Home / Curated / Algorithms / TART (Task-Aware Retrieval with Instructions)

TART (Task-Aware Retrieval with Instructions)

Algorithms AIMech:TARTTaskAwareRetrievalWithInstructions DRAFT

Aliases: Task-Aware ReTriever, Task-Aware Retrieval with Instructions, TART Multi-Task Instructed Retriever

A multi-task instruction-following information retrieval methodology that conditions document relevance scoring on both the user query q and an explicit natural-language task instruction t (specifying intent, domain, and retrieval unit), training either a dual-encoder s(t, q, d) = E([t; q])^T E(d) (TART-dual) or an encoder-only cross-encoder s(t, q, d) = FFN(E([t; q; d])) (TART-full) across diverse instruction-annotated tasks using random negatives, cross-encoder-denoised hard negatives d^HD, and cross-corpus instruction-unfollowing negatives d^UF.

Identity Notes & Concept Scope Introduced by Asai, Schick, Lewis, Chen, Izacard, Riedel, Hajishirzi, and Yih (Findings of ACL 2023, pp. 3994-4022, arXiv:2211.09260). Distinguished from standard zero-shot dense retrievers (such as Contriever or GTR) that condition only on the query q, from UPR (Sachan et al., 2022) which uses a fixed question-generation prompt to compute log p(q | d) rather than adapting relevance to arbitrary task instructions t, and from Promptagator (Dai et al., 2023) which trains separate task-specific retrievers on LLM-generated synthetic queries.

Claims & Grounded Evidence

3 Claims

Mechanism #tart-instruction-conditioned-scoring-and-unfollowing-negatives

Given a query q, document d, and natural-language instruction t encoding retrieval intent, domain, and unit, TART-dual scores relevance via inner product s(t, q, d) = E([t; q])^T E(d) trained with in-batch softmax cross-entropy loss, while TART-full scores top-K candidates via s(t, q, d) = FFN(E([t; q; d])) trained with binary cross-entropy over positive documents d^+, denoised in-domain hard negatives d^HD (cross-encoder score < 0.1), and instruction-unfollowing negatives d^UF mined from other tasks' corpora that match q topically but violate t.

supports Source: Task-aware Retrieval with Instructions · Locator: Section 4.1-4.2 (Negative documents selection, Figure 3) & Section 5.1-5.2 (Equations 1-2), pp. 3-6
supports Source: Task-aware Retrieval with Instructions · Locator: filter_denoised_hard_negatives, tart_dual_inbatch_loss, tart_full_bce_loss & main

Scope #tart-zero-shot-and-cross-task-pooled-retrieval

Trained on 37 instruction-annotated BERRI datasets, TART-full (initialized from T0-3B or FLAN-T5-XL encoders, 1.5B parameters) achieves 44.8 average nDCG@10 on BEIR and 75.7 Success@5 on LOTTE-pooled without task-specific parameter updates or synthetic target-corpus data generation, and reaches 59.1 average nDCG@10 on X^2-Retrieval (3.7M pooled cross-domain documents), outperforming Contriever+CE by +8.2 points.

supports Source: Task-aware Retrieval with Instructions · Locator: Section 6.2 (X^2-Retrieval), Section 7.1-7.2 (Tables 3-4), & Section 8.4 (Figure 8), pp. 7-11
supports Source: Task-aware Retrieval with Instructions · Locator: main (X^2-Retrieval instruction routing check)

Limitation #tart-dual-encoder-capacity-and-misleading-instruction-limitation

The 110M-parameter BERT-base bi-encoder variant (TART-dual) exhibits degraded zero-shot nDCG@10 compared to non-instructed Contriever-MS MARCO on several BEIR tasks (such as Touche-2020 and Climate-FEVER, 37.4 vs. 38.0 average) due to limited model capacity and lack of token-level query-document cross-attention, and providing inaccurate or mismatched domain/unit instructions at test time degrades TART-full accuracy.

supports Source: Task-aware Retrieval with Instructions · Locator: Section 7.1 (Table 3), pp. 8-9; Section 9, p. 11; Appendix D.2 (Table 16), pp. 21, 25

⚠️ Curated Limitations & Negative Results

Compact bi-encoder architectures (e.g., 110M-parameter TART-dual) have limited capacity to follow complex unseen instructions without cross-attention, underperforming TART-full on zero-shot transfer.
Sensitive to instruction specification: inaccurate or domain-mismatched instructions at inference time degrade retrieval accuracy.

Artificial Intelligence Ontology (AIO) Grounding

Mapped AIO Classes

AIO Target Class	Mapping Relation	Mapping Rationale
`AIO:ArtificialNeuralNetwork` Artificial Neural Network	related_match	Trains neural transformer dual-encoder and cross-encoder architectures conditioned on natural-language retrieval instructions.

Evidence Sources & Primary Literature

3 Sources

Source ID	Publication / Document	Type / Access	Reference & Checksum
`r746-primary-pdf`	Task-aware Retrieval with Instructions Akari Asai, Timo Schick, Patrick Lewis, Xilun Chen, Gautier Izacard, Sebastian Riedel, Hannaneh Hajishirzi, Wen-tau Yih	primary paper full_text	ID: https://dblp.org/search/publ/api?q=Task-aware+Retrieval+with+Instructions+Asai&h=5&format=xml sha256:e3780dcbd648…
`r746-arxiv-abs`	Task-aware Retrieval with Instructions (arXiv abstract) Akari Asai, Timo Schick, Patrick Lewis, Xilun Chen, Gautier Izacard, Sebastian Riedel, Hannaneh Hajishirzi, Wen-tau Yih	primary paper abstract_only	ID: https://arxiv.org/abs/2211.09260 sha256:3170436c052e…
`r746-metadata-cr`	Task-aware Retrieval with Instructions (Crossref metadata) Akari Asai, Timo Schick, Patrick Lewis, Xilun Chen, Gautier Izacard, Sebastian Riedel, Hannaneh Hajishirzi, Wen-tau Yih	primary paper abstract_only	ID: https://dblp.org/search/publ/api?q=Task-aware+Retrieval+with+Instructions+Asai&h=5&format=xml sha256:7fdf64373980…

Ontology & Grounding Topology

Open in interactive network graph →

Curation Provenance

Curator: Antigravity

Method: ai_assisted

Model: Gemini

Created: 2026-10-19T13:49:00Z

Updated: 2026-10-19T13:49:00Z

Source File: data/curated/TARTTaskAwareRetrievalWithInstructions.yaml · Schema: curated_concept.yaml

Explore across X-Mech

Related computational & biological Mechs: Embedding Map Network Graph ProteinTraitsMech (ESM-2 pLM) DUFMech (Domain pLM) TraitMech CultureMech All 14 X-Mech Repositories