I’ve deep-dived into this issue from the perspective of LLM attention mechanisms and RAG, and compiled a technical analysis along with my feedback record to Google on Hatena Blog. (Note: The forced prompt itself is kept private to avoid server overload).
https://t.co/OwBZu0iCqR
I’ve deep-dived into this issue from the perspective of LLM attention mechanisms and RAG, and compiled a technical analysis along with my feedback record to Google on Hatena Blog. (Note: The forced prompt itself is kept private to avoid server overload).
https://t.co/OwBZu0iCqR
I suspect the recent update that adds citation anchors for Knowledge in the chat is causing "context pollution." It feels like the AI is re-reading its own anchors, strengthening self-reference and triggering the endless loop.
@GeminiApp Regarding the Gemini Gem 1076 error: When the context gets long, is the reference to the Gem's Knowledge working too strongly, causing it to stop reading the rest of the chat context? (Is it stuck in an infinite loop where the Knowledge referencing never finishes?)
I’ve deep-dived into this issue from the perspective of LLM attention mechanisms and RAG, and compiled a technical analysis along with my feedback record to Google on Hatena Blog. (Note: The forced prompt itself is kept private to avoid server overload).
https://t.co/OwBZu0iCqR
@GeminiApp Regarding the Gemini Gem 1076 error: When the context gets long, is the reference to the Gem's Knowledge working too strongly, causing it to stop reading the rest of the chat context? (Is it stuck in an infinite loop where the Knowledge referencing never finishes?)
I suspect the recent update that adds citation anchors for Knowledge in the chat is causing "context pollution." It feels like the AI is re-reading its own anchors, strengthening self-reference and triggering the endless loop.
“TANK! Extended” 100万再生♪
聴いてくれたみんな、ありがとう。
目指せ、1億と2千回再生🐤
“TANK! Extended” has passed 1 million views♪
Thank you to everyone who listened.
Next stop: 120 million views🐤
#YokoKanno#CaptainDuckling#TANK
@levelsio flagged to workspace team, there should be an appeal process for this and in general, agree with the concerned given adversarial potential of this