How AI Discovered Hidden Information in Light

· Source: Discover AI · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems · Depth: Expert, extended

Summary

The "Briefcase" benchmark, released June 18, 2026, by Artificial Analysis, evaluates LLMs and VLMs on 91 multi-week knowledge work tasks, including data science and product management. It uses an Elo rating system, with Fable 5, Opus 4.8, GLM 5.2 max, and GPT 5.5 X high among top models. The benchmark reveals significant cost differences (Fable 5 at \$31 per task vs. GPT 5.5 high at \$3.68) and low factual accuracy, with Fable 5 achieving only a 3% 100% pass rate, primarily due to "incorrect reasoning." Research from Chinese University of Hong Kong, Shenzhen (June 25, 2026) indicates LLMs often rely on statistical pattern matching over deductive reasoning, leading to hallucinated causal paths. Concurrently, a multi-agent "Socratic" AI system from Hangzhou Institute of Advanced Study and Shanghai Institute of Optics (June 25, 2026) demonstrates autonomous scientific discovery in quantum optics. This system, using a critique agent to enforce physical laws, autonomously discovered that "speckle" patterns from multi-mode fiber imaging are high-dimensional feature representations, enabling object classification from chaotic optical data.

Key takeaway

For AI Scientists evaluating LLM capabilities for complex reasoning or designing autonomous discovery systems, recognize that current LLMs often rely on statistical pattern matching, leading to hallucinated deductive paths and low factual accuracy on real-world tasks. You should prioritize multi-agent architectures incorporating explicit physical or logical constraints, like the Socratic critique agent, to achieve reliable scientific discovery. This approach can transform seemingly chaotic data, such as optical speckle patterns, into valuable high-dimensional feature representations for robust classification and analysis.

Key insights

LLMs struggle with true deductive reasoning, but multi-agent Socratic AI can achieve autonomous scientific discovery by enforcing physical laws.

Principles

Method

A multi-agent system with a "critique agent" interrogates a "planner" agent, forcing hypotheses to satisfy physical laws and generating counter-examples, minimizing reasoning uncertainty.

In practice

Topics

Best for: AI Scientist, Research Scientist

Related on AIssential

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Discover AI.