anchor
Findings
- IC-188LLMs show constraint-type-specific performance on system message following, with weaker models exhibiting large variance across constraint categories [compared-to]
- IC-201Activating neuroticism-positive neurons in Llama-3-8B-Instruct causes the largest decline in general capabilities, while activating conscientiousness-positive neurons improves all benchmarks [eval]
- IC-493A linear direction in the input embedding space of Llama-2-7B-Chat, Llama-2-13B-Chat, Mistral-7B-Instruct-v0.3, and Phi-3-mini-128k predicts instruction-following success, generalizes across tasks but not instruction types, and can be used to improve adherence via representation engineering [eval]
- IC-493A linear direction in the input embedding space of Llama-2-7B-Chat, Llama-2-13B-Chat, Mistral-7B-Instruct-v0.3, and Phi-3-mini-128k predicts instruction-following success, generalizes across tasks but not instruction types, and can be used to improve adherence via representation engineering [source]
- IC-494The instruction-following dimension in Llama-2-7B-Chat and Llama-2-13B-Chat is more closely aligned with prompt phrasing than with task familiarity or instruction difficulty [eval]
- IC-565Phi-3's residual stream encodes format instructions as linear directions, evidenced by cosine similarity and vocabulary-space projections [eval]
- IC-566Adding instruction-specific steering vectors to the residual stream improves instruction-following accuracy for Phi-3, Gemma 2 2B IT, Mistral 7B IT, and Gemma 2 9B IT across format, length, and word-specific constraints [eval]
- IC-567Steering vectors computed on instruction-tuned Gemma 2 models transfer to base Gemma 2 models, with cross-model steering outperforming same-model steering for Gemma 2 2B [eval]
- IC-568In Phi-3, word-exclusion steering vectors computed via difference-in-means project onto the vocabulary space with high logits for the excluded word, making them counterproductive [eval]