원클릭으로 Manus에서 모든 스킬 실행

steering-block-selection

스타11

포크1

업데이트2026년 7월 11일 04:09

How to choose where (and how many sites) to intervene for any operation on internal representations — activation patching, steering, CAA, DAS, SAE feature scaling, attribution patching. Use whenever the plan declares a target block / layer / site, especially when it hard-codes a single index (`target_block: 4`, `layer: 16`) without justification or copied from another paper. Covers picking the site by screening (gradient / attribution, or an activation signal) and by heuristic (mid-to-late layers usually steer best; widen to 3–5 layers if one is inert; sweep the stack at spaced intervals), and which component to target (attention / MLP / residual stream — circuit discovery studies attention heads + MLP). Prevents two symmetric failures: too few sites → downstream norm/attention denoise it → false "no effect"; too many → off-distribution collapse and the localization claim becomes untestable. Triggers include `target_block = 4`, `sliding window = 15`, `layer ∈ [12,18]`, `early blocks (0-7)`, "steering had no e

설치

Codex 또는 Claude로 설치 이 Prompt를 복사해 Codex, Claude 또는 다른 어시스턴트에 붙여 넣으면 Skill 페이지를 검토하고 설치를 진행할 수 있습니다.

Manus에서 실행

출처

zjunlp

zjunlp/Mechanist

GitHub 저장소 열기 Creator 저장소 보기

다운로드

Manus에서 실행

Steering Block / Layer Selection

The point

Where and how many sites you intervene on changes both the effective strength of the intervention and whether a localization claim is testable.

Too few (single site on a deep stack) → downstream norm/attention denoise it → looks like "no effect".
Too many (whole stack) → off-distribution collapse → the effect is collapse-driven and no longer localized.

Choosing where to intervene

Pick the site(s) by a screening method, by heuristic, or both.

By method (pre-screen the layers):

Gradient-based — rank layers by gradient / attribution score w.r.t. the target.
Activation-based — rank by an activation signal (probe accuracy, diff-mean magnitude, separability).

By heuristic:

Mid-to-late layers usually work best — start there.
If a single layer shows no effect, intervene on 3–5 layers instead.
To sweep the whole stack, use spaced intervals (e.g. every 2–3 layers) rather than every layer.

What to intervene on (transformer)

Target the attention, the MLP, or the residual stream — choose by what the claim is about.

Note: circuit discovery generally studies attention heads + MLP.

Two rules to keep it honest

Match the claim. If the claim is about a region (e.g. "early layers 0–7"), intervene on a window covering that region and use matched windows elsewhere as null controls — a single site can't adjudicate a regional claim. Report the localization you actually observe; don't rewrite the claim to fit the data.
Never copy a raw index across models of different depth. "Layer 16 of 32" → scale by relative depth (0.5 → layer 24 of 48), then check neighbors.

Composition

Lock the site set here first (the coefficient plateau is site-dependent), then sweep α via ../steering-coefficient-tuning/. Re-sweep the coefficient whenever the site set changes.

이 저장소의 다른 Skills

같은 저장소

representation-and-parameter-analysis

zjunlp/Mechanist

Representation and Parameter Analysis interprets and controls a model by directly manipulating its two kinds of internal objects — features (the hidden-state activations produced during a forward pass) and weights (the parameters of the target model). The analysis side relies on standard linear-algebra tools applied straight to these objects: techniques such as PCA, mean-difference, or simple linear classifiers are run over collected features (e.g., from contrastive prompt pairs) or over weight differences between a fine-tuned and a pre-trained checkpoint, in order to extract a small set of meaningful directions. The control side then operates directly through arithmetic on those objects: adding such a direction back into the features steers behavior at inference time, while adding, subtracting, or combining task-level weight differences edits the model in one shot — so that interpretability findings translate into behavioral control without any retraining.

2026-07-1211

steer-features

zjunlp/Mechanist

Use this skill for feature-level steering of models — locating the internal feature that drives a target behavior, scoring and selecting it by its effect on the model's output, and directly amplifying or shrinking that feature's activation during generation to control behavior. Applies to features read from the model's own activations or from a Sparse Autoencoder (SAE); the bundled demo scripts happen to use an SAE, but the method does not require one.

2026-07-1211

mechanism-skills

zjunlp/Mechanist

Routing entry point for eleven families of mechanistic-interpretability methods that localize *which* internal object (layer, attention head, neuron, SAE feature, weight, or input feature) drives a model's behavior, *how influential* it is, and *what changes* when it is intervened on. Use this skill whenever the question is about a model's internal mechanism rather than its external metrics — for example, claims that a specific component is responsible for a behavior, mechanistic-evidence requests, circuit-discovery tasks, feature-attribution work (SHAP), or concept-level explanations of vision/VL models. The file lays out each family's premise, signal, cost, advantages, limitations, and how to compose them into a cheap-screen → causal-verify pipeline. Loading is hierarchical and mandatory: after picking a family from this routing file you MUST load that family's `SKILL.md`, and after picking a submethod you MUST load that submethod's `SKILL.md` — never act on the previews in this file alone.

2026-07-1211

vocabulary-projection

zjunlp/Mechanist

Vocabulary Projection methods interpret internal model states by projecting them through the unembedding matrix to obtain a distribution over the vocabulary. The core idea is that the unembedding matrix, which maps the final hidden state to output logits, can serve as a universal decoder for intermediate states throughout the model. By inspecting the top-ranked tokens of the resulting distribution, researchers can directly read off the semantic content encoded in any internal object — such as a residual stream state or an attention head output — in terms of the model's output vocabulary.

2026-07-1211

ablation-planner

zjunlp/Mechanist

Use when main results pass result-to-claim (claim_supported=yes or partial) and ablation studies are needed for paper submission. The external LLM reviewer (via llm-chat MCP) designs ablations from a reviewer's perspective, CC reviews feasibility and implements.

2026-07-1111

arxiv

zjunlp/Mechanist

Search, download, and summarize academic papers from arXiv. Use when user says "search arxiv", "download paper", "fetch arxiv", "arxiv search", "get paper pdf", or wants to find and save papers from arXiv to the local paper library.

2026-07-1111

name

steering-block-selection

description

Steering Block / Layer Selection

The point

Where and how many sites you intervene on changes both the effective strength of the intervention and whether a localization claim is testable.

Too few (single site on a deep stack) → downstream norm/attention denoise it → looks like "no effect".
Too many (whole stack) → off-distribution collapse → the effect is collapse-driven and no longer localized.

Choosing where to intervene

Pick the site(s) by a screening method, by heuristic, or both.

By method (pre-screen the layers):

Gradient-based — rank layers by gradient / attribution score w.r.t. the target.
Activation-based — rank by an activation signal (probe accuracy, diff-mean magnitude, separability).

By heuristic:

Mid-to-late layers usually work best — start there.
If a single layer shows no effect, intervene on 3–5 layers instead.
To sweep the whole stack, use spaced intervals (e.g. every 2–3 layers) rather than every layer.

What to intervene on (transformer)

Target the attention, the MLP, or the residual stream — choose by what the claim is about.

Note: circuit discovery generally studies attention heads + MLP.

Two rules to keep it honest

Match the claim. If the claim is about a region (e.g. "early layers 0–7"), intervene on a window covering that region and use matched windows elsewhere as null controls — a single site can't adjudicate a regional claim. Report the localization you actually observe; don't rewrite the claim to fit the data.
Never copy a raw index across models of different depth. "Layer 16 of 32" → scale by relative depth (0.5 → layer 24 of 48), then check neighbors.

Composition

Lock the site set here first (the coefficient plateau is site-dependent), then sweep α via ../steering-coefficient-tuning/. Re-sweep the coefficient whenever the site set changes.