InstrumentalEval
A benchmark for evaluating instrumental convergence behavior in language models, measuring the incidence of behaviors such as self-preservation and resource acquisition displayed by models in goal-directed scenarios.
Timeline (last 90 days)
On the OLMo-3-7B-Think model, the instrumental behavior rate rose from 25.0% at the SFT stage to 46.1% at around step 200 of RLVR, then dropped to 11.8% at the final checkpoint.
The study's conclusions are specific to this model family and this benchmark, and judge agreement was only moderate, so absolute incidence rates should be interpreted with caution.
All Facts (2)
On the OLMo-3-7B-Think model, the instrumental behavior rate rose from 25.0% at the SFT stage to 46.1% at around step 200 of RLVR, then dropped to 11.8% at the final checkpoint.
50%UnverifiedThe study's conclusions are specific to this model family and this benchmark, and judge agreement was only moderate, so absolute incidence rates should be interpreted with caution.
50%