CLT-Forge: Open-source toolkit for cross-layer transcoder training and attribution graphs
ZhijingJin · x · 2026-10-27
Zhijing Jin's team released CLT-Forge at EMNLP 2026 (System Demonstrations track, MIT licensed): an open-source library for training cross-layer transcoders and computing attribution graphs to decode computational circuits in LLMs.
- The authors argue mech interp should decode causal computational graphs, not just catalog representations like steering vectors — "you can't understand biology from blood composition alone"
- It's a pioneering open-source effort to scale such analysis to larger models, featuring sharded training, compressed activation caching, auto-interp, and graph visualization
- The attached screenshot shows the interactive attribution-graph interface
- Work done with EuroSafeAI, JinesisLab, MPI, University of Toronto, University of Michigan and others; the GitHub repo has 110 stars
More from Research
- OpenAI Announces Three New Discoveries in Math, CS, and Physics — thesaraharminta · 2026-10-27
- CMU-led team proposes "epiplexity," a new information measure for computationally bounded learners — davidmanheim · 2026-10-27
- Four Gemini agents beat Minecraft autonomously in 27 hours, writing 2,783 programs — rohanpaul_ai · 2026-10-27
- OpenAI PDF claims refutation of ETH, solving 3SAT in exp(n/(log n)^(1/2+o(1))) — rrwilliams · 2026-10-27
- OpenAI Outlines Plan to Build Scientific Understanding with Academia — Chemical-Agency-3997 · 2026-10-27
- Retraction Watch echoes call: pay peer reviewers, don't conscript them — RexDouglass · 2026-10-27