌎 Just as coding agents 10x’d engineering, revenue agents 10x customer work. With Rox, humans are evolving to orchestrators while agents manage the end-to-end customer lifecycle.
1
0
15
Shreya’s work is awesome and she’s an amazing research mentor! Any university would be lucky to have her as faculty :)
On my way to VLDB! 🇬🇧 I am on the job market this year, seeking tenure-track CS faculty positions. I will be giving a talk on DocETL and on a panel titled “Where Does Academic Database Research Go From Here?” I would love to meet folks; please reach out if you’re also attending!
1
2
41