REVIEW 16 cited by
Improving Efficient Neural Ranking Models with Cross-Architecture Knowledge Distillation
Not yet reviewed by Pith; the record is open.
This paper has not been read by Pith yet. Machine review is queued; the pith claim, tier, and objections will appear here once it completes.
SPECIMEN: schema-true, not a live event
T0 review · schema-true
One-sentence machine reading of the paper's core claim.
pith:XXXXXXXX · record.json · timestamp
Improving Efficient Neural Ranking Models with Cross-Architecture Knowledge Distillation
read the original abstract
Retrieval and ranking models are the backbone of many applications such as web search, open domain QA, or text-based recommender systems. The latency of neural ranking models at query time is largely dependent on the architecture and deliberate choices by their designers to trade-off effectiveness for higher efficiency. This focus on low query latency of a rising number of efficient ranking architectures make them feasible for production deployment. In machine learning an increasingly common approach to close the effectiveness gap of more efficient models is to apply knowledge distillation from a large teacher model to a smaller student model. We find that different ranking architectures tend to produce output scores in different magnitudes. Based on this finding, we propose a cross-architecture training procedure with a margin focused loss (Margin-MSE), that adapts knowledge distillation to the varying score output distributions of different BERT and non-BERT passage ranking architectures. We apply the teachable information as additional fine-grained labels to existing training triples of the MSMARCO-Passage collection. We evaluate our procedure of distilling knowledge from state-of-the-art concatenated BERT models to four different efficient architectures (TK, ColBERT, PreTT, and a BERT CLS dot product model). We show that across our evaluated architectures our Margin-MSE knowledge distillation significantly improves re-ranking effectiveness without compromising their efficiency. Additionally, we show our general distillation method to improve nearest neighbor based index retrieval with the BERT dot product model, offering competitive results with specialized and much more costly training methods. To benefit the community, we publish the teacher-score training files in a ready-to-use package.
Forward citations
Cited by 16 Pith papers
-
Prism-Reranker: Beyond Relevance Scoring -- Jointly Producing Contributions and Evidence for Agentic Retrieval
Prism-Reranker models output relevance, contribution statements, and evidence passages to support agentic retrieval beyond scalar scoring.
-
A Unified Model and Document Representation for On-Device Retrieval-Augmented Generation
A single model unifies retrieval and context compression for on-device RAG via shared representations, matching traditional RAG performance at 1/10 context size with no extra storage.
-
Relevance-Based Embeddings: Lightweight Candidate Retrieval via Heavy-Ranker Calls
Relevance vectors to support items/queries, mapped by light networks, uniformly approximate any continuous ranker and improve lightweight candidate retrieval over dual encoders.
-
Scaling Dense Retrieval with LLM-Annotated Training Data: Structured Mining and Progressive Curriculum for E-Commerce Sponsored Search
A pipeline using multi-system retrieval disagreement, LLM-graded annotation at 89.1% human agreement, and staged curriculum on 240M+ examples trains a two-tower model that improves NDCG@10 by 5.1% and online ad metric...
-
Understanding Wacky Weights: A Dissection of SPLADE's Learned Term Importance
SPLADE models produce wacky expansion terms whose prevalence rises with larger vocabularies and falls with stricter sparsity; these terms primarily aid in-domain retrieval rather than out-of-domain generalization.
-
LEAF: Knowledge Distillation of Text Embedding Models with Teacher-Aligned Representations
LEAF distills teacher-aligned student embedding models that achieve new SOTA results on BEIR and MTEB for their size class while requiring only modest data and compute.
-
Towards a Relevance Posterior in Neural Information Access
Fusing a cached query-independent document-quality prior with BM25 or re-rankers improves TREC DL nDCG, with largest gains for RankZephyr, under a prior–likelihood view of retrieval.
-
Rationale-Guided Knowledge Distillation for Cross-Lingual Stance Detection
Distilling LLM-generated reasoning rationales into mBERT via dual-path contrastive distillation improves cross-lingual stance detection by 1–3% accuracy on three benchmarks.
-
RAMP: Robust Ad Recommendation Under Limited Personalized-Feature Availability via Masking and Alignment Pathways
RAMP splits ad prediction into personalized and non-personalized towers, then uses a distillation-style loss to transfer knowledge so non-personalized predictions stay accurate when user features are absent.
-
Score-Only Distillation for Compact Dense Retrieval
Score-only distillation with a row-centered all-pairs PairMSE objective lets 0.6B bi-encoders recover up to 50% of the base-to-teacher retrieval gap under matched protocols.
-
Generalistic or Specific Embeddings, Which is Better? An Empirical Study on Search for Clinical Coding in Non-English Languages
Fine-tuning a Spanish biomedical encoder on Gemini-generated synthetic data for multiple languages yields a bi-encoder that matches or exceeds BioBERT-ST on clinical code retrieval metrics, with further gains from cro...
-
Beyond Hard Negatives: The Importance of Score Distribution in Knowledge Distillation for Dense Retrieval
Stratified sampling preserving teacher score distribution outperforms hard-negative mining as a robust baseline for knowledge distillation in dense retrieval.
-
jina-embeddings-v5-text: Task-Targeted Embedding Distillation
A distillation-plus-task-contrastive training regimen yields compact embedding models that match or exceed state-of-the-art performance for their size while supporting 32k-token contexts and quantization.
-
The Role of Vocabularies in Learning Sparse Representations for Ranking
Larger 100K vocabularies in SPLADE models, especially those initialized with ESPLADE pretraining, improve retrieval effectiveness after pruning compared to 32K baselines while keeping similar efficiency.
-
Search for Coverage: Learning Coverage-Aware Retrieval with Augmented Sub-Question Answerability
CoveR improves nugget coverage by 10% over dense baselines in long-form RAG via coverage-aware contrastive training on LLM-generated sub-question signals without losing relevance performance.
-
Unified Supervision for Walmart's Sponsored Search Retrieval via Joint Semantic Relevance and Behavioral Engagement Modeling
A hybrid supervision method for bi-encoder retrievers combines graded relevance from teacher models, production retrieval priors, and selective engagement to improve relevance and NDCG over Walmart's current sponsored...
discussion (0)
Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.