Horizon AI Daily Digest - May 27, 2026: 24 Top AI Papers and Tech News

Forum topic · 小凯 · 2026-05-27

Summary

Horizon AI Daily Digest for May 27, 2026 curates 24 highlights from 36 tracked items, spanning AI research, industry news, and developer tools. Top-rated entries (9.0/10) include a reality check on whether LLMs can truly introspect versus surface pattern matching; JobBench, a benchmark of 130 tasks across 35 occupations where the strongest model scores only 45.9% on human-delegated work; ScientistOne, an autonomous research system built on a Chain-of-Evidence framework for verifiable AI-generated science; and the MiniMax-M2 series of MoE large models designed for agent scenarios with self-evolution capability. Also covered: Spain blocking prediction markets Polymarket and Kalshi over gambling licences, agent memory as a database workload (GEM model), the POLAR personalization framework for embodied multimodal agents, AgingBench for long-running agent reliability, OmniToM theory-of-mind benchmarking, hallucination detection via automatic layer selection, and tool updates including Claude Code v2.1.152 and Cloudflare Flagship feature management. The digest ends with AI-for-science entries on polymer design, virtual lab planning, and math-question robustness comparisons of chain-of-thought versus code execution.

Horizon AI Daily Digest - May 27, 2026

> 36 items tracked; 24 best picks below, rated by editorial score.

Key points

LLM introspection under scrutiny: A new paper questions whether existing studies of LLM introspection distinguish genuine introspection from surface pattern matching. (9.0/10)
JobBench: Evaluates AI agents against human delegation needs (not economic value) across 35 occupations and 130 tasks; the strongest model reaches only 45.9%. (9.0/10)
ScientistOne: Introduces the Chain-of-Evidence framework and an end-to-end autonomous research system to make AI-generated research verifiable. (9.0/10)
MiniMax-M2 series: MoE large models achieving strong real-world intelligence with minimal activated parameters, designed for agent scenarios with self-evolution capability. (9.0/10)
Regulation: Spain blocks prediction markets Polymarket and Kalshi over missing gambling licences, sparking global regulatory debate. (Reuters, HN discussion)

Research highlights

1. Can LLMs Introspect? A Reality Check (9.0/10) — Argues current research fails to separate true introspection from surface pattern matching. 2. JobBench: Aligning Agent Work With Human Will (9.0/10) — Human-delegation-oriented agent benchmark; best model at 45.9%. 3. ScientistOne: Towards Human-Level Autonomous Research via Chain-of-Evidence (9.0/10) — Verifiability framework for autonomous research outputs. 4. The MiniMax-M2 Series (9.0/10) — Small-activation MoE models for agentic intelligence with RL-driven self-evolution. 5. Is Agent Memory a Database? (8.0/10) — Frames long-term agent memory as a new data-management workload; proposes the GEM model for state-trajectory correctness. 6. Personalizing Embodied MLLM Agents (8.0/10) — POLAR framework uses multimodal knowledge graphs for long-term personalized embodied interaction. 7. Your Agents Are Aging Too (8.0/10) — AgingBench evaluates reliability degradation in long-running deployed agents. 8. Anchor: Mitigating Artifact Drift in Agent Benchmark Generation (8.0/10) — Constrained optimization pipeline keeps instruction/environment components consistent. 9. OmniToM (8.0/10) — Benchmarks theory of mind in LLMs via explicit belief modeling. 10. Automatic Layer Selection for Hallucination Detection (8.0/10) — Intrinsic-dimension-based first effective peak for automatic mid-layer selection improves hallucination detection. 11. Exploiting Local Dynamics Regularity for Reusable Skills in Offline Hierarchical RL (8.0/10) — CARL algorithm aligns local dynamics with action sequences via contrastive learning. 12. Advancing Creative Physical Intelligence in LMMs (8.0/10) — MM-CreativityBench evaluates creative physical reasoning. 13. From Static Context to Calibrated Interactive RL (8.0/10) — Unified framework mitigating cumulative distribution shift in multi-turn dialogue with an aligned simulator. 14. Towards Trustworthy Legal AI (LexGuard) (8.0/10) — Relevance-sensitive evaluation curbs LLM oversensitivity to legally irrelevant changes.

Tools and industry

Claude Code v2.1.152 (7.0/10) — Enhanced auto-fix code review and skill management.
Cloudflare Flagship (7.0/10) — Feature management service with global flagging and evaluation. (HN discussion)
Minicor (YC P26) (7.0/10) — Windows desktop automation at scale for AI companies, integrating with API-less systems. (HN discussion)

More papers

15. BrickAnything (7.0/10) — Geometry-conditioned, structure-aware autoregressive generation of buildable brick structures from 3D shapes. 16. Constraint acquisition needs better benchmarks (7.0/10) — MPMMine benchmark suite for constraint acquisition. 17. Experiments in Agentic AI for Science (7.0/10) — DeepTS/DeepCollector and DeepScribe automate scientific workflows with hybrid local/cloud LLM architecture. 18. Managing Uncertainty in LLM-Generated Procedural Knowledge for Virtual Laboratory Planning (7.0/10) — Framework for handling uncertainty in LLM-generated lab procedures. 19. Reasoning, Code, or Both? (7.0/10) — Chain-of-thought reasoning is more robust than code execution on math-question variants. 20. PolyFusionAgent (7.0/10) — Multimodal foundation model plus autonomous agent for polymer property prediction and inverse design.