New theory explains why Anthropic's J-lens finds working-memory-like features in LLMs
burny_tech · x · 2026-10-27
A new arXiv paper by Mitchell Ostrow, Daoyuan Qian, Toni Liu, and Ila Fiete builds a mathematical theory of Anthropic's Jacobian lens (J-lens) interpretability tool.
- The paper shows J-lens preferentially detects "coherent features" — directions acting like leading, spatially coherent Lyapunov modes, where small perturbations amplify across layers, stay roughly constant across token positions, and remain consistent across contexts.
- Validated on a Bayes filter and a toy transformer optimally predicting HMM-generated tokens; modulating task parameters that change feature coherence predictively modulates J-lens response.
- The authors also introduce ω-J-lenses, a family of lenses emphasizing different temporal scales of information processing in LLMs.
27 pages, 19 figures, posted Oct 2026.
Related event: MIT Researchers Build Mathematical Theory for Anthropic's Jacobian Lens(3 posts)→
More from Research
- Researcher argues DoM steering is no more transparent than gradient-based vector selection — johnhewtt · 2026-10-27
- MPI's Alien Recombination method finds concept blends beyond human cognitive reach — profjamesevans · 2026-10-27
- Podcast: James Evans on why reasoning models simulate societies of thought — profjamesevans · 2026-10-27
- PikasekoHash-256: experimental AVX-512 hash function hits ~24 GB/s on one Zen 5 core — ssh4net · 2026-10-27
- Anthropic folder with possible second ETH refutation proof spotted in shared TCS repository — NoahSD · 2026-10-27
- Why an AI-disputed ETH matters: subexponential n/√log₂n scaling vs 2ⁿ, huge savings at large n — robleclerc · 2026-10-27