Lightnews — Scholar-powered news

Josef Woldense @woldense.bsky.social · 11h

Cool paper. I'm going to shamelessly plug my work here that also deals with LLMs for research

bsky.app/profile/wold...

Josef Woldense @woldense.bsky.social · Sep 8

Paper alert 📣

Rapid advances in AI has some believe that LLM agents can replace real participants in human-subject research. If true, this would be huge!

Following a growing body of research, we delve deeper into this topic and examine the merits of this claim.

🧵...

arxiv.org/abs/2509.03736

Are LLM Agents Behaviorally Coherent? Latent Profiles for Social Simulation

The impressive capabilities of Large Language Models (LLMs) have fueled the notion that synthetic agents can serve as substitutes for real participants in human-subject research. In an effort to evalu...

arxiv.org

1 3

Reposted by Josef Woldense

International Studies Association @isanet.bsky.social · 14d

Preparing for your upcoming #ResearchTalk? Sign up for two courses, taught by @woldense.bsky.social, to enhance your #Communication and apply #Storytelling principles to your #Presenting, and #Teaching skills! Open to both ISA Members and non-members. Register: buff.ly/PUXowRN

1

Josef Woldense @woldense.bsky.social · 20d

Congrats 🎉... looking forward to the coming research

1 1

Josef Woldense @woldense.bsky.social · 20d

Congrats!

Reposted by Josef Woldense

Vincent Arel-Bundock @vincentab.bsky.social · 21d

Whoa—my book is up for pre-order!

𝐌𝐨𝐝𝐞𝐥 𝐭𝐨 𝐌𝐞𝐚𝐧𝐢𝐧𝐠: 𝐇𝐨𝐰 𝐭𝐨 𝐈𝐧𝐭𝐞𝐫𝐩𝐫𝐞𝐭 𝐒𝐭𝐚𝐭 & 𝐌𝐋 𝐌𝐨𝐝𝐞𝐥𝐬 𝐢𝐧 #Rstats 𝐚𝐧𝐝 #PyData

The book presents an ultra-simple and powerful workflow to make sense of ± any model you fit

The web version will stay free forever and my proceeds go to charity.

tinyurl.com/4fk56fc8

9 84 270

Josef Woldense @woldense.bsky.social · 22d

This looks fascinating!

Bret Beheim @babeheim.bsky.social · 22d

How to quantify the impact of AI on long-run cultural evolution? Published today, I give it a go!

400+ years of strategic dynamics in the game of Go (Baduk/Weiqi), from feudalism to AlphaGo!

Miyagawa Shuntei's 1898 painting, "Playing Go (Japanese Chess)"

1

Josef Woldense @woldense.bsky.social · 24d

Congrats!

1

Josef Woldense @woldense.bsky.social · Sep 8

dongyeopkang.bsky.social

Josef Woldense @woldense.bsky.social · Sep 8

There is more in the paper, but broadly speaking, our results identify a deceptive problem: surface-level plausibility masking deeper failure modes. Agents appear internally consistent while concealing systematic incoherence.

Be careful when using LLMs as human substitutes. They might fool you.

1 1

Josef Woldense @woldense.bsky.social · Sep 8

Take pairs where one of the agents has a preference of 1. Next, take pairs where one of the agents has a preference of 5. Now compare them. You can see pairs with a 1 have lower agreement scores than pairs with a 5. This is consistent across preference gaps

1

Josef Woldense @woldense.bsky.social · Sep 8

Let me give you another one.

If we both equally dislike soda, our common ground should lead to high agreement. Not so with our agents.

1

Josef Woldense @woldense.bsky.social · Sep 8

The problem persists, even when we try to guard against the problem of sycophancy (column 3 of the graph).

(see paper for more info on sycophancy)

1

Josef Woldense @woldense.bsky.social · Sep 8

Our estimate suggests that the suppression of disagreement is quite large. Our counterfactual agreements scores (expected in the graph) are significantly lower than the observed ones, and this is across preference gaps.

(see paper for info in mean shift)

1

Josef Woldense @woldense.bsky.social · Sep 8

To do this, we adopt a simplifying assumption – agents should disagree at the same rate as they agree. We already know one end of this spectrum -- the amount of agreement when agents are aligned (gap = 0). We establish the disagreement side (gap = 4), by assuming it to be the inverse of agreement

1

Josef Woldense @woldense.bsky.social · Sep 8

When agents are aligned, they reach close to the highest agreement score. Yet, when maximally different (gap = 4), they come nowhere near the lowest score. It seems agreement is amplified while disagreement is dampened.

Is it possible to estimate how much disagreement is being suppressed? Yes!

1

Josef Woldense @woldense.bsky.social · Sep 8

Looking at the graph, it appears consistent with our expectations, the more closely aligned the agents (smaller preference gap between agents), the higher the agreement score.

But there is a problem. Can you spot it?

1

Josef Woldense @woldense.bsky.social · Sep 8

What are the results? Are agents internally consistent?

At first glance, yes. After a more thorough analysis, the answer is no.

1

Josef Woldense @woldense.bsky.social · Sep 8

How do we measure agreement level?

With the aid of an LLM judge, we score each conversation (strongly disagree = 1 – strongly agree = 5). This yields a set of agreement scores for a given preference pair. Using bootstrap sampling, we derive the distribution of average agreement scores (range)

1 1

Josef Woldense @woldense.bsky.social · Sep 8

We elicit the agents’ preference on a topic (1-5 scale), then pair them in a conversation to see if they follow through on their preferences.

Expectation: The more closely agents align in their preferences, the more strongly they will agree. The further apart, the more they disagree.

1

Josef Woldense @woldense.bsky.social · Sep 8

The basic intuition of internal coherence: If a person says they strongly prefer water over soda, we expect them to follow through on it. When offered both, they should select water, not soda.

How do we test for internal coherence?

1 1

Josef Woldense @woldense.bsky.social · Sep 8

Unlike other studies that look at how successfully LLM agents adopt human personas, we take a different approach and ask: Once a persona has been adopted, are agents internally coherent?

1 1

Josef Woldense @woldense.bsky.social · Sep 8

Paper alert 📣

Rapid advances in AI has some believe that LLM agents can replace real participants in human-subject research. If true, this would be huge!

Following a growing body of research, we delve deeper into this topic and examine the merits of this claim.

🧵...

arxiv.org/abs/2509.03736

Are LLM Agents Behaviorally Coherent? Latent Profiles for Social Simulation

The impressive capabilities of Large Language Models (LLMs) have fueled the notion that synthetic agents can serve as substitutes for real participants in human-subject research. In an effort to evalu...

arxiv.org

1 2 13

Josef Woldense @woldense.bsky.social · Sep 4

Congrats!! I see more GETGOV workshops on the horizon

1 2

Josef Woldense @woldense.bsky.social · Sep 4

Unfortunately, no. I will make my way to Europe next year though, so hopefully we'll get a chance to catch up

1 1

Josef Woldense @woldense.bsky.social · Sep 4

Congrats!!

1 1