hypothesis
active
hypothesis:r1-implicitly-recognizes-confusion-or-incoherence-when-activations-are-overly-perturbed-causing-it-to-stop-and-course-correctR1 implicitly recognizes confusion or incoherence when activations are overly perturbed, causing it to stop and course-correct
Working hypothesis for the oversteering reversion.
Neighborhood — ranked by edge-count
Papers (1)
paper
Findings (1)
finding
- Oversteering some R1 features makes the model revert to original behavior before outputs become incoherentassociated_withsupportsSecond steering insight, unlike non-reasoning models.
Concepts (1)
concept
- Rebalancing effectcitesProposed self-correction when activations are overly perturbed.
Hypotheses (1)
hypothesis
- Why rebalancing would appear in reasoning models specifically.
Related by similarity (8)
cosine ≥ 0.65 · no typed edgeEntities in the same semantic neighborhood but without a typed relation to this one — candidates for new edges or unrecognized duplicates.
- Central interpretive assertion of the post.
- Evidence for the training-prefix hypothesis.
- Attention sinks in R1 appear toward the end of the response prefix rather than at the start of responsesfinding0.757Suggests R1 does not model its true response as begun until after the prefix.
- Human data fine-tuning effect is distinct from synthetic emergent misalignment and likely caused by off-policy training
- Illustrates how even simple AI properties can be informative as sensitive negative-direction indicators
- The paper's central mechanistic explanation of why narrow fine-tuning causes broad misalignment
- Disanalogy with brains.
- First steering insight specific to R1.