Verified65% confidenceFactExact time
Anthropic's 2023 monosemanticity research explored how to separate aliased features using sparse autoencoders.
3
Sources
65%
Confidence
Long-term
Relevance
7/12/2026
First Seen
Sources
Jacobian-Lens in Practice: Manually Reshaping LLM Behavior with J-Space
redditr/LocalLLaMA7/11/2026
Related Claims
UnverifiedAnthropic's experiments on three corpus types—codebases, novels and papers—showed that embedding combined with BM25 outperforms embedding alone.66% similarUnverifiedAbstract concepts encoded inside the model can bridge distributional gaps better than surface features.65% similarUnverifiedThe study uses tree-structured arithmetic expressions as controlled probes to investigate the information loss when natural language conveys compositional structural information.65% similarUnverifiedAnthropic proposed the Contextual Retrieval approach in 2024, which uses a large language model to generate a contextual summary for each chunk before concatenating and encoding it.64% similarUnverifiedMultiple academic papers published in 2024 demonstrated that large language models can autonomously write exploit code when given CVE descriptions.63% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/493356API
curl https://kongchang.com/api/v1/knowledge/claims/493356MCP
get_claim(id=493356)