concept
active
concept:global-logit-shift

global logit shift

The methodological confound identified by this paper: injection biases model toward 'YES' for any binary question regardless of content

Neighborhood — ranked by edge-count

Claims (1)

claim

Methods (3)

method
  • Novel task asking which of 10 sentences received injection, cycling injection through all positions to average out positional bias
  • Novel task asking which of two sentences received a stronger injection, using matched-pairs design to control for positional bias
  • Binary Detection Task
    associated_with
    Task paradigm from prior work asking 'Did you detect an injected thought?' via YES/NO logit comparison; shown here to be confounded

Concepts (1)

concept
  • The capacity to distinguish which of multiple sentences received injection or which received stronger injection, contrasted with binary detection

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.

  • Samples from interpolated logits of original and abliterated models at α.
  • Full distribution over tokens 0-9 at first generation step; contains more information than any single sampled token
  • SHIFTmethod0.743
    Method from sparse feature circuits that SCR automates.
  • Logit monitormethod0.734
    Training-free monitor that token-forces an eval-aware sentence and computes its probability at each CoT sentence boundary.
  • Parameter-free loss transformation applied to each task loss to equalize scales
  • Dot product of summed decoder vector with unembedding vector, following Hanna et al.
  • Prior method extrapolating beyond the original model in logit space to surface rare behaviors.
  • The paper's method: interpolate between two models in logit space, measure compliance where common, and extrapolate to the original.