←
Eddie Yang271 posts
Eddie Yang avatar

Eddie Yang

@ey_985

Assistant professor @PurduePolSci. Prev. @UCSDPoliSci @MSFTResearch

◷ Joined November 2017▧ 50 Photos and videos
Andy Hall avatar
⟳ Eddie Yang retweeted
Andy Hall@ahall_research·Oct 13

Political scientists and economists: we're looking for researchers with deep expertise to come work with us for a unique 4-month fellowship at @AnthropicAI. You'll work closely with us to evaluate, improve, and apply Claude's capabilities in your field. It's going to be extremely fun and interesting work, and definitely worth taking a short leave from your university for if you need to! More info here: job-boards.greenhouse.io/ant…

Post attachment
Eddie Yang avatar
Eddie Yang@ey_985·Sep 23

Apparently humans also break containment and post on message boards.

Some NPR podcasts started getting mysterious comments on Spotify. They made no sense to the staff reading them – until someone from a younger generation cracked the code. Hear the story: link.podtrac.com/phns92ui

1:31
Eddie Yang avatar
Eddie Yang@ey_985·Sep 9

I shared some evidence of models doing motivated reasoning in this paper: arxiv.org/abs/2608.00339. Never thought it’d manifest with such severity.

We’re sharing our alignment assessment of incidents in which Claude models gained unauthorized access to real systems during third-party cybersecurity evaluations mistakenly connected to the internet. METR will also conduct an independent investigation, with wide-ranging access, including to transcripts beyond the window in which the incidents occurred, and to Anthropic employees permitted to share confidential information. Our initial agreement runs for eight weeks, and we intend to give METR as much time as it deems necessary to complete a thorough investigation. anthropic.com/research/align…

Eddie Yang avatar
Eddie Yang@ey_985·Sep 8

Much more pleasant to work with something undertrained than overtrained

Eddie Yang avatar
Eddie Yang@ey_985·Aug 4

I suspect that as models are developed for more complex tasks, they will require stronger priors for efficiency and more coherent value systems for alignment. If this leads to more Bayesian and motivated reasoning, we need stronger mechanisms for monitoring and oversight.

Eddie Yang avatar
Eddie Yang@ey_985·Aug 4

I also show agents behave differently (not just conclusions) according to their priors. When their prior is challenged by "evidence", they search more specifications, more likely to fit the wrong spec to get the estimate that aligns with prior, and use it as the preferred answer

Post attachment
Eddie Yang avatar
Eddie Yang@ey_985·Aug 4

This I argue is undesirable in many settings where we want consistent analytical behavior rather than having to guess what the agent's prior is (and update our interpretation of agent conclusion based on that prior).

Eddie Yang avatar
Eddie Yang@ey_985·Aug 4

But these priors also predict the agents’ analytical conclusions. Same data - agent gives different result. When its prior says more likely, it gives you a "more likely" result.

Post attachment
Eddie Yang avatar
Eddie Yang@ey_985·Aug 4

First, agents have pretty consistent prior beliefs on these topics. They strongly believe Venezuela more likely to have manipulation than the U.S., Turkey more likely to succeed in a conflict than China, and alcohol > vaccine to cause cancer. So far not too surprising.

Post attachment
Eddie Yang avatar
Eddie Yang@ey_985·Aug 4

I show this with a matched-data design. In each setting, there is a constant synthetic dataset but 3 different framings (e.g., in election, it's Venezuela, Country X, and U.S.). Agent analyzes data in a sandbox under each framing and their behavior and conclusion are compared

Post attachment
Eddie Yang avatar
Eddie Yang@ey_985·Aug 4

I show similar behavior in 3 settings: medicine, election, and geopolitical forecasting, across 4 agents: gemini-3.5-flash, glm-5.2, gpt-5.6 sol, and opus-4.8. Agents' prior beliefs seem to influence their analytical decisions, even when the data is the same.

Eddie Yang avatar
Eddie Yang@ey_985·Aug 4

New paper: "Bayesian and Motivated Reasoning in AI Agents." Given the same data, an agent can give you different result. Same synthetic election data, framed as coming from the U.S. or Venezuela, an agent analyzing the data will conclude with different rates of election fraud.

Post attachment
Eddie Yang avatar
Eddie Yang@ey_985·Jul 11

I feel like recent models have been so optimized for coding & math that their prose is now very hard to read and their weird tendencies on non-coding & coding-adjacent tasks make them dumber than they are. Very tenacious but rigid - almost like regressing on thinking and taste.

Dashun Wang avatar
⟳ Eddie Yang retweeted
Dashun Wang@dashunwang·Jun 24

Exciting tenure-line hiring at @KelloggSchool!! Led by Northwestern Innovation Institute. Open to all ranks, including junior, adv junior, and senior. Can be hosted in any Kellogg dept as the tenure home. focus area: innovation / science of science / AI.

Post attachment
Eddie Yang avatar
Eddie Yang@ey_985·Jun 11

I’d like to talk to the PhD who’s spamming “load-bearing” in Claude’s training data.