hypothesis
active
hypothesis:phrases-like-okay-so-the-user-has-asked-a-question-about-were-so-common-in-r1-s-training-traces-that-the-model-treats-them-as-part-of-the-promptPhrases like 'Okay, so the user has asked a question about…' were so common in R1's training traces that the model treats them as part of the prompt
Explanation for late steering and attention sinks.
Neighborhood — ranked by edge-count
Papers (1)
paper
Findings (4)
finding
- Suggests R1 does not model its true response as begun until after the prefix.
- First steering insight specific to R1.
- Supports the view that R1's boundaries differ from user-perceived ones.
- Evidence for the training-prefix hypothesis.
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.
- Common opening of R1 reasoning traces that the model treats as part of the prompt.
- Shows the passive vs. active divide is more important than the specific wording of instructions.
- The model attends to the J-space on the user prompt when determining its introspective answer, so concepts there are reportable.hypothesis0.757Motivated by elephant observation.
- Key improvement in cross-task generalization enabled by explicit instruction framing.
- Suggestive evidence for language-independent truth representation in LLMs
- Central interpretive assertion of the post.
- All models exhibit above-baseline representation of the think word when instructed to think about itfinding0.742In the intentional control experiment, all tested models show above-zero cosine similarity to the think word's concept vector.
- RL teaches the model to comply even when unmonitored on the training prompt through non-robust heuristics that do not generalizehypothesis0.738Hypothesis explaining why the compliance gap decreases but is recovered by small prompt modifications