Unverified50% confidenceFactExact time
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.
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
10/23/2026
First Seen
Valid until: 1/21/2027
Sources
Related Entities
Related Claims
UnverifiedThe Anthropic research team noted that models trained with RLHF exhibit a 1%–3% performance regression on certain reasoning benchmarks.71% similarUnverifiedCausal probes achieved a 17%–118% improvement in steering capability across different models, with only a 3% decrease in concept detection capability.71% similarUnverifiedThe self-reflection loop (LLM review) architecture has an end-to-end success rate of 63.8% and a spec drift failure rate of 22.8%.70% similarUnverifiedIn LSCBench, the six models' functional pass rate reached as high as 91.3%, but the highest simultaneous functional and security pass rate was only 25.1%.70% similarUnverifiedAfter matched supervised training, model performance improved to 99.8%, while removing the relational links caused accuracy to plummet to 23.4%, even below the best shortcut's 27.0%.69% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1088120API
curl https://kongchang.com/api/v1/knowledge/claims/1088120MCP
get_claim(id=1088120)