Unverified50% confidenceFactExact time
In an attention system, a first-order phase transition means that when a control parameter such as temperature or the variance of the keys crosses a critical threshold, the system suddenly jumps from one free-energy minimum to another, and the system may exhibit hysteresis—the threshold for degrading from the robust state and the threshold for recovering are not the same.
1
Sources
50%
Confidence
Long-term
Relevance
10/14/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedEven when temperature is set to 0, different conversation histories and system prompts may cause minute shifts in the attention weights of the Transformer self-attention mechanism, thereby producing different numerical outputs.68% similarUnverifiedAfter a model version update, its attention allocation mechanism, temperature sensitivity, and instruction parsing method may change, causing existing prompt strategies to become ineffective.67% similarUnverifiedAgent behavior is non-deterministic; even using the same prompt and temperature=0 settings, the same model's outputs at different points in time may exhibit subtle differences, influenced by factors such as floating-point arithmetic precision, batching strategies, and model version updates.66% similarUnverifiedAs task chains grow longer, the capacity limitations of the context window and attention decay affect the model's ability to accurately maintain the initial goal, which is one of the root causes of the 'goal drift' phenomenon in long-horizon tasks.66% similarUnverified在混合架构中，SSM 层与注意力层交替出现时，模型需要在层之间传递和恢复状态，频繁的状态存取会带来额外开销，在长序列生成场景下会制约解码吞吐量。66% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1023891API
curl https://kongchang.com/api/v1/knowledge/claims/1023891MCP
get_claim(id=1023891)