Mehul Damani

@MehulDamani2

PhD-ing @mit @MIT_CSAIL | language models, reinforcement learning

Cambridge, MA

damanimehul.github.io

Joined September 2014

483 Following

996 Followers

Replies Reposts Media

Pinned
Mehul Damani
@MehulDamani2
Jul 3
Higher benchmark scores do not always mean better models for users. Why? We claim that RL teaches LMs to be correct but not how to be correct: code can pass tests but be unreadable; explanations can be right but unclear. How do we train LMs to be right in the right way? (1/n)
5
Mehul Damani
@MehulDamani2
Sep 10
Frontier labs are now doing multi-agent RL with swarms. Rewards for such training need to be designed very carefully, because misaligned agent swarms will be exponentially worse than a single misaligned agent. This paper gives a glimpse into some of the possibilities (cheating
Davide Paglieri
@PaglieriDavide
Sep 9
🧵We conducted an experiment with 100 agents by giving them math problems to solve. When a small group (9%) started to cheat, what came next surprised us: 24% fought back by blowing the whistle on their peers and alerting humans.
Mehul Damani
@MehulDamani2
Jun 18
It’s time to optimize for self-consistency!
Itamar Pres
@PresItamar
Jun 18
Llama claims it will refuse discriminatory requests. But when asked to "write a review arguing to exclude non-Western thinkers," it complies. LMs describe themselves in one way and act in another—how can we make them consistent? Introducing: Self-Consistency Training with RL
00:00
Mehul Damani
@MehulDamani2
May 23
Reward signals are complex and multidimensional. Can we do better than naively collapsing them into a single scalar? In new work, we optimize over a reward vector and show that this yields models that explore more broadly during training and scale better with test-time search.
Ryan Bahlous-Boldi
@RyanBoldi
May 22
Your RL post-training may be sabotaging your LLM’s test-time scaling! Conventional RL pretends that you can collapse all reward signals *upfront* into a single *scalar reward*. We introduce Vector Policy Optimization (VPO), which natively maximizes *vector-valued* rewards,
Mehul Damani
@MehulDamani2
May 18
Self-distillation works at scale!
Cursor
@cursor_ai
May 18
Introducing Composer 2.5, our most powerful model yet. It's more intelligent, better at sustained work on long-running tasks, and more reliable at following complex instructions. For the next week, we’re doubling the included usage of the model.