Unverified50% confidenceFactExact time
The dense description method generates text descriptions as historical memory by calling a vision-language model every few frames, at the cost of being highly lossy, often glossing over the temporal transitions between actions and the trigger conditions for state switches.
1
Sources
50%
Confidence
Long-term
Relevance
10/15/2026
First Seen
Sources
Related Entities
Related Claims
Unverified大语言模型在生成查找目标文本时依赖记忆重建而非逐字符复制，在长上下文场景下会出现记忆漂移导致字符串替换静默失败，此现象被称为幻觉性精确72% similarUnverified文本生成阶段是memory-bound问题，速度受限于内存带宽；预填充阶段是compute-bound场景，可高度并行化70% similarUnverifiedOne common technical approach for large language models to implement memory is to write a summary of past conversations into the system prompt, allowing the model to read the learner's historical profile at the start of each session.69% similarUnverifiedTextual chain-of-thought alleviates the problem of models relying on language priors by decomposing visual questions into intermediate steps, but those intermediate steps require autoregressive token-by-token generation, which increases inference cost.69% similarUnverifiedCompared with RAG-based text memory, targeted replay can recover at the pixel level information that was never written into text during the description stage, at the cost of additional storage and read overhead, and the planner must accurately indicate the replay range.69% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1039506API
curl https://kongchang.com/api/v1/knowledge/claims/1039506MCP
get_claim(id=1039506)