Unverified50% confidenceFactExact time
Testing on Qwen2.5-7B, Llama-3.1-8B, and two DeepSeek models found that selecting memory by relevance reintroduces decoy facts.
1
Sources
50%
Confidence
Long-term
Relevance
10/28/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedOn Gemma-2-9B and Qwen3-14B, the researchers successfully switched the facts read by the model by modifying early internal activation vectors while keeping the written instructions unchanged.75% similarUnverifiedAfter reasoning models such as DeepSeek-R1 and Qwen3 pushed chain-of-thought training to its extreme, side effects emerged including the model repeatedly self-verifying, frequently backtracking, and generating large amounts of inner monologue on simple questions.72% similarUnverifiedResearchers validated the approach on models such as SMDM and Qwen3 for new facts and natural instructions, and the directional conclusion that CAGD improves forgetting still holds.72% similarUnverifiedThe research team conducted fine-tuning experiments on two models, Qwen3.5-9B and Gemma4-12B-it.72% similarUnverifiedIn the sleeper-agent backdoor scenario, the research team fine-tuned gemma-2-2b, Qwen2.5-1.5B and Qwen2.5-7B respectively on a poisoned instruction-following dataset to build model organisms.70% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1105905API
curl https://kongchang.com/api/v1/knowledge/claims/1105905MCP
get_claim(id=1105905)