Developments·Paper
VISTA: A Visual Harness for Reasoning in an Interactive World
A new arXiv paper introduces VISTA, a visual harness giving general-purpose multimodal models long-horizon vision for interactive environments. It maintains lossless visual memory of past observations and lets the model actively retrieve and reorganize visual input while reasoning. On ARC-AGI-3 it raised Claude Opus 5.0's Relative Human Action Efficiency from 40.68 to 100.00, beating first-time humans on all 25 public games.
0 engagement1 sourceThu, Oct 1, 2026, 05:59 PM
AI Overview
- VISTA gives multimodal models long-horizon vision with lossless visual memory, letting Claude Opus 5.0 score a perfect 100.00 Relative Human Action Efficiency on ARC-AGI-3, completing all 25 public games with 57.4% fewer actions than first-time humans.
- VISTA outperforms same-model baselines with minimal harnesses across three additional visual game and puzzle benchmarks.
Entities
Related
Developments·Paper
Nvidia research: harness, not model, drives long-horizon AI performance
Nvidia published research showing the software harness—tools, memory, and routing rules—matters more than the underlying model for long-horizon tasks. Using a custom harness with a supervisor component, they achieved a 100% score on ARC-AGI-3 with Claude Opus 5. A related arXiv paper formalizes this as Harness Continual Learning (HCL).
Nvidia published research showing the software harness—tools, memory, and routing rules—matters more than the underlying model for long-horizon tasks. Using a custom harness with a supervisor component, they achieved a 100% score on ARC-AGI-3 with Claude Opus 5. A related arXiv paper formalizes this as Harness Continual Learning (HCL).
0300 engagement3 sourcesarxiv, rssWed, Aug 19, 2026, 03:12 PM
Social·Community
OpenAI's Astra scored 62.7% and 99.9% on ARC-AGI-3 across two harnesses; Fortune reports five launch-page numbers quietly changed
Practitioners examining OpenAI's Astra benchmark results highlight a 37-point split on ARC-AGI-3: 62.71% on the Standard harness at max reasoning versus 99.95% with the Provider Adapter at high reasoning. The Ofox AI team notes Astra's 88.0% one-attempt SRE-Bench score against Sol's 68.7% with up to four attempts, while Fortune reported five numbers quietly changed on OpenAI's launch page after it went live. The discussion questions whether model rankings should account for harness state persistence and compaction.
Practitioners examining OpenAI's Astra benchmark results highlight a 37-point split on ARC-AGI-3: 62.71% on the Standard harness at max reasoning versus 99.95% with the Provider Adapter at high reasoning. The Ofox AI team notes Astra's 88.0% one-attempt SRE-Bench score against Sol's 68.7% with up to four attempts, while Fortune reported five numbers quietly changed on OpenAI's launch page after it went live. The discussion questions whether model rankings should account for harness state persistence and compaction.
11135 engagement4 sourcesredditFri, Sep 4, 2026, 02:38 AM
Developments·Paper
Tycho: A coding-agent system for interactive abstraction in ARC-AGI-3
Tycho is a coding-agent system that constructs and uses game-specific programmatic world models to solve ARC-AGI-3 tasks, which require inferring rules, hidden state, and goals under action constraints. It separates actionable observations from non-actionable frames to model, test, plan, and repair strategies.
Tycho is a coding-agent system that constructs and uses game-specific programmatic world models to solve ARC-AGI-3 tasks, which require inferring rules, hidden state, and goals under action constraints. It separates actionable observations from non-actionable frames to model, test, plan, and repair strategies.
0 engagement1 sourcearxivThu, Jul 30, 2026, 02:34 PM
Developments·Paper
Ablation study on ARC-AGI-3 agents reveals executable world models and verification are key
Researchers at arxiv published an ablation study of Codex-based agents solving ARC-AGI-3, testing four variants to isolate the contributions of executable world modeling, scheduled simplification, and exact replay verification. Using gpt-5.4 and gpt-5.5, they found that the flexible-interface executable world model without verification underperformed, while the fixed-interface verification treatment with simplification matched the full agent's performance, suggesting verification is critical.
Researchers at arxiv published an ablation study of Codex-based agents solving ARC-AGI-3, testing four variants to isolate the contributions of executable world modeling, scheduled simplification, and exact replay verification. Using gpt-5.4 and gpt-5.5, they found that the flexible-interface executable world model without verification underperformed, while the fixed-interface verification treatment with simplification matched the full agent's performance, suggesting verification is critical.
0 engagement1 sourcearxivThu, Jul 16, 2026, 08:18 PM
Developments·Benchmark
Schema harness claims 99% on ARC-AGI-3 public set with Fable+4.8
A new harness called Schema reportedly achieves 99% and 95.35% on the ARC-AGI-3 public set using Fable+4.8 and GPT 5.6 Sol respectively. The results are only on the public set, which is easier than the held-out set, so real performance remains unknown.
A new harness called Schema reportedly achieves 99% and 95.35% on the ARC-AGI-3 public set using Fable+4.8 and GPT 5.6 Sol respectively. The results are only on the public set, which is easier than the held-out set, so real performance remains unknown.
251 engagement1 sourceredditThu, Jul 16, 2026, 07:42 PM