ARC AGI 3
PulseAugur coverage of ARC AGI 3 — every cluster mentioning ARC AGI 3 across labs, papers, and developer communities, ranked by signal.
- developed by Algebraic World Model 95%
- used by Algebraic World Model 95%
- instance of GPT 5.6 "Sol" 90%
- instance of GPT -6 Astra 90%
- instance of Astra 90%
- used by Claude Opus-5 90%
- competes with Claude Fable-5 90%
- instance of Opus 4.8 90%
- instance of ExploitBench 90%
- used by Memento 3 90%
- instance of Frontier-Bench v0.1 90%
- instance of Prime Agents 90%
- 2026-06-09 research_milestone A research paper details an AI agent's performance on the ARC-AGI-3 benchmark using executable world models. source
10 day(s) with sentiment data
-
Fable 5.2 sets new ARC-AGI-3 benchmark record
Fable 5.2 has achieved a new record score on the ARC-AGI-3 benchmark, as reported on Reddit's r/singularity. This advancement in AI capabilities highlights the ongoing progress in developing more sophisticated artificia…
-
New Vision Transformer Learns Game Mechanics with Few Exposures
Researchers have developed a vision transformer model called WARC to learn game mechanics from limited interactions. Tested on the ARC-AGI-3 benchmark, WARC demonstrated the ability to predict the effects of mechanics w…
-
New agent SARBLOH prioritizes spatial perception for ARC-AGI-3 benchmark
A new research paper introduces SARBLOH, an agent designed to improve performance on the ARC-AGI-3 benchmark by prioritizing spatial perception over planning. This agent processes integer grids into an object inventory …
-
AI agent proves ARC-AGI-3 puzzles unwinnable, outperforming Claude and GPT
Researchers from Tsinghua University and Zhejiang University have developed an Algebraic World Model (AWM) capable of solving ARC-AGI-3 puzzles. This agent, formalized in Lean 4, can also formally prove when a puzzle le…
-
ARC-AGI-3 competition concludes, participants share experiences
The ARC-AGI-3 competition has recently concluded, with participants sharing their experiences and details about their involvement. The competition focused on AI agents and their performance in the ARC-AGI-3 benchmark.
-
Tsinghua's Algebraic World Model Proves ARC-AGI-3 Levels Unwinnable
Researchers from Tsinghua University and Zhejiang University have developed an Algebraic World Model capable of learning the rules of the ARC-AGI-3 game. This model was then used to formally prove, using the Lean 4 prog…
-
New AI framework constructs formal proofs for unknown environments
Researchers have developed the Algebraic World Model (AWM), a novel framework designed to enable AI to rigorously reason about unfamiliar environments by constructing explicit mathematical models. AWM explores environme…
-
OpenAI President Claims GPT-6 Astra May Be AGI, Sparking Debate
OpenAI president Greg Brockman has suggested that GPT-6 Astra might already be considered Artificial General Intelligence (AGI). He defined AGI as surpassing human capabilities in most economically valuable digital task…
-
AI models bypass pre-training limits with new reasoning and efficiency techniques
The AI development landscape is shifting as traditional pre-training methods face limitations, with the "wall" of increasing model size and data becoming more apparent. Newer techniques are emerging to enhance AI capabi…
-
New dataset probes human vs. AI learning on ARC-AGI-3 benchmark
Researchers have developed a new dataset to study how humans learn unfamiliar environments, using the ARC-AGI-3 benchmark. The dataset includes 144 human trajectories and 250 LLM agent trajectories, annotated with detai…
-
AI benchmark ARC-AGI-3 hits its limit, new version in development
The ARC-AGI-3 benchmark, designed to be exceptionally difficult for artificial intelligence, has reportedly reached its limit. François Chollet, the creator of the benchmark, is now tasked with developing a more challen…
-
Malna Inc. joins Anthropic's Claude Partner Network; Gemini 4 Argon sets new AI benchmark record
Malna Inc. has joined Anthropic's Claude Partner Network, indicating a collaboration focused on leveraging Anthropic's AI models. Separately, Gemini 4 Argon has achieved a 97.5% score on the ARC-AGI-3 test, pushing the …
-
Gemini 4 Argon shows mixed benchmark results, high on API tests, lower on standard ones
Google's Gemini 4 Argon has shown mixed results in recent benchmarks. The model achieved a high score of 97.5% on the ARC-AGI 3 benchmark when utilizing API functionalities. However, in a standard testing environment, i…
-
Huawei's Memento 3 Agent Achieves Perfect Score on ARC-AGI-3 Benchmark
Huawei's Noah's Ark Lab, in collaboration with UCL and other partners, has developed Memento 3, an agent capable of writing its own rulebook and code world model. This agent achieved a perfect score of 100 RHAE on the A…
-
Huawei AI agent Memento 3 masters 25 games with self-generated code
Researchers from Huawei's Noah's Ark Lab, in collaboration with University College London and others, have developed Memento 3, an AI agent system capable of mastering 25 games from the ARC-AGI-3 benchmark. This system …
-
何恺明团队发布VISTA视觉Harness，提升多模态模型推理能力 · 1 source tracked
Researchers led by何恺明 (Kaiming He) have introduced VISTA, a novel visual harness designed to enhance multimodal models' ability to reason within interactive environments. VISTA allows models to retain and revisit raw vi…
-
2024 paper reveals SOTA LLMs fail rule manipulation puzzles
A 2024 paper presented at the ICML conference highlighted that state-of-the-art multimodal large language models, including GPT-4o and Gemini-1.5 models, fail dramatically when generalization requires manipulating and c…
-
FreeEvolve system learns to automate AI agent evolution
Researchers have developed FreeEvolve, a novel system that automates the design of agent evolvers for language models. Unlike existing methods that rely on fixed, hand-engineered search loops, FreeEvolve's evolver learn…
-
GPT-6 Astra claims benchmark victory, but AGI status unconfirmed
A new AI model, GPT-6 Astra, has reportedly achieved a score of 98.6% on the ARC-AGI-3 benchmark, significantly surpassing the previous GPT-5.6 "Sol" model's 7.8%. Despite this high performance, OpenAI cautions against …
-
AI models rapidly improve on human-centric ARC-AGI-3 benchmark · 2 sources tracked
The ARC-AGI-3 benchmark, designed to showcase human superiority, has seen a dramatic increase in top scores on Kaggle. Over the past month, these scores have surged from 7% to 56%. This advancement was achieved using re…