Research map / Large language models / Agents, code and math
ADAPT: Hybrid Prompt Optimization for LLM Feature Visualization
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes ADAPT, a hybrid beam-search-initialized, gradient-guided prompt optimizer for feature visualization of Gemma 2 2B SAE latents, which outperforms GCG, BEAST and EPO on activation metrics and human-rated interpretability.
Problem
Feature visualization for LLM activation directions (e.g., SAE latents) requires discrete prompt optimization that is prone to local minima and poorly served by existing adversarial prompt optimizers.
Approach
Combine beam-search initialization (including middle-inserts) with adaptive mixing of GCG gradient-guided swaps and gradient-free logit-swaps, plus group-based diversity-preserving slot selection and a scheduled fluency penalty.
Key results
- ADAPT achieves higher activation than GCG on 70.4% of 486 latents (Wilcoxon p < 1e-28) and beats BEAST on 84.0% and EPO on 78.6%.
- For GCG, activation ratio correlates with human-rated interpretability (Pearson r = 0.645, p < 1e-5), and ADAPT's human ratings are significantly higher than baselines (p < 1e-8).
- Fluency penalties with higher lambda hindered GCG optimization across the hardness-weighted set, and did not significantly improve interpretability (lambda=1.0 vs GCG).
Surprising findings
- Higher fluency penalty weights hinder GCG optimization across the latent set, undermining the hypothesis that fluency smooths the loss landscape.
- GCG gradient estimates have poor signal-to-noise: improvement candidates are nearly uniformly distributed across gradient rank bins, and the estimate is biased toward certain common tokens.
Limitations
- Evaluation limited to residual stream SAEs from a single model family (Gemma 2 2B).
- Human evaluation relies on explanations written by the authors and a small set of 60 latents with three raters.
- Activation ratio is only a proxy for human interpretability.
Experimental setup
Datasets: Neuronpedia Gemma Scope 16k residual stream SAE latents and dataset examples
Benchmarks: Gemma Scope 16k residual stream latents (486 latents across 6 layers) hardness-weighted and balanced 50-latent sets
Metrics: activation ratio vs top maximally activating example activation rank among top 15 MAEs human interpretability rating (1-5) Wilcoxon signed-rank test Krippendorff's alpha
Baselines: GCG BEAST EPO dataset maximally activating examples
Builds on: GCG BEAST Evolutionary Prompt Optimization Gemma Scope
Closest papers by meaning
- Multi-Layer Visual Feature Fusion in Multimodal LLMs: Methods, Analysis, and Best Practices (CVPR 2025 · similarity 0.87)
- Compress to Impress: Efficient LLM Adaptation Using a Single Gradient Step on 100 Samples (NeurIPS 2025 · similarity 0.86)
- Drag-and-Drop LLMs: Zero-Shot Prompt-to-Weights (NeurIPS 2025 · similarity 0.86)
- VIPAMIN: Visual Prompt Initialization via Embedding Selection and Subspace Expansion (NeurIPS 2025 · similarity 0.86)
- IPO: Interpretable Prompt Optimization for Vision-Language Models (NeurIPS 2024 · similarity 0.86)
- Exploring the Role of Large Language Models in Prompt Encoding for Diffusion Models (NeurIPS 2024 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.