chinmaya_mohan's profile picture. applied research @CapitalOne, previously @StanfordAILab / @StanfordHAI

Chinmaya Andukuri

@chinmaya_mohan

applied research @CapitalOne, previously @StanfordAILab / @StanfordHAI

Chinmaya Andukuri reposted

One of my takeaways from #COLM2025 was that people are thinking a lot about user simulation (have been thinking about this myself in the context of tutoring!) Really exciting to see this work on the topic 🤩

Simulating user–AI conversations helps us understand how LMs work in multi-turn settings. Prompting LMs like GPT-4o to simulate users is common, but their assistant nature makes it hard to replicate user behavior. We introduce User LMs - trained to be users, not assistants.

tareknaous's tweet image. Simulating user–AI conversations helps us understand how LMs work in multi-turn settings.

Prompting LMs like GPT-4o to simulate users is common, but their assistant nature makes it hard to replicate user behavior.

We introduce User LMs - trained to be users, not assistants.


have been enjoying dipping my toes into `verifiers` and @PrimeIntellect environments hub - just pushed an eval environment for MultiChallenge (@scale_AI) to the Environments Hub my env: app.primeintellect.ai/dashboard/envi… main page: scale.com/leaderboard/mu…


if you’re at @COLM_conf, come say hi tomorrow and talk to us about LM self-improvement + clarification!

Presenting this tomorrow at @COLM_conf! Poster 36 (11:00 AM-1:00 PM). We’ll have a demo—come along if you want to try our models and talk about multi-turn dialogue!



Chinmaya Andukuri reposted

Constitutional AI showed LMs can learn to follow constitutions by labeling their own outputs. But why can't we just tell a base model the principles of desired behavior and rely on it to act appropriately? Introducing SAMI: Self-Supervised Alignment with Mutual Information!


Chinmaya Andukuri reposted

Excited to share OffTheRails: A moral reasoning benchmark beyond trolley problems! We present a simple prompting pipeline for generating moral reasoning evaluations with language models using causal templates 🔵→🟠

jphilippfranken's tweet image. Excited to share OffTheRails: A moral reasoning benchmark beyond trolley problems!

We present a simple prompting pipeline for generating moral reasoning evaluations with language models using causal templates 🔵→🟠

Chinmaya Andukuri reposted

Language models struggle to search, not due to an architecture problem, but a data one! They rarely see how to search or backtrack. We show how LLMs can be taught to search by representing the process of search in language as a flattened string, a stream of search (SoS)!


Chinmaya Andukuri reposted

Multi-turn interactive RL should be a bigger focus. Current methods are not well-suited for this - i.e. PPO can't train with user in the loop generally and offline Q-learning still does not work at scale. It's interesting to see more work in that direction.

When prompting language models to complete a task, users often leave important things unsaid. Can language models teach themselves to ask clarifying questions? In STaR-GATE, we explore LMs' ability to self-improve by rewarding the model for generating useful questions!



Chinmaya Andukuri reposted

New work where language models learn to ask questions? So they can better understand user needs? With an amazing method name? Oh, yes!

When prompting language models to complete a task, users often leave important things unsaid. Can language models teach themselves to ask clarifying questions? In STaR-GATE, we explore LMs' ability to self-improve by rewarding the model for generating useful questions!



really enjoyed working on STaR-GATE! thanks to the team + many others for helpful discussions 🚀 check out the arXiv here: arxiv.org/abs/2403.19154

When prompting language models to complete a task, users often leave important things unsaid. Can language models teach themselves to ask clarifying questions? In STaR-GATE, we explore LMs' ability to self-improve by rewarding the model for generating useful questions!



United States Trends

Loading...

Something went wrong.


Something went wrong.