Continual Harness: An Efficient Self-Improving Agent on ARC-AGI-3 by @sethkarten from @PrimeIntellect
> The heavy test-time learning required by the benchmark (ARC-AGI-3) pushes agents to form an internal world model of the rules and mechanics that updates with new evidence.
When I saw ARC-AGI-3 was out, I knew it would be a great testbed to show how Continual Harness could evolve through interactions. It's really inspiring to see the process of agents autonomously updating the harness. Long-horizon RL with CH will be the next!
@GregKamradt@chengshuai_shi@xinran_liang For future readers that might be interested: check out our first step towards ARC-AGI-3 with Continual Harness! https://t.co/bYCj1ex5xs
@GregKamradt@chengshuai_shi@xinran_liang That's actually a fascinating question I'm also thinking about! Happy to chat more, but I think proper scaffolding + environment design (i.e., observation, action space, reward) could be a good start to scale RL on ARC-AGI-3 : )
We've solved another piece of the generalist gaming agent puzzle!
Mario requires different skills than Pokemon: reactive navigation, spatial reasoning, and safe exploration rather than long-term memory and zero-sum reasoning.
Check out our new paper on finetuning VLMs to beat Mario via PPO. 👇
Excited to finally share Odysseus! A fun project pushing VLM agents toward 100+ turn decision-making in games. I really enjoyed exploring what makes VLM-based RL work under long-horizon tasks. Huge thanks to the amazing team, and can't wait to see where we take this next!
🔥 Excited to share our new paper:
🚀 Odysseus: Scaling VLMs to 100+ Turn Decision-Making in Games via Reinforcement Learning
🎮 We study how to make RL stable and effective for training VLM agents in long-horizon, visually grounded environments — using the video game Super Mario Land as a testbed.
📜 Paper: https://t.co/SPgs9dhtBz
🔗 Project page (w/ video demos): https://t.co/zwAclr3W8M
Today's state-of-the-art VLAs struggle to generalize zero-shot to new robot embodiments, despite training on extensive multi-embodiment data.
We introduce Language-Action Pre-training (LAP) and LAP-3B — the first VLA to achieve substantial zero-shot transfer to unseen real-world robot embodiments, through simply aligning action representation with language.
Everything is open-sourced! Try it out on your own robot:
🌐 https://t.co/dWuDtbV0Tn
Glad to see that @Kimi_Moonshot K2.5 uses AdvancedIF as a benchmark for IF evaluation. Rubric-based methods can go beyond verifiable rewards and become the next learning signal for more general applications🚀
Special thanks to my mentor @YunHeHe17 and all collaborators @metaai
Check our new rubric-based benchmark and RL for advanced LLM instruction following! Real-world conversations can be very complicated, and rubric-based verification can be more robust and interpretable for training and evaluation! Really enjoyed the collaboration with Meta folks!
🚀 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 ↓
🤔 Feel like your AI is bullshitting you? It’s not just you.
🚨 We quantified machine bullshit 💩
Turns out, aligning LLMs to be "helpful" via human feedback actually teaches them to bullshit—and Chain-of-Thought reasoning just makes it worse!
🔥 Time to rethink AI alignment.
@tourzhao Interesting work! Our Self-MoA paper in the previous workshop (https://t.co/iwqASuQv4u) also has very similar findings, which confirms that the MoA performance is rather sensitive to the quality, and mixing outputs from a single LLM is already very effective in various tasks.