Aaron Roth

1-panel SMBC comic update with the caption "Science pro tip: you can prove anything you want using regression to the mean" where a scientist explains he can up student grades by smearing peanut butter and googly eyes on the foreheads of the worse performing ones.

If you oppose this it's because you want line go down, but I want line go up.

COMIC ◆ www.smbc-comics.com/comic/protocol
PATREON ◆ www.patreon.com/c/ZachWeiner...
STORE ◆ smbc-store.myshopify.com

I was the AEA's President Elect as the first Trump term began. We worried about government statistics then, and appointed a committee that included members (such as Google's Chief Economist) who might be in a position to help make reliable statistics available if the government went dark.
#econsky

Tried this out; generally very impressed with the quality of the feedback!

This was also part of a fun exploration we did at the same time in how to incorporate modern AI tools in the process of doing theoretical and experimental science. We learned a lot about that too --- we'll write more at some point.

The paper is here: arxiv.org/abs/2507.09683 This is joint work with @mkearnsphilly.bsky.social and Emily Ryu, from a fun visit in June.

Nevertheless, we show that in any DAG information will be aggregated at the end of a path of length D up to excess error ~ 1/sqrt{D}. And there are distributions such that no DAG with depth D can support information aggregation to error < 1/D. So depth is a fundamental parameter.

No! Even though y = x1 + x2, x1 is uninformative on its own, since it is independent of y. Hence A's predictor will be the constant function f_A(x1) = 0. This is all B sees, so learns nothing about x1. The best predictor he can learn from x2 has MSE 1/2 -- information was lost.

An example to see why the problem is delicate/interesting: Suppose there are two agents, A and B, arranged in a line A->B. A observes x1 and B observes x2. x1 and y are iid standard Gaussians, and x2 = y-x1. So there is a perfect linear predictor: y = x1 + x2. Can B learn it?

When is this enough for "information aggregation" to take place? i.e. when will someone in the network be able to produce a predictor that is as accurate as if they had trained a regression model on ALL features, even though nobody sees all features at once?

Nobody sees other's observations, just their parents' predictions. They can use these as features (together with their direct observations) to solve a regression problem. Their own predictions will be observed by their children. This is how information spreads in the network.

Suppose we have a regression problem that many people want to solve, but information is distributed across agents --- different agents see different subsets of features. And the agents are embedded in a network (DAG). When one makes predictions, they are observed by its children.

Intersectional Fairness in Reinforcement Learning with Large State and Constraint Spaces In traditional reinforcement learning (RL), the learner aims to solve a single objective optimization problem: find the policy that maximizes expected reward. However, in many real-world settings, it ...

Come say hi at our #ICML poster today during Poster Session 1 (W-600)! Joint work with @ericeaton.bsky.social @marcelhussing.bsky.social @optimistsinc.bsky.social @aaroth.bsky.social @mkearnsphilly.bsky.social!

arxiv.org/abs/2502.11828

Swap Regret and Correlated Equilibria Beyond Normal-Form Games Swap regret is a notion that has proven itself to be central to the study of general-sum normal-form games, with swap-regret minimization leading to convergence to the set of correlated equilibria and...

Delighted (and honestly a little bit stunned) that our paper “Swap Regret and Correlated Equilibria Beyond Normal-Form Games” was just awarded both the “Best Paper” and “Best Student Paper” at EC! arxiv.org/abs/2502.20229

Agentic LLM tooling like Windsurf is amazing. But you can already tell from the interface (you can pick from two dozen models from half a dozen different model providers, seamlessly) that AI is going to be a commodity. Not at all clear that LLM developers will get the surplus.

Amazing set of hires, congratulations!

Decimating (many times over!) scientific funding and (worse!) restricting international scientists from coming here will give up our enviable status as the world center for science and technology. Even if these policies are reversed 4 years down the line it might be too late. 5/5

But for the same reason it was near impossible for anyone to overtake the US while we were the agreed upon meeting point for scientific talent, it will be near impossible to regain this status if the center moves elsewhere. This is why the current moment is so dangerous. 4/

This is why for the last 75 years, the United States has dominated scientific discovery and why the world-beating tech companies were founded and headquartered in the US: this has been where the talent is. This was a huge source of American economic (and military) power. 3/

This makes it hard for any other country to compete: while the US has been the global center for science, no other country has even been able to retain their -own- best students domestically, let alone attract the best students worldwide. 2/

The United States has had a tremendous advantage in science and technology because it has been the consensus gathering point: the best students worldwide want to study and work in the US because that is where the best students are studying and working. 1/

Check out the terrific set of EC 2025 accepted papers! ec25.sigecom.org/program/acce...

What is the complaint? It sounds like it is finding plenty of duplicates, just as it says on the box.

In some sense the bigger deal is perhaps the lower bound --- I don't think it was previously known that you couldn't get bounds diminishing at a polynomial rate with epsilon even for high dimensions, though of course this was suspected given the lack of algorithms.

A surprising new parameter regime, but right, not a pure improvement. An exponentially better dependence on the dimension at a cost of an exponentially worse dependence on the error parameter. It mirrors recent new swap regret algorithms (one by the same author).

High dimensional online calibration in polynomial time In online (sequential) calibration, a forecaster predicts probability distributions over a finite outcome space $[d]$ over a sequence of $T$ days, with the goal of being calibrated. While asymptotical...

Wow! arxiv.org/abs/2504.09096

Referring to Godels incompleteness theorem in any context other than mathematical logic is also a red flag

Thanks for the pointer, I'll check it out!

The paper is here: arxiv.org/abs/2504.06075 and is joint work with the excellent @ncollina.bsky.social , @iraglobusharris.bsky.social , @surbhigoel.bsky.social , Varun Gupta, and Mirah Shi!

Just as our last paper generalized Aumann's agreement theorem, this paper tractably generalizes "information aggregation" theorems for Bayesian reasoners. Our results lift back to the classic Bayesian setting and give the first distribution free information aggregation theorems.

For any function classes H(A), H(B), H(J) satisfying this weak learning condition, we show how two parties can collaborate to be as accurate at H(J), by only needing to solve a small number of squared error regression problems on their own data over H(A) and H(B) respectively.