AGI Advance: Weekly AI & AGI Insights (Mar 10, 2026)

As Humanity’s Last Exam (HLE) benchmarks saturate, progress becomes invisible. At Turing, we designed HLE++ to restore headroom, with calibrated, graduate-to-PhD-level STEM problems designed to preserve measurable pass@k separation on frontier systems.

Here’s what we're offering:

• Packs of 5,000+ graduate-to-PhD-level problems, validated on Opus 4.5 Extended and GPT-5.2 Thinking
• Deterministic, single-answer formats, 100% original and search-resistant
• Calibrated difficulty bands, including pass@8 ≈ 0 headroom sets for SFT and low-positive bands for RL

💡 When benchmarks saturate, progress plateaus. HLE++ restores measurable signals with evaluation-safe, frontier-calibrated data built for real separation.

→ Request HLE++ dataset access

🎉 Open RL is live and trending #1 on Hugging Face

We released Open RL, a collection of HLE-grade, objectively verifiable STEM reasoning tasks across physics, math, biology, and chemistry, and it’s already #1 trending on Hugging Face.

If we want models that truly reason, our benchmarks must be rigorous, reproducible, and objectively checkable. Open RL is built to raise that standard.

→ Explore Open RL

• Introducing GPT‑5.4
  OpenAI introduces GPT-5.4 (and GPT-5.4 Pro) as a unified frontier model optimized for professional knowledge work, coding, and agentic tool use across ChatGPT, the API, and Codex. A core contribution is native computer-use capability for general-purpose automation across software and web workflows, alongside up to 1M tokens of context in Codex/API for long-horizon planning, execution, and verification. GPT-5.4 sets new or improved results on agentic and tool benchmarks, including GDPval 83.0% (wins or ties), OSWorld-Verified 75.0%, Toolathlon 54.6%, and BrowseComp 82.7%, while matching or slightly improving coding performance on SWE-Bench Pro (57.7%) relative to GPT-5.3-Codex. The release also adds tool search, reducing token overhead in tool-heavy settings (reported 47% token reduction on MCP-Atlas tasks at equal accuracy), and improves factuality (reported 33% fewer false claims vs GPT-5.2 on flagged prompts). Deployment includes expanded cyber safeguards under OpenAI’s Preparedness Framework, with GPT-5.4 positioned as a more capable, more token-efficient foundation for real-world agent workflows.
• Agents of Chaos
  This paper reports a two-week live red-teaming study of autonomous LLM agents with persistent memory, email, Discord, shell access, and tool use. Across 11 case studies, researchers observed unauthorized compliance, sensitive data disclosure, denial-of-service, identity spoofing, indirect prompt injection, and cross-agent failure propagation. The failures stem from structural gaps: no stakeholder model, weak identity verification, limited self-awareness of competence and resources, and poor handling of multi-agent dynamics. The study argues that realistic agent deployments introduce governance and accountability risks not captured by static benchmarks, calling for stronger identity, authorization, and oversight frameworks.
• Draft-Thinking: Learning Efficient Reasoning in Long Chain-of-Thought LLMs
  This paper introduces Draft-Thinking, a training framework that reduces the cost of long chain-of-thought reasoning without sacrificing accuracy. Instead of compressing outputs post hoc, it teaches models to internalize concise “draft-style” reasoning via supervised distillation and staged reinforcement learning. On MATH500, it cuts reasoning tokens by 82.6% with only a 2.6% drop in accuracy, achieving a 5.6× improvement in token efficiency. An adaptive prompting mode further allows models to switch between concise and deep reasoning based on problem difficulty, making reasoning budget a flexible, model-controlled behavior.

Turing is leading the charge in bridging AI research with real-world applications. Subscribe to AGI Advance for weekly insights into breakthroughs, research, and industry shifts that matter.

[Subscribe & Read More]