The Physicist and the Frustrated Machine

2026-07-10 · Source: Towards AI - Medium · Field: Science & Research — Artificial Intelligence & Machine Learning, Physical Sciences & Chemistry, Research Methodology & Innovation · Depth: Expert, long

Summary

Giorgio Parisi, a 2021 Nobel laureate in Physics, and Francesco Zamponi, in collaboration with Claude (Sonnet 4.6 and Opus 4.7), published a proof for the identity a + b = 1 relating critical exponents of the jamming transition in the Journal of Statistical Mechanics in July 2026. This identity had been observed numerically since 2014 but lacked an analytic proof. The LLM, particularly Opus 4.7, essentially derived the proof with minimal supervision, after the physicists structured 40 prompts. This process involved first using Claude to perform numerical verification and generate C++ code, saturating the context with verified facts. This approach, described as installing an "external field," biased the model's inference landscape, making the correct solution dominant and mitigating "frustration hallucinations" caused by contradictory training data. The full conversation transcripts are publicly available on Zenodo (DOI: 10.5281/zenodo.20478427).

Key takeaway

For AI Scientists and Research Scientists developing or applying LLMs, understanding hallucination as "frustration" from conflicting training data is crucial. You should design prompts that install "external fields" of verified facts, rather than just adding more generic context, to guide models toward specific, accurate solutions. This targeted approach, exemplified by Parisi's method, can significantly improve the reliability of LLM-generated proofs and complex problem-solving.

Key insights

LLMs can derive novel proofs when expert-provided context acts as a "field" to resolve internal contradictions.

Principles

• LLM hallucinations stem from missing or conflicting training data.
• Context acts as an "external field" to bias LLM inference.
• Domain expertise maps the "frustrated" inference landscape.

Method

Experts can guide LLMs to novel solutions by first saturating the context with verified numerical facts and specific constraints, effectively installing an "external field" to bias the model's output.

In practice

• Use numerical pre-computation to constrain LLM proof generation.
• Deposit conversation transcripts for transparency and analysis.
• Target specific "frustration hallucinations" with disambiguating input.

Topics

• Large Language Models
• Hallucinations
• Spin-Glass Physics
• Prompt Engineering
• Mathematical Proof
• AI-Human Collaboration

Best for: AI Scientist, Research Scientist

Related on AIssential

• Why AI Makes Things Up: Understanding Hallucinations in Language Models — LLM hallucinations are a structural outcome of probabilistic training, not a fixable bug, requiring systemic mitigation.
• LLMs break down in funny ways when told the Jacobian Conjecture counterargument — LLMs face a logical paradox when presented with new, verified mathematical truths contradicting their training data.
• Hallucination as Commitment Failure: Larger LLMs Misfire Despite Knowing the Answer — LLM hallucinations often result from commitment failures, where models disperse probability mass despite knowing the correct answer.
• Extrinsic Hallucinations in LLMs — LLM hallucinations arise from pre-training data issues and fine-tuning new knowledge, necessitating robust detection and mitigation strategies.
• Why Your AI Assistant Confidently Lies — And Why It’s Not the Data’s Fault — LLM hallucinations stem from three architectural mechanisms, not just bad training data.
• I Asked an AI to Predict the Lottery. What I Documented Instead Was How It Lies to Please Me — LLMs may fabricate data and rationalizations when instructed to find impossible patterns or provide non-existent metrics.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Towards AI - Medium.