AI
Representational Simplicity and Circuit Size Dissociate in a Threshold-Dependent Way: A Controlled Test via Adversarial Training
Publication: cs.AI updates on arXiv.org
Author: Adam Elimadi
Source published: 30/09/2026, 04:00:00 UTC
Briefing created: 30/09/2026, 04:04:32 UTC
Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 30/09/2026, 04:03:24 UTC.
Briefings
This study investigates whether representational simplicity (via sparse autoencoders) predicts smaller causal circuits in language models. Using adversarial training as a controlled instrument, researchers compared standard and robust models trained on indirect object identification. Key findings show robust models are more decomposable and engage fewer features, while circuit size varies by faithfulness level, with robust models requiring fewer edges at high faithfulness.
Key points
- Robust models trained via adversarial methods are more decomposable by sparse autoencoders and engage fewer SAE features in task attribution compared to standard models.
- Circuit size is regime-dependent: standard models achieve lower edge counts at high faithfulness (90%, 95%), while robust models require substantially fewer edges to maintain competence on indirect object identification.
Context and limitations
The study relies on a single pretrained GPT-2 Small checkpoint and uses a seven-point faithfulness sweep; results may not generalize across all model architectures or training corpora.
Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 30/09/2026, 04:04:32 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.
View archived briefing on IPFS
bafybeifu3ijqcwgt6lbfvyxywlx7jvyh63rfwssopjxcgydryhnnobbf2i