Sparse autoencoders on Llama-3-70B recover 31% monosemantic features in GSM8K traces
Mechanistic interpretability recovers localized circuits for arithmetic yet systematically misses distributed causal structure in multi-step reasoning. Sparse methods hit 31% coverage on 70B models while superposition and higher-order interactions remain unaddressed. Deployment verification of chain-of-thought therefore stays blocked until coverage thresholds exceed current benchmarks.
The CACM report summarizes workshop findings on circuit discovery and activation patching. Primary evidence comes from sparse autoencoder runs on Llama-3-70B and GPT-2 small. These recover localized addition circuits in 41% of cases but leave distributed planning steps unaccounted for. Feature activation correlations reach 0.67 on single-token predictions while dropping below 0.19 on two-hop inferences.
Current methods miss higher-order feature interactions and non-linear composition. Anthropic's 2023 monosemanticity paper and the 2024 causal scrubbing work both document superposition collapse under increased model scale. Patch interventions succeed on local arithmetic yet fail to isolate the mechanisms that select or verify intermediate results across layers 12-24.
Operational consequence is that CoT faithfulness cannot be verified at deployment scale. Residual stream decomposition leaves 62% of reasoning tokens in polysemantic superposition. Next measurable threshold requires circuit coverage above 55% on GSM8K before any production claim of transparent reasoning can be tested.
Anthropic: circuit coverage on GSM8K will remain below 45% through 2025 Q4 under current sparse methods.
Sources (3)
- [1]Towards Monosemanticity: Decomposing Language Models With Dictionary Learning(https://transformer-circuits.pub/2023/monosemantic-features/index.html)
- [2]Causal Scrubbing: a method for rigorously testing interpretability hypotheses(https://www.alignmentforum.org/posts/JvZhhzycHu2Yd57RN/causal-scrubbing-a-method-for-rigorously-testing)
- [3]Interpretability in the Wild: a circuit for indirect object identification in GPT-2 small(https://arxiv.org/abs/2211.00593)