Disagreement-Guided Steering
An inference-time intervention method that uses disagreement among multilingual retrieval results to construct a direction vector. Vector addition is then applied to correct the language bias of query representations in a shared multilingual index, without fine-tuning model weights.
Timeline (last 90 days)
The research team proposed a disagreement-guided steering method, which obtains a direction vector by comparing the representation difference between the documents retrieved by that language alone and the consensus documents of other languages, and adds it to that language's query vector at inference time.
Experimental results show that disagreement-guided steering is most effective at reducing same-language document crowding, while remaining competitive on recall.
Language mean removal performs better than disagreement-guided steering at preserving ranking quality, and there is a tradeoff between the two.
For recall-prioritizing scenarios, disagreement-guided steering is more suitable; for scenarios with high ranking-precision requirements, mean removal may be the safer choice.
The key difference between disagreement-guided steering and post-hoc debiasing methods such as language mean removal, subspace removal, and per-language rotation is that it uses the dynamic signal of disagreement among retrieval results rather than global statistics of languages.
All Facts (5)
The research team proposed a disagreement-guided steering method, which obtains a direction vector by comparing the representation difference between the documents retrieved by that language alone and the consensus documents of other languages, and adds it to that language's query vector at inference time.
50%UnverifiedExperimental results show that disagreement-guided steering is most effective at reducing same-language document crowding, while remaining competitive on recall.
50%UnverifiedLanguage mean removal performs better than disagreement-guided steering at preserving ranking quality, and there is a tradeoff between the two.
50%UnverifiedFor recall-prioritizing scenarios, disagreement-guided steering is more suitable; for scenarios with high ranking-precision requirements, mean removal may be the safer choice.
50%UnverifiedThe key difference between disagreement-guided steering and post-hoc debiasing methods such as language mean removal, subspace removal, and per-language rotation is that it uses the dynamic signal of disagreement among retrieval results rather than global statistics of languages.
50%