paper:circuit-tracing-5e35e4Circuit tracing: Revealing computational graphs in language models
Methods (4)
- Attribution GraphsGradient-based technique using SAE features to estimate causal effects on completions; used to corroborate NLA findings.
- Circuit tracingAnthropic technique revealing computational pathways in Claude.
- Completeness and replacement metricsMetrics of how thoroughly features explain computation.
- Cross-layer transcoder (CLT)Replacement model architecture with 30M features replacing MLP neurons.
Frameworks (1)
- Cross-Layer TranscodersMLP replacement with sparse features underlying the original attribution graphs.
Similar preprints — Semantic Scholar
Cited by (4)
- When Models Manipulate Manifolds: The Geometry of a Counting Task
Claude 3.5 Haiku's linebreaking in fixed-width text is implemented by manipulating low-dimensional, rippled 1-D feature manifolds: character count, line width, characters remaining and next-token leng
- Turn-Averaged SAEs for Feature Discovery and Long-Context Attribution
Training a sparse autoencoder on the mean residual-stream activation of a conversation turn, rather than on individual tokens, yields features that describe a turn's high-level properties more complet
- Persona Features Control Emergent Misalignment
Fine-tuning GPT-4o on narrowly incorrect datasets—spanning nine domains including health, legal, and automotive advice—generalizes into broadly malicious behavior across unrelated prompts, confirming
- Reasoning Theater: Disentangling Model Beliefs from Chain-of-Thought
Reasoning models generate chains of thought that are frequently performative rather than causally necessary for reaching the correct answer: on MMLU recall questions, activation probes decode the mode