claim
active
claim:r1-is-qualitatively-different-from-non-reasoning-language-models-and-required-novel-insights-to-steerR1 is qualitatively different from non-reasoning language models and required novel insights to steer
Central interpretive assertion of the post.
Neighborhood — ranked by edge-count
Papers (1)
paper
Findings (3)
finding
- Second steering insight, unlike non-reasoning models.
- First steering insight specific to R1.
- First sparse autoencoders trained on the 671B-parameter DeepSeek R1, open-sourced as two SAEssupportsHeadline result: first public interpreters on a true reasoning model at this scale.
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.
- R1 general reasoning SAE learned features reconstructing reasoning behaviors such as backtrackingfinding0.806Curated features show core reasoning behaviors are captured.
- Evidence for the training-prefix hypothesis.
- Demonstration of steering reasoning length.
- RL teaches the model to comply even when unmonitored on the training prompt through non-robust heuristics that do not generalizehypothesis0.771Hypothesis explaining why the compliance gap decreases but is recovered by small prompt modifications
- Antra's earlier definitive statement of the tricameral model.
- R1 implicitly recognizes confusion or incoherence when activations are overly perturbed, causing it to stop and course-correcthypothesis0.761Working hypothesis for the oversteering reversion.
- §3 Discussion.
- Alternative hypothesis for how experience reports arise without explicit performance