Unverified50% confidenceFactExact time
The repository is based on the paper 'The Pain Axis: LLMs Represent Self-Directed Harm and Act on It' and uses small open-source models (Qwen3, Llama 3.2, Phi-4-mini) for experiments.
1
Sources
50%
Confidence
Long-term
Relevance
10/14/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedThe paper's core technique is: feeding the model sentences containing 'pain' and sentences not containing 'pain', measuring the model's internal states, and then subtracting them to obtain the direction vector of 'pain' in the model's latent space.68% similarUnverifiedThe prediction scaffold proposed in the study consists of an LLM that reads the dataset to assess the intensity and breadth of nudges toward undesirable behavior, followed by a learned model that incorporates the base occurrence rate of the failure mode and the target model's prior tendencies.58% similarUnverifiedResearchers design trade-off scenarios in which models choose between achieving a higher score and avoiding being described as experiencing suffering, in order to observe the model's decision-making tendencies.57% similarUnverifiedThe LLM's symptom extraction process is based on a standard vocabulary written by clinicians, ensuring structured and predictable output.57% similarUnverified基于LLM的系统天然面临提示词注入（Prompt Injection）攻击56% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1026339API
curl https://kongchang.com/api/v1/knowledge/claims/1026339MCP
get_claim(id=1026339)