---
title: 'FiD-LightSP: Source-Pointer Re-Ranking'
url: https://www.emergentmind.com/topics/source-pointer-re-ranking-fid-light-sp
type: topic
---

# FiD-LightSP: Source-Pointer Re-Ranking

Source-Pointer Re-Ranking (FiD-Light$^{SP}$) is an extension of the FiD-Light retrieval-augmented text generation model that introduces explicit source pointer mechanisms for re-ranking retrieved passages, yielding substantial gains in combined answer generation and provenance retrieval, particularly under strict latency and efficiency constraints. The method leverages compressed passage representations and end-to-end sequence-to-sequence training to achieve state-of-the-art results on the KILT benchmark with near-zero added computational cost [2209.14290].

## 1. FiD-Light Model Architecture and Data Flow

FiD-Light is based on the Fusion-in-Decoder (FiD) encoder–decoder paradigm, which encodes each retrieved passage independently using a T5-style Transformer encoder, denoted as $T_E$. The encoder input for passage $i$ is $e_i = T_E([\text{“query:”}; q; \text{“context:”}; p_i])$, where $q$ is the query, $p_i$ the passage, and $e_i \in \mathbb{R}^{(|q|+|p|)\times D}$ with $D$ as the hidden size. In the vanilla FiD model, these per-passage encodings are fully concatenated, yielding a decoder input $o = T_D([e_1; e_2; \ldots; e_n])$ for $n$ passages, resulting in input sequences of up to $O(40 \cdot (|q|+|p|)) \approx 5,000$–$10,000$ vectors.

FiD-Light replaces this full concatenation with a lossy compression function $f_k(e_i) \in \mathbb{R}^{k \times D}$ for each passage, where $k \ll |q| + |p|$. The decoder input becomes $\hat{o} = T_D([f_k(e_1); f_k(e_2); \ldots; f_k(e_n)])$. The typical instantiation is “take the first $k$ vectors of $e_i$,” though learned pooling or attention over passage encodings is plausible.

## 2. Source Pointer Mechanism

FiD-Light$^{SP}$ (“FiD-Light with Source Pointers”) incorporates pointer tokens as re-ranking cues, extending the input format as $e_i = T_E([\text{“query:”}; q; \text{“index:”}; i; \text{“context:”}; p_i])$ to mark each passage. At training, the target output concatenates all relevant passage indices (in arbitrary order) with the answer: $\hat{t} = [\text{“index:”}; r_1; \ldots; r_R; \text{“text:”}; t]$, with $\{r_1,\ldots, r_R\} \subseteq \{1,\ldots,n\}$ the gold-relevant indices and $t$ the answer string.

At inference, the decoder freely emits index tokens, defining a pointer set $\{\hat{r}_1, \ldots, \hat{r}_m\}$. These indices serve not as explanations but as actionable signals for document re-ranking.

## 3. Passage Re-Ranking via Decoded Pointer Sets

To robustly select top provenance items, FiD-Light$^{SP}$ leverages the generated pointer set to reorder the original retrieval list $C = [p_1, p_2, ..., p_n]$. The re-ranked list is
$$
\hat{C} = [\hat{r}_1, \hat{r}_2, \ldots, \hat{r}_m, \ldots \text{the remaining passages in }C\text{ in original order}],
$$
where $\hat{r}_j$ appear in the output. Alternately, each pointer can be viewed as casting a "vote" via $s_i = \sum_{t \in T_{\text{ptr}}} \mathbb{1}[y_t = i]$, with $T_{\text{ptr}}$ the pointer token positions. Passages are ranked by $s_i$ as an alternate implementation.

Unlike previous FiD-Ex approaches, which directly generate indices, FiD-Light$^{SP}$ mitigates the problem of raw index under-generation, especially when the gold set size $R>1$ (as in multi-hop questions).

## 4. End-to-End Joint Pointer and Text Training Objective

End-to-end fine-tuning is performed on the full encoder–decoder stack with a single cross-entropy loss over the concatenated index and answer token output:
$$
\mathcal{L}(\theta) = - \sum_{t \in T_{\text{ptr}}} \log p_\theta(\text{pointer}_t \mid y_{<t}, x, P) - \sum_{t \in T_{\text{text}}} \log p_\theta(y_t \mid y_{<t}, x, P).
$$
In practice, indices are packed as ordinary tokens prior to answer tokens, using a standard sequence-to-sequence cross-entropy formulation without auxiliary weighting. The model learns to emit correct indices before switching to answer generation.

## 5. Computational Efficiency and Efficacy Trade-Offs

FiD-Light$^{SP}$ achieves significant efficiency improvements due to passage compression. With $k=64$, decoder compute for 40 passages drops by approximately $6\times$, and the total query latency is reduced by over $2\times$ (under T5-Base), even while all 40 passages are still retrieved, hence maintaining high recall. This design allows leveraging larger T5 backbones (T5-Large, T5-XL) for further quality improvements at the same or reduced latency relative to FiD-Base.

Table: Decoder Input Scale

| Model            | Passages ($n$) | Passage Compression ($k$) | Relative Decoder Cost |
|------------------|----------------|--------------------------|----------------------|
| FiD-Base         | 40             | $|q|+|p|$ (~100–250)      | 1.0                  |
| FiD-Light$^{SP}$ | 40             | 8/32/64                  | $\approx$0.17–0.66   |

*All numbers refer to relative decoder input size per query as derived from model configurations.*

## 6. Empirical Results on KILT Benchmark Tasks

FiD-Light$^{SP}$ was evaluated on seven knowledge-intensive tasks (HotpotQA, NQ, TriviaQA, FEVER, T-REx, zsRE, WOW) in the KILT benchmark. Representative combined KILT-score results (text score × R-Precision) demonstrate:

- **T5-Base, $k=64$:** NQ: 45.6 K-EM, HotpotQA: 25.6 K-EM, TriviaQA: 57.6 K-EM, FEVER: 80.6 K-AC, T-REx: 76.0 K-AC, zsRE: 81.1 K-AC
- **T5-Large, $k=32$:** NQ: 49.9, HotpotQA: 28.2, TriviaQA: 61.4, FEVER: 82.1, T-REx: 76.7, zsRE: 84.1
- **T5-XL, $k=8$:** NQ: 51.1, HotpotQA: 29.2, TriviaQA: 63.7, FEVER: 84.5, T-REx: 76.3, zsRE: 84.0

Parentheses in the original data denote gains over the prior state of the art. FiD-Light$^{SP}$ sets new state-of-the-art R-Precision for nearly all tasks, attributed specifically to its pointer-based re-ranking regime.

Ablation studies reveal that direct index decoding in FiD-Ex frequently under-selects relevant indices, especially when $R > 1$ (multi-hop cases). For instance, T5-Base HotpotQA doc R-Precision increases from 23.7% to 24.1%, TriviaQA from 32.1% to 37.6%, and FEVER from 77.1% to 78.1% when switching from FiD-Ex to FiD-Light$^{SP}$ re-ranking. Pareto-frontier analysis on latency vs. KILT-score shows FiD-Light$^{SP}$ variants strictly dominating the FiD baseline: equal or better quality at 2–3$\times$ lower inference cost or much higher effectiveness at iso-cost.

## 7. Practical and Methodological Significance

The Source-Pointer Re-Ranking mechanism transforms FiD-Light into a dual-purpose generator and re-ranker. This approach introduces virtually no additional computation—merely a minimal increase in target tokens—while delivering robust, multi-passage provenance pointing and integrating seamlessly into the compressed-encoding architecture. Its design addresses several challenges inherent to retrieval-augmented generation, such as efficient handling of long input sequences, joint answer and provenance evaluation, and robust multi-hop provenance under variable $R$. FiD-Light$^{SP}$ thus provides an effective, efficient unified solution for high-precision retrieval-augmented generation tasks [2209.14290].

Source: https://www.emergentmind.com/topics/source-pointer-re-ranking-fid-light-sp