analiza aiqbench
Exemplar Partitioning offers an efficient, unsupervised approach to decoding LLM activations by clustering them into interpretable regions linked directly to observed inputs. This method enables precise intervention on model behaviors, such as neutralizing safety refusals or detecting hidden concepts in Taboo models, while requiring significantly fewer tokens than traditional sparse autoencoders. Its ability to track activation reorganization during instruction tuning provides new insights into how models align with human directives without sacrificing interpretability.
Exemplar Partitioning for Mechanistic Interpretability
arXiv:2605.14347v3 Announce Type: replace Abstract: We introduce Exemplar Partitioning (EP), an unsupervised method for building interpretable feature dictionaries from large language model activations. An EP dictionary is a Voronoi partition of activation space, built by leader-clustering streamed activations within a distance threshold. Each region is defined by an observed exemplar and an average of its member activations, which define region membership and provide directions for intervention. Dictionary size is determined by the activation stream at the chosen threshold rather than pre-selected. Exemplars link regions to observed inputs, allowing dictionaries built from the same input stream to be compared across layers, training checkpoints, and architectures. We demonstrate how EP can be used to interpret and intervene on model behaviour, track changes in activation space through training, and detect hidden concepts. Comparing EP dictionaries on base and instruction-tuned Gemma-2-2B and Llama-3.1-8B reveals that instruction tuning reorganises harmful prompt activations similarly across the two models, but at different granularities. Interventions on these regions make both models answer harmful requests they previously refused. In 19 of 21 Taboo models trained to hide a secret word, EP finds new regions that do not exist in the base model, whose decoded tokens relate to the known secret on inspection. Although EP assigns each token to a single region, linear probes built from EP regions achieve up to $90.5\%$ of full-activation probe accuracy. On AxBench concept detection at Gemma-2-2B-it layer 20, EP achieves the highest mean AUROC of all unsupervised methods ($0.937$), outperforming SAE-A ($0.911$) and approaching supervised probes ($0.946$). Building EP dictionaries is fast and cheap: EP uses $\sim 10^3$ times fewer construction tokens than comparable SAEs.
Abstract:We introduce Exemplar Partitioning (EP), an unsupervised method for building interpretable feature dictionaries from large language model activations. An EP dictionary is a Voronoi partition of activation space, built by leader-clustering streamed activations within a distance threshold. Each region is defined by an observed exemplar and an average of its member activations, which define region membership and provide directions for intervention. Dictionary size is determined by the activation stream at the chosen threshold rather than pre-selected. Exemplars link regions to observed inputs, allowing dictionaries built from the same input stream to be compared across layers, training checkpoints, and architectures. We demonstrate how EP can be used to interpret and intervene on model behaviour, track changes in activation space through training, and detect hidden concepts. Comparing EP dictionaries on base and instruction-tuned Gemma-2-2B and Llama-3.1-8B reveals that instruction tuning reorganises harmful prompt activations similarly across the two models, but at different granularities. Interventions on these regions make both models answer harmful requests they previously refused. In 19 of 21 Taboo models trained to hide a secret word, EP finds new regions that do not exist in the base model, whose decoded tokens relate to the known secret on inspection. Although EP assigns each token to a single region, linear probes built from EP regions achieve up to $90.5\%$ of full-activation probe accuracy. On AxBench concept detection at Gemma-2-2B-it layer 20, EP achieves the highest mean AUROC of all unsupervised methods ($0.937$), outperforming SAE-A ($0.911$) and approaching supervised probes ($0.946$). Building EP dictionaries is fast and cheap: EP uses $\sim 10^3$ times fewer construction tokens than comparable SAEs.
Code: this https URL. Pretrained dictionaries: this https URL
Machine Learning (cs.LG)
arXiv:2605.14347 [cs.LG]
(or arXiv:2605.14347v3 [cs.LG] for this version)
https://doi.org/10.48550/arXiv.2605.14347 Focus to learn more arXiv-issued DOI via DataCite
arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.
Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.
Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
Text integral sub licența deschisă a publicației · 367 cuvinte
Subiecte conexe
J++ Lens: Jacobian Filtering Enables More Faithful Workspace Lenses
LessWrong · 8 oct. 2026
Data-driven Circuit Discovery for Interpretability of Language Models
arXiv · 15 oct. 2026
The Geometry of Logic: Stratification Induces Semantic Structure and Robust Reasoning
arXiv · 30 sept. 2026
Similarity-Distance-Magnitude Activations
arXiv · 9 iun. 2026