A retriever can find passages that resemble a question yet rank the passage containing the answer below them. End-to-end learning addresses this mismatch by training retrieval and answer production together. The goal is not to change nearest-neighbor search itself: training changes the text encoder so passages useful to the reader move closer to the queries that need them. section 3.0 section 6.8
In a separately assembled pipeline, a fixed retriever selects documents using similarity, whether or not they help the reader answer. Joint training closes that loop: the reader’s performance supplies a learning signal for the retriever. This matters when a passage repeats the question’s terms but another passage supplies the explanation the answer requires. section 3.0 section 6.8
Learning from a missing word
REALM connects a neural retriever to a language-model reader during pre-training. Given “The monarch butterfly migrates to [MASK] for the winter,” the retriever can fetch evidence about migration to Mexico; predicting “Mexico” then rewards useful retrieval. The text supplies its own answer key, without human relevance labels. Masking a name or date makes retrieval valuable, whereas masking a function word may let the reader succeed from grammar alone and give the retriever little signal. section 3.2
Joint learning can use different objectives and incur different costs:
- REALM jointly pre-trains retriever and reader using masked-token prediction as the signal for retrieval usefulness. section 3.2
- Atlas uses Perplexity Distillation (PDist): it measures how likely the reader finds the target answer with each retrieved document, then trains retriever scores toward that distribution. Other objectives use decoder attention, marginalize over documents, or measure the damage from removing one; removal is more faithful to multi-document inference but costs an extra forward pass per document. section 6.2.2
- Canonical RAG jointly fine-tunes retriever and generator and weighs contributions across retrieved documents. Naive RAG instead puts retrieved chunks into an off-the-shelf model’s prompt without joint training, trading potential task-specific gains for lower training and deployment complexity. section 4.7
Table 6.3 Training objectives
|
Training Objective
|
Core Mechanism
|
Computational Cost
|
Key Insight
|
|
Attention Distillation (ADist)
|
Distills decoder cross-attention scores into the
retriever.
|
Medium
|
Uses the reader's internal attention as a proxy for
document relevance.
|
|
EMDR²
|
Marginalizes over documents to maximize the target
output's likelihood.
|
Medium
|
Treats documents as latent variables in an end-to-end
probabilistic model.
|
|
Perplexity Distillation (PDist)
|
Trains a retriever to predict the LM's perplexity for each
document.
|
Low
|
A simple, direct, and efficient way to align retriever
scores with LM utility.
|
|
LOOP
|
Measures importance by the performance drop when a
document is removed.
|
High
|
More closely aligns with multi-document inference but is
expensive.
|
Compare four ways to turn reader behavior into retriever supervision, including their relative computational costs. section 6.2.2
Finding the explanation, not just the figure
For “Why did operating margins decline?”, a similarity-based retriever may rank a passage listing margin figures above the CFO’s causal explanation. In an Atlas-style workflow, a small set of question–answer examples lets PDist favor the explanatory passage because it makes the target answer easier for the reader to produce. The retriever learns to value evidence such as “driven by” rather than merely matching the question’s numbers. section 6.7
Joint training is a choice to justify, not a default requirement. In an Atlas few-shot ablation, a fixed retriever scored 42.2, PDist scored 45.7, and omitting joint pre-training scored 30.0; pre-training mattered more than fine distinctions among retriever losses. Start with a simpler pipeline when it meets the task, and evaluate joint training when answer-bearing passages repeatedly lose to lookalikes. Jointly optimizing an entire retriever–reranker–generator pipeline remains computationally expensive and an open direction. section 6.2.2 section 4.7 section 15.8.5



