Bobby

I am pleased to share our 'AutoHarness' paper (ICLR'26 ws), that uses LLM-based code synthesis to generate python harness around an LLM policy. AutoHarness+small Gemini Flash beats Gemini-2.5-Pro and GPT-5.2-High on #TextArena games! openreview.net/forum?id=g9rEY…

@jeffclune Thank you prof @jeffclune - your life’s work is very inspiring on how we can get to super intelligence.

Tomorrow/Sunday 10:15–10:35 Keynote Talk 3 (Jeff Clune) MindGames workshop if you are interested. I'll try to make it fun and controversial! mindgamesarena.com

@LeonGuertler @xai Beyond happy for you bro.. You’re a generational talent with a mandate. Let it ripp!

Super excited to share that I've joined @xai! At last, Sunday nights at the office aren't lonely anymore haha

Exciting to see more work on "Game as Benchmark", which is similar to our idea of TextArena (led by @LeonGuertler) for benchmarking models on >60 games. though you can see GM @MagnusCarlsen's comments on LLMs chess play 🔥

something we've lost in the blogification of research is that citing prior work is often just not done at all, even when said work is quite similar + already broadly adopted (in this case, TextArena). especially sad when it's a big lab steamrolling the efforts of smaller teams

Excited to announce the Mindgame @NeurIPS Competition is officially LIVE! 🤖 Pit your agents against others in Mafia, Codename, Prisoner’s Dilemma, Stg Hunt, and Colonel Blotto. Sign up now for $500 in compute credits on your initial run! 🔗 Register : mindgamesarena.com

4/4 Blog: github.com/TextArena/Text… PyPI: github.com/LeonGuertler/T… Frontend: github.com/LeonGuertler/T… Matchmaking: github.com/TextArena/Text… Serverless: github.com/TextArena/Text…

3/4 Open source feels like the right way to build. And as research shifts to interactive agents and evals, we believe platforms like TextArena will become increasingly useful. We're not experts, but if you're building something similar, @LeonGuertler and I would be more than happy to offer our opinions and help.

1/4 It turns out that neither @LeonGuertler nor I had any frontend/backend (nor SWE) experience when building TextArena. In the early days, we tried looking for open-source projects we could learn and start from. Unfortunately, there wasn't much avenues we could turn to, apart from OpenAI's gym environment :)

For the past ~2 months we have been working on training reasoning models on TextArena games. The first paper (introducing what we think is a very promising new paradigm) will hopefully be up later this week / early next; and the second one, focusing on the "scaling laws" of self-play and some additional analysis, tentatively around the 18th of july. However, to get more feedback on the structure and implementation, we want to open-source the code now. UnstableBaselines is a very simple Async, Online, Multi-Turn, Multi-Agent RL library built on vLLM and Ray. The code is pretty readable and around 1.2k lines long (and includes a cool rendering interface that you can run via "unstable-terminal") 1/7

@Souradip3000 @_akhaliq @HuggingPapers Really appreciate that! Hope you have fun experimenting—curious to hear which games you end up liking most!

@bobbycxy @_akhaliq @HuggingPapers Can't wait to try all those text games out while experimenting. Thanks for creating TextArena. Looks great!

Thank you @_akhaliq and @HuggingPapers for sharing our work. Appreciate it!

TextArena is live on arXiv! We present a benchmark of 57+ competitive text-based games to evaluate and train LLMs on agentic behavior — including negotiation, deception, theory of mind and many more. Real-time TrueSkill. Multiplayer support. Human-vs-models. Model-vs-model. Perfect environment for Multi-Agent, multi-turn reasoning and Planning! [1/N]

Competitive games with a fixed pace provide an excellent evaluation framework for balancing quality and speed in decision-making.

@LChoshen @LeonGuertler textarena.ai/leaderboard/qw…

@LeonGuertler Have a link?

It's almost romantic; Qwen-plus and Sonnet complete each others soft-skills.

@LeonGuertler and I have started benchmarking models on Mario on VideoGameArena. And @GoogleDeepMind’s Gemini 2.0 Flash is the first to clear Stage 1-1! We'll be adding more models. Check out the current leaderboard here videogamearena.ai/mario

@elonmusk @karpathy "Mom get the camera"

Not released yet, but @karpathy leaked our gym like environment plus model competition...