Seth Karten

Super cool eval and event!

Lidar is great for cars but we dont have a high utility sensor equivalent for grasping. Which yes is ultimately measuring friction but i think that misses the point. We need Dense Verification to signal closed loop autonomy

Id say grasping is mainly a sensor issue

Diagram of LLM Economist: left—grid of persona‑conditioned worker agents; center—planner LLM sends tax schedule; right—social‑welfare ‘hill‑climb’.

🚀 New preprint!
🤔 Can one agent “nudge” a synthetic civilization of Census‑grounded agents toward higher social welfare—all by optimizing utilities in‑context? Meet the LLM Economist ↓

Special thanks to my collaborators Wenzhe Li, Zihan Ding, Samuel Kleiner, Yu Bai, Chi Jin and to Cooperative AI foundation for great feedback at the 2024 summer school

A sandbox for mechanism design: iterate incentive schemes inside large‑scale simulacra before touching the real world.
Thoughts? RT if you think generative agents can design policy. 🔄❤️

Timeline plot: planner changes each tax year; welfare remains elevated across turnovers.

Democratic alignment: in a special case, periodic citizen voting can fire the planner. Leader turnover keeps welfare high and prevents policy drift—central nudging plus decentralized oversight in one sandbox.

Bar chart comparing social welfare for U.S. baseline, LLM planner, Saez schedule; LLM nearly matches Saez.

Centralized nudging: the planner’s marginal taxes beat U.S. statutory rates and approach Saez on aggregate welfare (almost double vs baseline).

Synthetic behavioral policies → we sample workers from 2023 ACS skills & demographics, then let each agent verify its own bounded rational utility from individualized preferences, enabling counterfactual reasoning.

LLM Economist: Large Population Models and Mechanism Design in Multi-Agent Generative Simulacra We present the LLM Economist, a novel framework that uses agent-based modeling to design and assess economic policies in strategic environments with hierarchical decision-making. At the lower level, b...

We turn economic policy into in‑context RL:
• 1 000+ scalable persona‑agents with bounded‑rational utilities
• Planner agent designs marginal‑tax nudges
• Arbitrary economic simulacra in scenarios such as voting
📄 arxiv.org/abs/2507.15815
💻 github.com/sethkarten/L...

Diagram of LLM Economist: left—grid of persona‑conditioned worker agents; center—planner LLM sends tax schedule; right—social‑welfare ‘hill‑climb’.

🚀 New preprint!
🤔 Can one agent “nudge” a synthetic civilization of Census‑grounded agents toward higher social welfare—all by optimizing utilities in‑context? Meet the LLM Economist ↓

😅probably still confusing i guess. We are working on a getting started tutorial video/livestream

Are you on the discord? That is the best place to ask questions.
For visual learners, I recommend playing the games yourself to gain familiarity with why it is hard.

Hopefully the compression didnt ruin the text :(

Open review doesnt seem public yet but here are the titles

Full rules, practice ladder, and timeline 👇
pokeagent.github.io
Updated discord link due to massive enrollment! discord.gg/E2DuX5FWF7

Huge shout-out to my collaborators: Jake Grigsby, @stephmilani.bsky.social , Kiran Vodrahalli, Amy Zhang, Fei Fang, Yuke Zhu, and Chi Jin

GitHub - sethkarten/pokeagent-speedrun: Official repository of the NeurIPS 2025 Competition: The PokeAgent Challenge: Competitive and Long-Context Learning at Scale. (Track 2, Speedrunning) Official repository of the NeurIPS 2025 Competition: The PokeAgent Challenge: Competitive and Long-Context Learning at Scale. (Track 2, Speedrunning) - sethkarten/pokeagent-speedrun

⏱️ Speedrun baseline (Track 2) — pokeagent-speedrun
mGBA wrapper + tool-use prompts for Emerald any % agents.
GitHub → github.com/sethkarten/p...

🧩 LLM baseline (Track 1) — PokeChamp
Prompt-only GPT + minimax search hits 76 % win-rate vs. scripted bots.
GitHub → github.com/sethkarten/p...
🤖 RL baseline (Track 1) — MetaMon
Population-based training with action masking for competitive play.
GitHub → github.com/UT-Austin-RP...

Banner reading “PokéAgent Challenge @ NeurIPS 2025” with two panels: Track 1 – Competitive Pokémon Battle Bots, Track 2 – Long-Horizon RPG Gameplay. Call-to-action: “Create video-game AI! Win prizes! Live now at pokeagent.github.io.”

🚀 Launch day! The NeurIPS 2025 PokéAgent Challenge is live. @neuripsconf.bsky.social
Two tracks:
① Showdown Battling – imperfect-info, turn-based strategy
② Pokemon Emerald Speedrunning – long horizon RPG planning
5 M labeled replays • starter kit • baselines.
Bring your LLM, RL, or hybrid agent!

GitHub - sethkarten/pokechamp: Official repository of the spotlight ICML 2025 paper, PokeChamp: an Expert-level Minimax Language Agent. Official repository of the spotlight ICML 2025 paper, PokeChamp: an Expert-level Minimax Language Agent. - sethkarten/pokechamp

Full paper on OpenReview → openreview.net/pdf?id=SnZ7SKykHh
Code & starter kit → github.com/sethkarten/pokechamp

🚀 5 days until my ICML spotlight poster!
Key insights we’ll unpack:
• Base LLM + test-time planning
• Game-theoretic scaffolding
• Context-engineered opponent prediction
• Comparative LLM-as-judge (relative > absolute)

Catch me Thu Jul 17, 4:30-7 PM PT👇

Heading to #ICML2025 next week! If you’re into all things API (Artificial Pokémon Intelligence) from our PokéChamp spotlight to the upcoming NeurIPS PokeAgent Challenge, LLM-agent scaffolding & reasoning, or mechanism-design nudging, let’s connect. DMs open!

Congrats, Simon!

Also the Pokemon Agent challenge by @sethkarten.ai @stephmilani.bsky.social and others!

pokeagent.github.io

I do think that cross-posting is worth it still because the audience on this site isnt necessarily on the other

Social media is usually started by young people and academics are an older crowd so it mainly just follows pre-existing social graphs. I joined academic twitter originally for the discoverability, which seems quite lost here

Social media takeoff is hard. Bluesky still lacks the capability to compete with twitter

Thank you!!

Excited to announce that I will be spending the summer at @Waymo on the simulation realism team! I’ll be working on learning to generate simulated worlds.
🚙🚙🚙
Send me a message if youre in the bay area and want to chat!