Unverified50% confidenceConceptExact time
Position bias means the judge model's judgment of information in the context is affected by where that information appears: content near the end is often given higher weight, while key warnings that appear early are diluted in long conversations.
1
Sources
50%
Confidence
Long-term
Relevance
10/15/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedThe drawback of the context condensation strategy is that it may lose details deemed unimportant, leading to misunderstandings or repeated mistakes in subsequent conversations.74% similarUnverifiedInput perturbation robustness is a core metric of decision models, including synonym rewrites, punctuation changes, insertion of irrelevant sentences, spelling errors, and swapping the positions of conditions; a high flip rate under perturbation renders threshold policies meaningless.74% similarUnverifiedPrecisely locating mention boundaries is more challenging than judging passage topics, as reflected in the performance gap between mention-level and passage-level metrics.73% similarUnverifiedThe causes of the lost-in-the-middle phenomenon are usually attributed to the decay effect of positional encoding, the fact that beginning and end content in training data is more often associated with answers, and the dilution of the attention softmax over long sequences.72% similarUnverifiedPlacing the question before the context causes the attention mechanism to assign higher weights to relevant tokens, allowing attention-score-based KV eviction strategies to easily identify important tokens, but in real inference scenarios the question is often given only after the context has been processed.72% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1036221API
curl https://kongchang.com/api/v1/knowledge/claims/1036221MCP
get_claim(id=1036221)