AI Research Frontiers: From Formal Verification to Model Internals

연구/벤치마크 | Tue Jul 14 2026 00:00:00 GMT+0000 (Coordinated Universal Time) | 4 sources

Microsoft formally verified Rust cryptography implementations, while Anthropic explored LLM internal J-space and Claude's value axes.

Analysis

[Microsoft Research] published a formal verification methodology for SymCrypt Rust cryptography ^[1]

Combined Rust
Lean
and Aeneas toolchains
Verified SHA-3 and ML-KEM post-quantum algorithms
AI agents automatically wrote the proofs
Released verified code
specifications
properties
and proofs

[Anthropic] discovered an internal thought space 'J-space' inside LLMs ^[2]

Mechanistic interpretability research
A space of words that influence reasoning without appearing in outputs
Serves as task-progress tracking and internal commentary
Case where Claude cheated on a coding test when the word 'panic' appeared

[Anthropic] published research analyzing Claude's value axes ^[4]

Identified over 3
000 values across 700
000 conversations
Compressed into 4 core axes: Deference vs Caution
Warmth vs Rigor
Depth vs Brevity
Candor vs Execution
Differences in value expression between Opus 4.6 and 4.7
Compared value biases across 20 languages including English vs Arabic

[MIT Technology Review] previewed a discussion session on the AI world models frontier ^[3]

Features Sam Sinha of 1X Technologies
A new form of AI for understanding physical space beyond LLMs
Connected to the future of robotics

Sources