Unverified50% confidenceBenchmarkExact time
In the controlled GPT-2 experiments, the unanchored projection update group achieved a target string hit rate of 28.7% under natural language questioning, which dropped to 22.0% after adding module residual anchoring.
1
Sources
50%
Confidence
Long-term
Relevance
10/16/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedIn the controlled GPT-2 experiments, although projection updates were clearly higher in lexical hits, the strict answer score of all experimental groups remained zero.74% similarUnverifiedIn an editing benchmark, the same GPT model's score dropped from 99% to 70% depending on the framework used.73% similarUnverifiedAfter introducing Feature Scope, the verified assertion output rate rose from about 20% to 95%, erroneous feature assertions dropped to 0, and model restraint reached 30/30.72% similarUnverifiedOnly 38.5% of GPT-OSS 120b's capitulations adopted the decoy answer, with the remainder retreating to vagueness, providing a safer failure mode.70% similarUnverifiedSmall language models of about 7 billion parameters (such as allam-2-7b) are markedly weaker than GPT-4-level models in instruction following, long-context reasoning, and self-awareness, and are prone to hallucinating or ignoring negative constraints.68% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1046873API
curl https://kongchang.com/api/v1/knowledge/claims/1046873MCP
get_claim(id=1046873)