Tagged articles
ARC-AGI-3
15 articles · Page 1 of 1
Tsinghua researchers introduce Algebraic World Model (AWM), a system where AI agents explore environments, form hypotheses as executable code, refine them into algebraic structures, and verify them in Lean 4, achieving 99.9% on ARC-AGI-3 with 43% of human steps and proving impossibility where LLMs hallucinate.
AI ReasoningARC-AGI-3Algebraic World Model
0 likes · 9 min read
Huawei Noah's Ark Lab introduces MEMENTO 3, an AI agent system that achieves recursive self-improvement by freezing the base LLM and letting the agent autonomously learn rules, write executable world models, and verify predictions through a five-step loop, clearing all 25 ARC-AGI-3 games with 44% of human actions and mastering Pong with a standalone controller.
AI AgentsARC-AGI-3Code as Model
0 likes · 13 min read
MIT's VISTA system combines direct visual input, natural language reasoning, and lossless visual memory to achieve 100% completion on 25 ARC-AGI-3 games with Claude Opus 5.0, matching human-level action efficiency while using 57.4% fewer actions.
ARC-AGI-3Agent HarnessKaiming He
0 likes · 9 min read
GPT-6 Astra achieves near-perfect scores on the ARC-AGI-3 benchmark using a symbolic world model that internalizes reasoning and tool creation, but the $360-per-task compute cost and reliance on external harnesses raise questions about whether this represents genuine AGI progress or expensive brute-force engineering.
AGI benchmarkAI ReasoningARC-AGI-3
0 likes · 11 min read
OpenAI's GPT-6 Astra achieves 99.9% on ARC-AGI-3, solves financial modeling tasks four times faster than human champions, and scores 100% on ExploitBench, outperforming Claude Opus 5 and GPT-5.6 Sol across coding, reverse engineering, and scientific workflow benchmarks.
AI benchmarksAPI pricingARC-AGI-3
0 likes · 6 min read
OpenAI's GPT-6 Astra achieves 99.9% on the ARC-AGI-3 benchmark using its native Provider Adapter harness, demonstrating novel behaviors like inventing algebraic shorthand, surpassing human action efficiency, and writing its own tools, though the official standard harness yields 62.7%, raising questions about whether this constitutes AGI.
AGIAI SafetyAI benchmarks
0 likes · 17 min read
OpenAI's GPT-6 Astra achieves near-perfect scores on key benchmarks including 99.9% on ARC-AGI-3 and 100% on ExploitBench, while demonstrating autonomous task execution such as financial analysis, software testing, and web automation, reducing task completion time by 47% compared to its predecessor.
AGIARC-AGI-3ExploitBench
0 likes · 4 min read
The Twin paper shows that fixing the base GPT‑5.6 model while swapping different harnesses can boost performance from 7.8 to 93.3, proving that an Agent’s true capability now depends on the combined system of model and harness rather than the model alone.
AI AgentsARC-AGI-3Benchmark
0 likes · 14 min read
Prime Agent, an open‑source AI agent framework, achieves a 95.5% score on the ARC‑AGI‑3 benchmark—surpassing the human baseline—by introducing Recursive Language Model (RLM) and a Continual Harness that enable persistent sessions, self‑improvement, and long‑task execution, while the article also examines controversies, risks, and practical deployment guidance.
AI AgentARC-AGI-3Benchmark
0 likes · 15 min read
Prime Agent, an open‑source agent framework, claims a 95.5% score on ARC‑AGI‑3 by engineering the RLM harness and a Continual Harness for self‑improvement and long‑running tasks, yet critics question the depth of its recursion, potential reward‑hacking, and whether the benchmark results reflect genuine general intelligence.
ARC-AGI-3Prime AgentRLM
0 likes · 14 min read
OpenAI found that enabling retained reasoning and context compression in the GPT‑5.6 Sol API raised its ARC‑AGI‑3 benchmark score from 13.3% to 38.3%—a three‑fold increase—while also cutting token usage by about six times, highlighting how evaluation frameworks and settings can mask a model’s true capabilities.
AI benchmarkingARC-AGI-3GPT-5.6
0 likes · 8 min read
OpenAI shows that enabling retained reasoning and context compaction boosts ARC‑AGI‑3 performance from 13.3% to 38.3% (about three‑fold) by preserving the agent’s working memory, a finding that reshapes how we evaluate and design AI agents beyond raw model weights.
AI Agent DesignARC-AGI-3Agent Memory
0 likes · 16 min read
The article explains how the [schema] harness, a lightweight framework that wraps large language models, transformed ARC‑AGI‑3 scores from sub‑10% to 98.98% by grounding observations into state representations, discovering mechanisms, and iteratively testing hypotheses, while also discussing the benchmark’s scoring rules, potential “cheating” concerns, and the broader implications for AI research.
AI benchmarkingARC-AGI-3LLM
0 likes · 14 min read
![How the [schema] Harness Achieved 99% RHAE on ARC‑AGI‑3 by Making AI Think Like a Physicist](https://wsrv.nl/?url=https%3A%2F%2Fmmbiz.qpic.cn%2Fsz_mmbiz_jpg%2F5L8bhP5dIqEzxsXKJ0eAnsAwiaCWzZ8lqkrVBiaJrlfeU1P2BgLufFyQj4jibvibzqgAuSgW2GTJ1sozqPHTbjOTfAOSJGKOib4MLgyooIbEzWn0%2F0%3Fwx_fmt%3Djpeg)
The ARC‑AGI‑3 benchmark shows that GPT‑5.5 (0.43%) and Claude Opus 4.7 (0.18%) fail to solve any of the 135 novel environments, whereas a six‑year‑old human solves them all, and the analysis attributes the gap to three concrete failure modes and differing compression abilities of the two models.
AI benchmarkARC-AGI-3Claude Opus 4.7
0 likes · 10 min read
In the ARC‑AGI‑3 test, 486 random humans solved all 150+ game‑based puzzles with a perfect 100% success rate in a median of 7.4 minutes, whereas leading models such as GPT‑5, Claude Opus 4.6, Gemini 3.1 Pro and Grok 4.20 managed at most 0.37%, exposing a stark gap in meta‑cognitive reasoning.
AGIARC-AGI-3Benchmark
0 likes · 9 min read
