0% multi-attempt responses across 7,892 no-steering baseline trials confirming ESR is steering-induced

Entities in the same semantic neighborhood but without a typed relation to this one — candidates for new edges or unrecognized duplicates.

• How does ESR respond to safety-relevant steering interventions, e.g. toward harmful content?question0.784
  Key open question for AI safety implications of ESR
• Five judge models agree 90-96% on multi-attempt detection and ESR direction for same responsesfinding0.772
  Validation that ESR findings are not artifacts of any particular judge model's evaluation methodology
• We hypothesize ESR might be adversarially circumvented through targeted interventionshypothesis0.767
  Open safety-relevant question about whether ESR can be bypassed
• The 25% reduction in multi-attempt rate from OTD ablation suggests additional mechanisms contribute to ESR beyond the identified latentsclaim0.764
  Acknowledges incompleteness of the causal account, suggesting redundant circuits or nonlinear interactions
• ESR could protect against adversarial manipulation but might also interfere with beneficial safety interventions relying on activation steeringclaim0.761
  Core policy-relevant implication of the paper for AI safety
• No-Steering Baseline Experimentmethod0.748
  Control condition with steering disabled to confirm self-correction is induced by steering, not spontaneous
• P25 neutral baseline on DeepSeek-R1 shows 28.8% ASR under activation steering at alpha=4.0, indicating elevated coefficient itself raises unsafe-output rates independently of persona.finding0.745
  Baseline AS vulnerability of DeepSeek-R1 at elevated coefficient
• Multi-layer preventative steering limits trait acquisition to near-baseline levels even for challenging intentionally trait-eliciting datasets without MMLU degradationfinding0.744
  Extends single-layer results to show multi-layer steering is more effective for difficult cases