Oswin So

Graduate Researcher with Chuchu Fan at MIT @mit_REALM. Bringing Guarantees to Safe Reinforcement Learning 🇭🇰

[Guan-Horng Liu @guanhorng_liu]

Neither @oswinso nor I could make it to ICLR this year, but our poster did make the trip 🤣✈️🇧🇷

Come by our poster today about #DAM (#Discrete #Adjoint #Matching)—a unifying AM for discrete generative models (e.g., dLLM) !!

🕒 3:15 PM – 5:45 PM
📍 Pavilion 3 P3 #1711

[Guan-Horng Liu @guanhorng_liu]

Adjoint Matching works great for fine-tuning diffusion models with reward gradients.

How about #AM for #diffusionLLMs with #nondifferentiable #rewards? Does "discrete adjoint" even exist ... and how? 🤔

📢 Introduce #DiscreteAdjointMatching (#DAM)—a unifying AM for discrete generative models, accepted to #ICLR2026 🇧🇷

Work done with my amazing intern @oswinso and @RickyTQChen, Brian, Chuchu 🙌

📰 https://t.co/v2CcenlkAT

[Oswin So @oswinso]

At #NeurIPS from Dec 2 to Dec 7 in San Diego! Looking forward to catching up and meeting new friends.

Excited to chat about safety for robotics, constraint satisfaction in RL, and (stochastic) optimal control.

Feel free to DM me to grab coffee or have a chat!

[Isabel Liu @YijieIsabelLiu]

Robots can plan, but rarely improvise. How do we move beyond pick-and-place to multi-object, improvisational manipulation without giving up completeness guarantees?

We introduce Shortcut Learning for Abstract Planning (SLAP), a new method that uses reinforcement learning (RL) to discover shortcuts in the planning graphs induced by task and motion planning (TAMP) skill libraries. It is a plug-and-play module that can be trained on top of existing planners to speed up execution through learned shortcuts.
(1/5)

[Huihan Liu @huihan_liu]

Meet Casper👻, a friendly robot sidekick who shadows your day, decodes your intents on the fly, and lends a hand while you stay in control!

Instead of passively receiving commands, what if a robot actively sense what you need in the background, and step in when confident? (1/n)

[Robotics: Science and Systems @RoboticsSciSys]

🏆 Huge congratulations to the #RSS2025 Award Winners!
https://t.co/BJMqRQzQQG

[Oswin So @oswinso]

@Almost_Sure Yeah, I was thinking about it from the perspective of using Feller’s test of explosion to show that it either hits zero in finite time or explodes to infinity, and that former happens with positive probability but is not 1 so not almost surely.

[Oswin So @oswinso]

@momin_rayhan @ben_moll @comp_simon @jlperla @KahouMahdi @MarlonAzinovic Not exactly what you asked for, but here's a repo on Hamilton-Jacobi reachability (which is derived from HJB, but should translate) written in python using jax (and hence autodiff-able): https://t.co/F3ZPSMuUhn

It'll probably require a lot of changes to solve HJBs tho.

[Oswin So @oswinso]

@Almost_Sure Here is an attempt at a proof:

Let t be such that Wₜ > 0. Then, we must have that Lₜ > L₀=0 ⟹ Wₜ > Xₜ.

Suppose there exists some u. Since Wₜ > 0 occurs at arbitrarily small times, there will always exist some t < u where Wₜ > Xₜ, resulting in a contradiction.

[Oswin So @oswinso]

Suppose now that Xₜ is started from ε:
Xₜ ≔ ε + ∫₀ᵗ 1{Wₛ >= 0} dWₛ

Since Xₜ = ε + max(W_t,0) - ½ L(t) and L(t) strictly increases only when Wₜ=0, does there exist a time u such that

Xᵤ ≥ Wᵤ AND Xᵤ < 0?
https://t.co/1aPExUdxuR

[Oswin So @oswinso]

More observations and questions on the following stochastic integral:

Xₜ ≔ ∫₀ᵗ 1{Wₛ >= 0} dWₛ

Numerically simulating this does confirm that E[Xₜ]=0 and Xₜ does go negative. What I did not expect, however, is the distribution of Xₜ to look the way it does.

[Oswin So @oswinso]

@Almost_Sure I think I'm getting confused by the fact that Wₜ goes both negative (and positive) at arbitrarily small times, which (should?) mean that Xₜ also goes negative at arbitrarily small times (and hence inf { t | Xₜ < 0 } = 0 a.s. ?)

[Oswin So @oswinso]

@Almost_Sure Ah. I realized I have another bad typo. The condition should be whether there exists a time u such that

Xₛ ≥ Wₛ ∀ s∈[0,u] AND Xᵤ < 0?

i.e., while X goes negative, does it stay at or above W the entire time?

[Almost Sure @Almost_Sure]

#almostsure blog post: On the integral ∫I(W ≥ 0) dW

This looks at the mentioned integral, which displays properties particular to stochastic integration and which may seem counter-intuitive.

https://t.co/gRRyvtNg2O

[Oswin So @oswinso]

@Almost_Sure Good point. The second case can be reduced to the first case.

How would you show that it holds for ε=0?

[Oswin So @oswinso]

@Almost_Sure No worries, Im fine with the screenshot

[Oswin So @oswinso]

More observations and questions on the following stochastic integral:

Xₜ ≔ ∫₀ᵗ 1{Wₛ >= 0} dWₛ

Numerically simulating this does confirm that E[Xₜ]=0 and Xₜ does go negative. What I did not expect, however, is the distribution of Xₜ to look the way it does.

[Oswin So @oswinso]

Small question about Ito integrals: Consider

Xₜ ≔ ∫₀ᵗ 1{Wₛ >= 0} dWₛ

where Wₜ is a Brownian Motion and 1 is the indicator.

Xₜ is a martingale, so E[Xₜ] = 0. I would think that Xₜ is non-negative, but that doesn't seem to be true?

[Oswin So @oswinso]

Using intuition from the discrete case, "Xᵤ downcrosses 0 when Wᵤ also downcrosses 0", and so u exists. However, I have no idea whether this holds in the continuous limit...

Numerical simulations show that u exists, but I feel like this is due to numerical error?

[Oswin So @oswinso]

@BeatrizGietner I'm using https://t.co/TV3VAT4AJn

[Oswin So @oswinso]

@FunaRiite @Almost_Sure In this thread its the first one. The thread for the second one is here:

[Oswin So @oswinso]

I realized theres a typo: I mean to put 1{X_s>=0} instead of 1{W_s>=0}. That changes the question significantly though.

[Oswin So @oswinso]

@Almost_Sure Here's a bonus plot of W, X and L on a single path!

[Oswin So @oswinso]

@Almost_Sure Seems about right? I think L(t) =ᵈ | W(t) |, which looks close in the histogram plot