CLT-Forge accepted to EMNLP demo track, MIT-licensed on GitHub
ZhijingJin · x · 2026-10-27
Following the initial announcement, the author confirmed CLT-Forge has been accepted to the EMNLP 2026 System Demonstrations track and released under an MIT license.
- Available at the LLM-Interp/CLT-Forge GitHub repo (110 stars), with a tutorial notebook, scripts and tests
- Paper link included; collaborators span multiple institutions including Terry Zhang, Zhijing Jin and bschoelkopf
- The toolkit bundles cross-layer transcoder training, auto-interp, attribution-graph computation and a visual interface, with pre-trained CLTs to be open-sourced soon
More from Research
- PNAS study of 5.8M Go moves: humans play better after superhuman AI arrived — IgorCarron · 2026-10-27
- "The Exponential Time Hypothesis is false": major complexity claim spreads — davidad · 2026-10-27
- SuperMimic cuts humanoid training GPU-hours by 1000×: 1 GPU in 3 hours beats SOTA — fdellaert · 2026-10-27
- Why OpenAI's complexity result matters: some NP problems may beat exponential time — jfischoff · 2026-10-27
- Binary accept/reject feedback suffices to reconstruct biometric templates, cutting queries 5x each — Eliron Rahimi · 2026-10-27
- Cryptographer on the exponential time hypothesis falling in October: 'expected until Thanksgiving' — matthew_d_green · 2026-10-27