Research map / Large language models / Retrieval and knowledge
ShiftRAG: Bypassing the Textual Bottleneck via Decoupled Learning and Continuous Soft Tokens
Youmin Ko, Jihong Jeong, Hyunjoon Kim
Read the paperCheck your idea against 137,217 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes ShiftRAG, a retrieval-augmented generation approach that passes continuous soft tokens from the retriever to the generator instead of only raw text, with decoupled learning of retrieval and generation.
Problem
Retrieval-augmented generation still hallucinates during complex reasoning; existing pipelines pass only raw text to the generator, discarding discriminative retriever signals, and unified retrieval-generation models suffer objective conflicts.
Approach
Bypasses the textual bottleneck by feeding continuous soft tokens derived from retrieval into the generator, while decoupling the learning objectives of retrieval and generation to avoid objective interference.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add layer-wise gap-aware contrastive loss instead of InfoNCE (ablation) | ShiftRAG encoder · MuSiQue retrieval | Recall@10: 0.6140 -> 0.6516 | Improved |
| add soft orthogonal alignment loss (ablation) | ShiftRAG encoder · MuSiQue retrieval | Recall@10: 0.6493 -> 0.6516 | Improved |
| add semantic signal translator soft knowledge tokens (ablation) | ShiftRAG decoder · MuSiQue QA | EM: 0.2450 -> 0.3050 | Improved |
| add decoder-mode LoRA tuning (ablation) | ShiftRAG decoder · MuSiQue QA | EM: 0.2650 -> 0.3050 | Improved |
| use soft tokens instead of verbalized relevance scores (ablation) | ShiftRAG · MuSiQue QA | EM: 0.2650 -> 0.3050 | Improved |
| verbalize relevance scores as text versus nothing (ablation) | ShiftRAG · MuSiQue QA | EM: 0.2450 -> 0.2650 | Improved |
| separate learning of encoder and decoder instead of joint learning with shared parameters (ablation) | ShiftRAG · HotpotQA QA | EM: 0.1300 -> 0.5610 | Improved |
| separate learning instead of joint learning with separate parameters (ablation) | ShiftRAG · HotpotQA QA | EM: 0.4730 -> 0.5610 | Improved |
| vary number of soft knowledge tokens from 3 to 7 (ablation) | ShiftRAG · QA benchmarks | QA performance: max variation 1.43% | No clear effect |
| unified LLM backbone sharing (GritLM, OneGen) versus decoupled | unified baselines · HotpotQA retrieval | Recall@2: 0.5890 -> 0.6598 (ShiftRAG) | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- NV-Embed-v2 — NV-Embed: Improved Techniques for Training LLMs as Generalist Embedding Models (ICLR 2025) · better on 9 of 9 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Retrieval is Not Enough: Enhancing RAG through Test-Time Critique and Optimization (NeurIPS 2025 · similarity 0.89)
- RankRAG: Unifying Context Ranking with Retrieval-Augmented Generation in LLMs (NeurIPS 2024 · similarity 0.88)
- Cooperative Retrieval-Augmented Generation for Question Answering: Mutual Information Exchange and Ranking by Contrasting Layers (NeurIPS 2025 · similarity 0.88)
- Retrieval from Within: An Intrinsic Capability of Attention-Based Models (NeurIPS 2026 · similarity 0.88)
- Accelerating Inference of Retrieval-Augmented Generation via Sparse Context Selection (ICLR 2025 · similarity 0.87)
- RAPID: Long-Context Inference with Retrieval-Augmented Speculative Decoding (ICML 2025 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
