zjunlp/Mechanist - 第 2 页

eap-ig

Use this skill to discover and analyze computational circuits within autoregressive transformer language models via attribution methods like Edge Attribution Patching with Integrated Gradients (EAP-IG), enabling mechanistic interpretability studies and…

原文语言：英语

automatic-circuit-discovery

Activates when working on mechanistic interpretability of transformers, specifically for automated circuit discovery in transformer models using computational graph editing and analysis.

原文语言：英语

circuit

Circuit discovery methods localize the minimal subgraph of a transformer's computational graph (a set of nodes such as attention heads and MLPs, together with the edges that connect them) that is causally responsible for a specific model behavior. Rather than…

原文语言：英语

crosscoder-learning

Use this skill when working with sparse autoencoders (SAEs), crosscoders, dictionary learning on neural network activations, training SAEs/crosscoders from scratch, loading pretrained dictionaries, caching model activations, or comparing model internals…

原文语言：英语

ica-lens

Use this skill when applying Independent Component Analysis as a training-free interpretability lens — decomposing a target activation site (residual stream, MLP output, attention-head output, or any cached hook point) into maximally non-Gaussian directions…

原文语言：英语

saelens

Use this skill when working with Sparse Autoencoders (SAEs) for mechanistic interpretability of language models, including training SAEs, loading pre-trained SAEs, analyzing neural network features, or integrating SAEs with TransformerLens, HuggingFace…

原文语言：英语

feature-dictionary-learning

Feature Dictionary Learning methods address the polysemanticity of neuron-level units by decomposing a dense internal activation (e.g. a residual-stream state or MLP output) into a sparse weighted sum of directions drawn from a large over-complete dictionary.…

原文语言：英语

transcoder-circuits

Use this skill when working with transcoder-based circuit analysis of large language models, including training transcoders, analyzing MLP sublayers, reverse-engineering LLM circuits, and creating feature dashboards for interpretability research.

原文语言：英语

inputs-and-layer-wise-states

Analyze and visualize layer-wise gradient behaviors in LLMs during fine-tuning for fast vs slow thinking tasks, calculate gradient statistics, and understand training patterns across different model layers

原文语言：英语

intermediate-outputs

Use this skill when working with circuit discovery in language models, mechanistic interpretability, activation patching, attribution patching, or Layer-wise Relevance Propagation (LRP) for neural network analysis

原文语言：英语

linguistic-regions-llm

Use this skill when working with linguistic region analysis in Large Language Models, including data preprocessing for multilingual training, region-based model training with DeepSpeed, and extracting/visualizing linguistic regions in transformer models

原文语言：英语

gradient-detection

Gradient Detection methods localize influential internal objects by scoring them with the sensitivity of a scalar target $F(x)$ (e.g., a logit, margin, or loss) with respect to an object $o_j$: $s_j(x) = \phi(\nabla_{o_j} F(x), o_j)$, where common…

原文语言：英语

dynamic-components

Identify and manipulate language-specific neurons in multilingual Large Language Models (LLMs) to understand and control language-specific behaviors in models like LLaMA-2, BLOOM, OPT, Mistral, and Phi-2

原文语言：英语

layer-wise-representation

Use this skill when you need to enhance the truthfulness of Large Language Models (LLMs) or reduce hallucinations in model outputs. This skill provides TruthX, an inference-time method that edits LLM internal representations to control truthfulness and…

原文语言：英语

static-parameters

Analyze and manipulate massive values in LLM attention mechanisms, particularly for understanding contextual knowledge processing in transformer models with RoPE

原文语言：英语

clip-dissect

Use this skill when you need to automatically describe or interpret the functionality of individual neurons in deep neural networks (DNNs) using CLIP-based semantic analysis, perform mechanistic interpretability research on vision models, dissect…

原文语言：英语

multi-modal-interpretability

Multi-Modal Interpretability methods explain *vision* and *vision-language* models by linking each internal unit (a convolutional neuron, attention head, residual-stream channel, or SAE feature) to a *natural-language concept* drawn from an external concept…

原文语言：英语

zennit-crp

Use this skill when working with Concept Relevance Propagation (CRP) and Relevance Maximization for explainable AI in PyTorch models, including generating concept-conditional heatmaps, feature visualizations, attribution graphs, and identifying which latent…

原文语言：英语

deep-neural-feature-ansatz

Use this skill when working with the Deep Neural Feature Ansatz (DNFA) — verifying feature learning in neural networks, training fully connected networks on image/tabular datasets, computing Neural Tangent Kernels (NTK) and Neural Network Gaussian Processes…

原文语言：英语

recursive-feature-machines

Use this skill when working with Recursive Feature Machines (RFM) for kernel-based machine learning, feature learning, or when implementing backpropagation-free models that learn features similarly to neural networks.

原文语言：英语

tp4-feature-learning

Use this skill when working with infinite-width neural networks for feature learning, replicating Word2Vec or MAML experiments from the Tensor Programs series (TP4), or implementing infinite-width limits (GP, NTK, muP) for meta-learning and word embedding…

原文语言：英语

convrfm

Use this skill when working with convolutional neural feature ansatz (CNFA), deep convolutional recursive feature machines (Deep ConvRFM), or tasks involving kernel-based feature learning with convolutional architectures, VGG networks, patch-based Jacobian…

原文语言：英语

neural-feature-learning

Neural Feature Learning methods open the black box of *how* a trained neural network discovers useful representations from data, by deriving the directions along which each layer's effective input geometry is reshaped during training. The unifying object is…

原文语言：英语

residual-stream-states

Use this skill when working with LLM hallucination detection, probing internal representations of language models, analyzing model correctness, or conducting intrinsic evaluation experiments on models like Mistral and Llama-3

原文语言：英语

sae-feature-activation-state

Use this skill when working with Sparse Autoencoders (SAEs) for feature analysis, particularly for studying feature splitting, absorption, and attribution in language models. Activate for tasks involving SAE feature ablation, probing experiments, or analyzing…

原文语言：英语

probing

Probing methods interpret model signals by training an auxiliary predictor (often linear) to decode a labeled property y from an internal vector $z$ (e.g., the residual stream state $x_l$ at layer $l$). Operationally, probing treats the model as a frozen…

原文语言：英语

task-vectors

Use this skill when working with task arithmetic for editing neural network models, including creating task vectors from pre-trained and fine-tuned checkpoints, combining them via arithmetic operations (negation, addition, analogies), and applying them to…

原文语言：英语

representation-engineering

Use this skill when working with Representation Engineering (RepE) for AI transparency, monitoring, or controlling internal representations of large language models including truthfulness detection, emotion control, harmlessness steering, and memorization…

原文语言：英语

steering-vectors

Work with Llama 2 model steering using Contrastive Activation Addition (CAA) for behavioral control, including generating steering vectors, evaluating model behavior, and analyzing activation patterns

原文语言：英语

fastshap

Use this skill when you need to train amortized Shapley value explainers using FastSHAP, generate real-time local feature importance explanations for machine learning models (tabular or image), train surrogate models for feature masking, or understand how…

原文语言：英语

shap

Use this skill when working with SHAP (SHapley Additive exPlanations) to explain machine learning model predictions, compute feature importance, generate SHAP values for tree ensembles (XGBoost, LightGBM, CatBoost, scikit-learn), deep learning models…

原文语言：英语

shap

SHAP (SHapley Additive exPlanations) is a unified, game-theoretic framework for local feature attribution. Treating a model's input features as players in a coalition game, SHAP assigns each feature its Shapley value — the average marginal contribution of…

原文语言：英语

attention-head-output

Use this skill when you need to analyze and interpret the internal workings of large language models layer by layer, visualize hidden states and predictions across transformer layers, or understand how models like Llama-3.1-8B and Qwen-2.5-7B make predictions…

原文语言：英语

neuron-value-weight

Analyze transformer feed-forward layers as key-value memories, extract activations, identify trigger examples, and compute key-value agreement in transformer language models

原文语言：英语

residual-stream-state

Use this skill when working with transformer model interpretability, analyzing layer-by-layer predictions, training tuned lenses to understand intermediate representations, or peeking into iterative computations of transformers

原文语言：英语

monitor-experiment

Monitor running experiments, check progress, collect results. Use when user says "check results", "is it done", "monitor", or wants experiment output.

原文语言：英语

next-round

Between-round transition helper for multi-round /auto exploration. Archives the just-finished round's outputs into rounds/round_<N>/ and drafts the next round's task.md with recommended behavior-source/mechanism params, by reading the orchestrator-owned…

原文语言：英语

notify

Draft a research-progress briefing and dispatch it through whatever notification service the user has already configured. Channel-agnostic — this skill only drafts the briefing text and scans for a configured service; it does not hard-code or recommend any…

原文语言：英语

novelty-check

Verify research idea novelty against recent literature. Use when user says "novelty check", "check novelty", or wants to verify a research idea is novel before implementing.

原文语言：英语

research-refine-pipeline

Run an end-to-end workflow that chains `research-refine` and `experiment-plan`. Use when the user wants a one-shot pipeline from vague research direction to focused final proposal plus detailed experiment roadmap, or asks to build a pipeline, do it…

原文语言：英语