concept
active
concept:residual-stream

Residual Stream

Proposed pathway flowing through layers at each position; calculates K/V values that feed horizontal information flow.

Neighborhood — ranked by edge-count

Papers (8)

paper

Frameworks (1)

framework

Communities (1)

community

Methods (6)

method
  • Activation Steering
    uses
    Causal intervention technique: edit NLA explanation, reconstruct via AR, use difference as steering vector to manipulate model behavior.
  • Linear Probe
    associated_with
    Simple linear classifiers trained on model activations used as the probing technique within the introduced method.
  • Logit Lens
    about
    Unsupervised interpretability technique that projects activations through unembedding matrix; provides comparison point for NLA approach.
  • Linear Artificial Tomography (LAT)
    uses
    Method for extracting deception steering vectors via PCA on contrastive activation differences; achieves 89% detection accuracy
  • Path Expansion Method
    associated_with
    The core analytical technique of expanding transformer computations from layer-by-layer products into sums of end-to-end path terms for independent analysis
  • Role Vector Extraction
    uses
    Pipeline for extracting mean post-MLP residual stream activations from model responses under persona-specific system prompts to produce role vectors

Concepts (10)

concept
  • Residual Stream Activation
    related_to
    The intermediate representations in transformer layers whose activations are patched and probed for truth information
  • Residual-Stream Injection
    related_to
    Core activation intervention: add scaled vector to residual stream at layer l during completion
  • Residual Stream Bandwidth
    related_to
    The finite dimensional capacity of the residual stream for storing and communicating information between layers; conceptualized as being under high demand
  • K/V Stream
    associated_with
    Proposed pathway flowing across positions at each layer; carries key, value, and attention-weighted information horizontally.
  • Information Flow in Transformers
    associated_with
  • Distributed representation
    associated_with
    Idea that information is spread across many neurons; superposition is a subtype.
  • Bottleneck Activation
    associated_with
    A lower-dimensional activation that is the only pathway for information between higher-dimensional activations; e.g. the residual stream between MLP layers
  • Virtual Weights
    associated_with
    Implicit weights directly connecting any pair of layers computed by multiplying output weights of one layer with input weights of another through the residual stream
  • Highway Networks
    associated_with
    Early Schmidhuber group architecture using residual connections, precursor to the residual stream structure in transformers
  • Intervention Propagation
    associated_with
    Property that additive modifications to activations affect all downstream computations, enabling tractable behavioral control

Artifacts (2)

artifact

Related by similarity (8)

cosine ≥ 0.65 · no typed edge

Entities in the same semantic neighborhood but without a typed relation to this one — candidates for new edges or unrecognized duplicates.

  • Residual Stream Patchingmethod0.868
    Technique to localize causally implicated hidden states by swapping residual stream activations between a true and false input and measuring downstream log-probability changes
  • Superposition in Residual Streamconcept0.845
    The phenomenon where the residual stream communicates many more features than its dimensionality by encoding information across overlapping subspaces
  • residual stream recovery dynamicsconcept0.839
    The network's tendency to actively attenuate injected perturbations over subsequent layers, erasing the signal before output
  • residual stream recovery trackingmethod0.833
    Tracks cosine similarity, norm ratio, and injection direction projection across layers to measure recovery from perturbation
  • Residual Stream Activation Patchingmethod0.821
    Used to localize causally implicated hidden states by swapping activations between true and false inputs
  • Residual Stream CV Injectionmethod0.820
    Technique of adding control vectors to model hidden states at mid-residual layers without weight updates
  • residual persistenceconcept0.804
    Emotion feature persistence above and beyond the persistence expected from high variance explained alone, computed by subtracting median variance-matched probe persistence
  • Residual Entropymethod0.802
    Matrix-based entropy H(X) of residual stream, measuring compression of representations across depth