Unverified50% confidenceOpinionExact time
Long prompts triggering cross-user leakage is usually related to how the inference engine manages context length, batching strategy, and KV cache.
1
Sources
50%
Confidence
Long-term
Relevance
10/27/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedSelection leakage refers to the search loop itself being able to read the final evaluation score, causing the model to overfit to the holdout set when the number of iterations is large enough, and is closely related to the multiple comparisons problem and p-hacking.71% similarUnverifiedAn overly long system prompt occupies the context window and affects the model's understanding of user input.69% similarUnverified防御推理链窃取的策略包括限制logprobs等细粒度输出、对摘要进行随机化或差分隐私处理、部署异常查询检测机制68% similarUnverifiedAlgorithmic recommendation often leads to information cocoons and accidental exposure; even if a user does not actively follow a topic, the algorithm may push posts containing spoilers onto their timeline due to relevance.68% similarUnverifiedContext truncation is the most common source of confounding in long-input experiments: when the concatenated prompt exceeds the model's context window or the server-side max_tokens limit, the system usually silently discards content from the tail or the head.68% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1104793API
curl https://kongchang.com/api/v1/knowledge/claims/1104793MCP
get_claim(id=1104793)