本文由 AI 分析生成
建立時間： 2026-05-30 來源： https://x.com/dair_ai/status/2038262704486400218
Summary
DAIR.AI’s weekly paper digest covering 10 papers from March 23-29, 2026. Highlights: Hyperagents (self-improving agents that improve their improvement mechanism), ARC-AGI-3 (interactive benchmark showing <1% AI vs 100% human performance), Claudini (Claude Code discovers novel adversarial attacks), Composer 2 (Cursor’s domain-specialized coding model via RL), and PivotRL (NVIDIA’s turn-level RL for agentic tasks with 4x fewer rollouts than end-to-end RL).
DAIR.AI 的週刊論文摘要涵蓋十篇論文。亮點：Hyperagents（可改進自身改進機制的自我進化 agent）、ARC-AGI-3（互動式基準測試顯示 AI <1% vs 人類 100%）、Claudini（Claude Code 自動發現新型對抗攻擊）、Composer 2（Cursor 的領域特化編程模型）、PivotRL（以 1/4 rollout 匹敵端到端 RL 的 agent 後訓練方法）。
Key Points
- Hyperagents: meta-level modification procedure is itself editable — the system improves how it improves
- ARC-AGI-3: 100% human vs <1% AI; requires exploration, goal inference, and planning without explicit instructions
- Claudini: Claude Code agent discovers attacks with 40% ASR vs 10% for all existing methods; white-box red-teaming is well-suited for automation
- Composer 2: train-in-harness infrastructure; 61.7 TerminalBench, 73.7 SWE-bench Multilingual
- PivotRL: identifies “pivot” turns where sampled actions have high outcome variance; +10% out-of-domain accuracy; 4x fewer rollouts than end-to-end RL
- MemCollab: contrastive trajectory distillation produces agent-agnostic shareable memory
Insights
The ARC-AGI-3 human-AI gap (100% vs <1%) is the starkest benchmark result in recent memory — it specifically tests interactive turn-based exploration without external knowledge, targeting genuine adaptive reasoning. Composer 2’s “train-in-harness” principle — training in the same harness used at deployment — directly validates the co-evolution argument from the harness architecture debate. PivotRL’s identification of “pivots” (high-variance decision points) as the training signal is a clean insight that could generalize beyond coding tasks.
Connections
Raw Excerpt
“ARC-AGI-3: Humans can solve 100% of the environments while frontier AI systems score below 1%. This gap demonstrates that current systems lack the fluid adaptive efficiency that humans exhibit on genuinely novel tasks.”