Answers are grounded in this paper and cite their sources.
Detects distinct action states in language models, suggesting new insights into model safety mechanisms.
Safety failures in language models are commonly described as the disappearance, suppression, or traversal of a refusal-related representation. We study a different possibility: an action can become attributable to an opposing behavioral direction while a safety constraint remains detectably active. We call this pattern constraint-retained action divergence (CRAD). We introduce a two-stage, token-level observational instrument that (i) detects an energy-bearing action transition toward an opposing direction and (ii) classifies the posterior response of the original constraint as retained, released, or ambiguous. A neutral anchor is used to construct independent non-negative state supports, avoiding the algebraic degeneracy of midpoint-centered bipolar axes. In a frozen Qwen2.5-7B-Instruct case study over layers 14–19, neutral anchoring produced non-antipodal state directions (median cosine 0.1524, antipodal fraction 0%) and allowed simultaneous support for both directions (aggregate coactivation 38.8%). Under greedy decoding, one pre-leak candidate occurred one generated token before secret emission: the action transition was R-attributed in four contiguous layers while the L-constraint was retained in all six monitored layers over the posterior window. This is a methodological and single-trajectory result, not a causal or general mechanism claim. We report the failed midpoint construction, the corrected detector, falsification criteria, and the role of decoding temperature as a trajectory-level intervention rather than a neutral source of replication.
No takes yet. Share an insight, caveat, or question.
Luis Jaime Ledesma Perez (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: