hypothesis
active
hypothesis:finetuning-improves-capabilities-primarily-by-enhancing-existing-circuits-rather-than-creating-new-mechanismsFinetuning improves capabilities primarily by enhancing existing circuits rather than creating new mechanisms
Prakash et al. evidence.
Source paper
extracted_from(2025) · Lee Sharkey · Bilal I. Chughtai · Joshua Batson · Jack W. Lindsey +25
Neighborhood — ranked by edge-count
Papers (1)
paper
Findings (1)
finding
- Supported by further finetuning and causal interventions.
Related by similarity (8)
cosine ≥ 0.65 · no typed edgeEntities in the same semantic neighborhood but without a typed relation to this one — candidates for new edges or unrecognized duplicates.
- Key interpretive conclusion from the dissociation between attempt rate and improvement rate in fine-tuning experiments
- During pretraining the model learns distinct harm categories and during finetuning wires them into general harmful-request features.hypothesis0.749Suggested by dictionaries trained on pretraining data having few refusal features.
- Key claim about the value of the introspection stage, supported by both prefill attack and adversarial prompting experiments
- Conclusion from Experiment 3 and HH intent analysis.
- The projection of the activation change induced by finetuning onto a persona vector direction, used to quantify how much a model shifts along a trait
- Technique used to impose guardrails on base LLMs, analogized to censorship on the simulator's range of simulacra
- Future work hypothesis about extending SOO to direct value alignment
- Claim supported by Perspectives scenario results showing near-100% accuracy post-fine-tuning