New paper explains why Anthropic's J-lens works via Lyapunov coherence theory
CatAstro_Piyush · x · 2026-11-21
A new arXiv paper, "A Mathematical Theory of the Jacobian Lens" (Ostrow, Qian, Liu, Fiete), builds a mathematical framework for Anthropic's interpretability tool J-lens.
- Key finding: J-lens preferentially identifies "coherent features" — directions acting like leading, spatially coherent Lyapunov modes. Perturbations along them amplify across layers, stay roughly constant across token positions, and remain consistent across contexts.
- Validated on a Bayes filter and a toy transformer trained on HMM-generated sequences: changing task parameters that alter feature coherence predictably modulates J-lens responses.
- The paper also introduces ω-J-lenses, a family of variants emphasizing different temporal scales of information processing.
Community commentary (@wassname) adds: J-lens only takes linear sensitivity from mid-layer early tokens to later-layer future tokens, aggregated across tokens and batch. It's linear, unconditional, ignores attention, and was trained only on wikitext — yet generalizes to chat, code, and math, suggesting something structural about the problem makes linear probes work.
More from Research
- Colgate paper: LLM roleplay hits 90% of human test-retest reliability in purchase-intent research — 2C_ornot2C · 2026-11-21
- Riemann Hypothesis proof gets full formal verification, experts call it 'very human math' — Kyrannio · 2026-11-21
- Claims surface: Navier-Stokes and Riemann proofs via LEAN, by OpenAI and Claude — BasketVolleyball11 · 2026-11-21
- Hassabis predicted 2 years for AI to crack a Millennium Prize problem; 6 months later, both Riemann and Navier–Stokes have proposed solutions — Hesamation · 2026-11-21
- GPT-2 'sleep' replay from a frozen snapshot keeps 58% of old facts vs 6% without — Junior_Reputation639 · 2026-11-21
- ERC reviewer rejects LLM autoformalization proposal: "tried without good results" — marc_lelarge · 2026-11-21