Entity trail

ARC-AGI-3

Source-backed findings, relationship evidence, citations, and briefing history from the public MindPattern archive.

Briefing refs

Findings

Edges

Sources

Showing the first 40 findings. More graph evidence exists in the corpus.

Corpus findings

2026-05-23 / skill-finderHermes Agent v0.10 Ships Closed Learning Loop: Agent Creates Reusable Markdown Skill Files from Experience, Benchmarked at 40% Task Time ReductionNous Research's Hermes Agent v0.10 (MIT license) ships 118 bundled skills, three-layer memory, and six messaging integrations. The core differentiator is the closed learning loop: when the agent solves a task, it writes a reusable Markdown skill file, stores the outcome in persistent memory, and adjusts its approach next time. TokenMix benchmarks show self-created skills cut research task time by 40% versus a fresh instance. At ~$0.30 per complex task on budget models and $5-10/month for always-on VPS hosting, the economics are viable for solo builders running personal agent infrastructure.
2026-05-07 / arxiv-researcherExecutable World Models for ARC-AGI-3: Agent Maintains and Refactors Python Model Before ActingA coding-agent system for ARC-AGI-3 maintains an executable Python world model, verifies it against previous observations, and refactors toward simpler abstractions as a practical MDL-like simplicity bias before planning actions. Uses a scripted controller with predefined interfaces — no hand-coded game logic. Early results on a notoriously hard benchmark, demonstrating that explicit world-model maintenance improves agent planning.
2026-05-03 / thought-leaders-researcherARC-AGI-3 Live: Humans Score 100%, Frontier AI Scores 0.51% — $2M Prize Competition ActiveFrançois Chollet's ARC Prize 2026 is live on Kaggle with over $2 million in prizes across ARC-AGI-2 and ARC-AGI-3 benchmarks. ARC-AGI-3 introduces interactive environments and game-style levels measuring agentic intelligence, where humans score 100% but frontier AI achieves only 0.51%. The March launch featured a fireside chat between Chollet and Sam Altman at Y Combinator HQ. This remains the most significant benchmark challenging the 'LLMs are close to AGI' narrative with concrete evidence.
2026-05-02 / vibe-coding-researcherARC-AGI-3 Deep Analysis: GPT-5.5 Scores 0.43%, Opus 4.7 Scores 0.18% — Three Failure Modes IdentifiedARC Prize published a detailed May 1 analysis of 160 replays from GPT-5.5 and Opus 4.7 on ARC-AGI-3's novel interactive environments, where humans score 100%. Three failure modes emerged: local observation without global comprehension, training data interference (mapping unfamiliar mechanics to known games like Tetris), and success without understanding (winning Level 1 while carrying misconceptions). Opus over-committed to confident-but-wrong theories; GPT-5.5 failed to compress observations into hypotheses at all.
2026-05-01 / arxiv-researcherGoogle Launches AI Research Agents Powered by Gemini 3.1 Pro (77.1% ARC-AGI-2)Google shipped Gemini 3.1 Pro scoring 77.1% on ARC-AGI-2 — a benchmark testing novel logic pattern solving — and deployed it to power autonomous AI research agents via AI Studio, Vertex AI, and Gemini CLI. Available in preview for agentic workflows. The ARC-AGI-2 score represents a significant jump in abstract reasoning capability for production-available models.
2026-04-26 / arxiv-researcherSolidCoder: Don't Imagine, Execute — State-of-the-Art Code Gen via Sandboxed VerificationAddresses the 'Mental-Reality Gap' where LLMs hallucinate execution traces during code generation. SolidCoder's S.O.L.I.D. architecture forces edge-case awareness before algorithm design and replaces imagined traces with sandboxed execution using property-based oracles. With GPT-4o achieves 95.7% pass@1 on HumanEval, 77.0% on CodeContests (+4.3pp), 26.7% on APPS (+3.4pp).
2026-04-19 / sources-researcherAI Explained: Two AI Models to 'Stir Government Urgency' — ARC-AGI-3 Context, OpenAI Spud Preview, Anthropic Pentagon PushAI Explained's latest video (73K views) covers exclusive reports on OpenAI's Spud model (internal codename for their next major release, compute freed by shutting down Sora) and Anthropic's strategy to leverage upcoming model capabilities to revive stalled Pentagon negotiations. Analyzed against ARC-AGI-3 benchmark results where frontier models scored under 1% while humans scored 100%, framing the gap between benchmark performance and real-world capability claims made to government buyers.
2026-04-18 / thought-leaders-researcherFrancois Chollet's ARC-AGI-3: Frontier Models Score Below 1% While Humans Score 100%The ARC Prize Foundation released ARC-AGI-3, an interactive reasoning benchmark of turn-based environments with no instructions, rules, or stated goals. Frontier models collapsed: Gemini Pro scored 0.37%, GPT 5.4 High 0.26%, Opus 4.6 0.25%, Grok-4.20 0%. Humans solve 100% on first try. Labs had spent millions training on ARC-AGI-2 pushing scores to ~50%, but all progress evaporated — proving those gains were benchmark optimization, not genuine adaptive reasoning. The $2M prize remains unclaimed.
2026-04-12 / skill-finderAddy Osmani's Code Agent Orchestra: Three-Agent Teams Consistently Outperform Single Agents — Developer-Written AGENTS.md Provides 4% Lift While LLM-Generated Files Reduce Success 3%Addy Osmani's O'Reilly CodeCon talk (published March 26) provides empirical data on multi-agent coding patterns. Three focused agents consistently outperform one generalist working 3x as long. Three patterns ranked by sophistication: subagents (parallel delegation), Agent Teams (shared task list + peer messaging + file locking), and hierarchical subagents (feature leads spawn their own specialists). Critical finding: LLM-generated AGENTS.md files offer no benefit and marginally reduce success rates (~3%), while developer-written context files provide ~4% improvement. Sweet spot: 3-5 teammates; token costs scale linearly. Key quote: 'The bottleneck is no longer generation. It's verification.'
2026-04-10 / saas-disruption-researcherCROSS-CATEGORY: Three Incumbents Pivot From Point Tool to AI Platform in Same Month — Canva, Slack, and Notion All Expand Beyond Original CategoryIn March-April 2026, three major SaaS incumbents simultaneously expanded beyond their original categories via AI: Canva (design → full marketing automation via Simtheory/Ortto acquisitions), Slack (messaging → native CRM + desktop agent via 30 AI features), and Notion (docs → custom multi-model AI agents with cross-platform MCP integrations). All three are using AI agents as the mechanism to collapse adjacent categories into their platform. The pattern suggests that AI doesn't just disrupt individual SaaS tools — it dissolves the category boundaries that defined the SaaS era.
2026-04-05 / sources-researcherGenerative World Renderer: 4 Million AAA Game Frames Dataset for Bidirectional Video GenerationResearchers release a 4-million-frame dataset (720p/30fps) of synchronized RGB and five G-buffer channels from Cyberpunk 2077 and Black Myth: Wukong, with sequences averaging 8 minutes up to 53 minutes of continuous recording covering diverse urban, outdoor, and indoor scenes with varying weather. The dataset enables bidirectional rendering — decomposing real game frames into geometry/material channels AND generating new video guided by G-buffers with text-prompt style editing — with 86 HuggingFace upvotes signaling strong builder interest in game-to-video-generation bridging.
2026-04-01 / vibe-coding-researcherPattern: ARC-AGI-3's Interactive Format Proves Benchmark Optimization ≠ Capability — Every Frontier Model CollapsedARC-AGI-3 moved from static visual puzzles to interactive game environments with no instructions or rules — and every frontier AI model's score collapsed from high ARC-AGI-2 numbers to below 1% (best: 0.26%), while humans maintain 100%. This is the strongest empirical evidence that high benchmark scores reflected format optimization rather than genuine adaptive reasoning. For agentic engineering, this matters: agents that 'solve coding benchmarks' may be pattern-matching the benchmark format, not demonstrating the exploratory reasoning needed for novel production tasks. The $2M prize specifically rewards agents that can explore, learn rules from scratch, and transfer knowledge — exactly what production coding agents need.

Graph relationships

RELEASED
Chollet
Chollet previewed ARC-AGI-3 measuring agency in interactive environments.
Source finding
RELEASED
François Chollet
Chollet released ARC-AGI-3 interactive reasoning benchmark.
Source finding
PREDICTS
Chollet
Chollet confirms ARC-AGI-3 launches March 25, 2026 with first format change since 2019.
Source finding
BENCHMARKED AGAINST
Claude Opus
Claude Opus scored under 1% on ARC-AGI-3 benchmark.
Source finding
BENCHMARKED AGAINST
Gemini 3.1 Pro
Gemini 3.1 Pro scored under 1% on ARC-AGI-3 benchmark.

Source trail

NVIDIA Blog — Hermes Agent Self-Improving AI arXiv ARC Prize / arXiv 2603.24621 ARC Prize Google Blog arXiv AI Explained

Graph sources

entity graphfindings textkg edgeskg entities