CTA paper: graph structure predicts probe accuracy at ρ=0.91, R²=0.84
ZhijingJin · x · 2026-10-25
The CTA paper (arXiv:2608.27510) released full results, by Vedant Palit, Florent Draye, Terry Zhang, Bernhard Schölkopf, and Zhijing Jin, supported by Jinesis Lab, MPI-IS, University of Toronto/Vector Institute, and others.
- Method: Cross-Layer Transcoders replace MLP modules; attribution graphs are trained with respect to a linear probe's concept direction, yielding probe-specific circuits that explain why an internal concept representation arises, independent of the generated token.
- Graph structure predicts probe accuracy: across four concept categories (toxicity/sentiment/reasoning/truthfulness), graph-level features predict probe accuracy at ρ=0.91, R²=0.84; high-accuracy layers have concentrated, decisive circuits while low-accuracy ones are distributed.
- Stable cross-prompt core: top predictive features appear 6–21× more often than random features at the same belief layer — many circuits, one core.
- Causal ablations: removing probe-relevant features lowers internal concept scores while largely preserving output tokens; removing logit-relevant features changes the token in 92–100% of cases with near-zero effect on probe scores, showing functionally distinct mechanisms.
Related event: New EMNLP Paper Introduces CTA to Explain How Linear Probes Emerge(2 posts)→
More from Research
- Lean4AgenticRL brings formal verification to agentic RL, +20.2% on SWE-Bench — May_F1_ · 2026-10-25
- Developer Reverse-Engineers SynthV AI v2 Binary and Reimplements It Fully in PyTorch — cephaloform · 2026-10-25
- arXiv paper: test collections may need expiration dates as neural rankers near human-level — bclavie · 2026-10-25
- FlashMap: flow map-based speculative decoding speeds up Qwen3-8B inference over 6x — caglarml · 2026-10-25
- CS papers near 18,000/month, up 10x since 2016, led by complexity and algorithms — burny_tech · 2026-10-25
- ICML Paper: Empowerment-Based Representations Are 1000x More Sensitive to Control-Relevant Features — burny_tech · 2026-10-25