Unverified50% confidenceFactExact time
In thehumanbox, action selection prioritizes survival needs and then executes an ε-greedy policy based on the learned values.
1
Sources
50%
Confidence
Long-term
Relevance
10/23/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedIn the ε-greedy policy, the agent chooses an action at random with probability ε and chooses the action with the highest current estimated Q-value with probability 1-ε.67% similarUnverifiedProcess-obsessed people regard results as a byproduct of the process rather than the sole driving force, and this prioritization makes them more resilient when they encounter setbacks.66% similarVerifiedChoosing a model cannot rely solely on benchmark scores; the model's willingness to execute real-world tasks is the key practical value.65% similarUnverifiedReasoning hijacking refers to an attacker tampering with intermediate goals or premises in the agent's reasoning process so that the model 'believes' it is executing a reasonable plan; the final action itself may be fully compliant, with the anomaly appearing only early in the reasoning trace.62% similarUnverifiedIf the reward model is misaligned with the true experimental objective, the harder the search works, the further it deviates; when using reward-guided search, one should first verify the degree of alignment between the reward model and the true objective.62% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1088914API
curl https://kongchang.com/api/v1/knowledge/claims/1088914MCP
get_claim(id=1088914)