Unverified50% confidenceFactExact time
Injecting vectors into the residual stream is a commonly used causal intervention technique in mechanistic interpretability, acting directly on the model's intermediate representations rather than at the vocabulary level.
1
Sources
50%
Confidence
Long-term
Relevance
9/16/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedModel steering influences model output behavior during inference by directly modifying the activation vectors of intermediate layers, without modifying weights.76% similarUnverified与传统可解释性方法（如特征重要性、显著图）不同，机制可解释性追求对模型内部计算的完整逆向75% similarUnverified"Correlation does not equal causation" is the fundamental limitation of traditional interpretability methods that infer model behavior by observing statistical associations between activation values and outputs.73% similarUnverifiedThe study found that even when prompts explicitly instruct the model to answer from a prospective perspective, the model is still influenced by outcome information already present in the context, and physical truncation is the effective means of blocking hindsight bias.71% similarUnverifiedBy keeping the scene constant and only replacing the key object in controlled comparative experiments, the model's physical reasoning capabilities can be observed under controlled conditions.71% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/924759API
curl https://kongchang.com/api/v1/knowledge/claims/924759MCP
get_claim(id=924759)