---
title: 'FACTUM: Framework for Citation Trustworthiness'
url: https://www.emergentmind.com/topics/factum-framework-for-attesting-citation-trustworthiness-via-underlying-mechanisms
type: topic
---

# FACTUM: Framework for Citation Trustworthiness

FACTUM (Framework for Attesting Citation Trustworthiness via Underlying Mechanisms) is a methodology for mechanistic detection of citation hallucination in long-form Retrieval-Augmented Generation (RAG) systems. It quantitatively characterizes the trustworthiness of generated citations by probing and aggregating internal model signals, specifically decomposing the contributions of attention and feed-forward network (FFN) pathways, as well as their geometric alignment. FACTUM directly addresses the challenge of “citation hallucinations,” where a language model confidently cites sources that do not actually support the claims being made. By leveraging four interpretable mechanistic scores—Parametric Force, Context Alignment, Beginning-of-Sentence Attention, and Pathway Alignment—FACTUM offers scale-robust, quantifiable, and explainable guidance for both citation attribution and quality control in RAG pipelines [2601.05866].

## 1. Formal Problem Definition: Citation Hallucination in RAG

FACTUM is designed for RAG systems in which a user query $x$ and a set of retrieved documents $D = \{d_1, ..., d_m\}$ yield a generated long-form response $R$ containing inline citations of the form “[Source:i]”, where $i$ indexes into $D$. Each citation-index token $t_i$ in the output is associated with a binary ground truth label $y_i \in \{0,1\}$:

- $y_i=1$ if document $d_{t_i}$ supports the directly-preceding factual claim.
- $y_i=0$ otherwise (i.e., hallucinated or mis-attributed citation).

The detection task is to estimate $P(y_i=1)$ at each citation token, using the available inputs and internal model states. The goal is accurate, token-level identification of correct versus hallucinated citations, in contrast to black-box or post-hoc truthfulness scoring [2601.05866].

## 2. Mechanistic Scores: Definitions and Intuitions

FACTUM decomposes the Transformer's residual updates at each generation step $i$ and each layer $l$ into multi-head attention (MHA) and FFN pathway components. Four mechanistic scores are defined:

**a) Parametric Force Score (PFS):**

Measures the magnitude of the FFN update vector, quantifying how strongly parametric (model-memory) information is introduced.

- Definition: $PFS^{(l)}(t_i) = \| v_{ffn,i}^{(l)} \|_2$  
Where $v_{ffn,i}^{(l)} = x_{post-ffn,i}^{(l)} - x_{pre-ffn,i}^{(l)}$.

High PFS suggests model output is being “pushed” toward stored (parametric) facts.

**b) Context Alignment Score (CAS):**

Assesses how well the generated citation token semantically aligns to retrieved evidence, via attention pathways.

- Definition: $CAS^{(l,h)}(t_i) = \cos(c_i^{(l,h)}, h_{t_i}^{(L)})$  
Where $c_i^{(l,h)} = \sum_{t_j \in T_C} A_{i,j}^{(l,h)} \cdot h_{t_j}^{(L)}$, and $T_C$ are context tokens.

High CAS indicates strong coupling between the citation and supporting document context.

**c) Beginning-of-Sentence Attention Score (BAS):**

Quantifies attention mass allocated to the sentence-initial token (information synthesis “sink”).

- Definition: $BAS^{(l,h)}(t_i) = A_{i, 1}^{(l,h)}$

High BAS is associated with greater internal synthesis prior to citation generation.

**d) Pathway Alignment Score (PAS):**

Computes the cosine similarity between MHA and FFN update vectors.

- Definition: $PAS^{(l)}(t_i) = \cos(v_{attn,i}^{(l)}, v_{ffn,i}^{(l)})$

PAS $\approx$ +1: cooperative pathway usage, PAS $\approx$ 0: orthogonal, PAS $\approx$ -1: antagonistic pathway updates.

## 3. Practical Pipeline for Score Measurement

Extracting mechanistic scores in FACTUM entails instrumenting the Transformer at inference time:

- Register hooks on all MHA and FFN sublayers; capture attention weights ($A_{i,\cdot}^{(l,h)}$) and pre/post-FFN hidden states at the citation-index token position.
- For each response, identify citation tokens $t_i$ and compute PFS, CAS, BAS, and PAS for every layer $l$ and attention head $h$ as applicable.
- Aggregate scores through component-pruning (ranking heads/layers by correlation with ground-truth), head-aggregation (mean/std over heads), and layer-aggregation (summary statistics: mean, slope, frequency bandpower).

A reference pseudocode implementing these steps is provided in [2601.05866] and is reproduced here for clarity:

```python
for each generated response R:
    identify all citation-index tokens t_i
    for each t_i:
        for each layer l:
            for each head h:
                A = attention_weights[l][h][i]
                BAS[l,h] = A[0]
                context_mask = is_context_token
                c = sum_j (A[j] * h_final[j]) over j ∈ T_C
                CAS[l,h] = cosine(c, h_final[i])
            x_input = hidden_state_before_layer[l][i]
            v_attn = hidden_state_after_attn[l][i] - x_input
            x_pre_ffn = hidden_state_after_attn[l][i]
            x_post_ffn = hidden_state_after_ffn[l][i]
            v_ffn = x_post_ffn - x_pre_ffn
            PFS[l] = norm2(v_ffn)
            PAS[l] = cosine(v_attn, v_ffn)
```

Feature vectors composed from these mechanistic readouts form the basis for downstream classification (e.g., via Logistic Regression, Explainable Boosting Machines, LightGBM).

## 4. Model Scale Dependence and Mechanistic Signature Evolution

Citation trustworthiness signals in FACTUM are not static; their structure changes with model scale:

- For Llama-3.2-3B, correct citations exhibit statistically significant increases in all four FACTUM scores (CAS, BAS, PFS, PAS), supporting a “full-agreement” signature.
- For Llama-3.1-8B, optimal performance is obtained by pruning to the top 25% most predictive heads/layers; in this regime, correct citations show strong PFS and BAS but either variable or decreased PAS, implying more specialized/orthogonal information transfer between MHA and FFN.
- CAS for larger models is also less reliable: signal directionality reverses or weakens for correct versus hallucinated citations.

These observations confirm that citation attribution is an emergent, scale-sensitive phenomenon. Regular re-calibration of the component selection and aggregation strategy is advised after model updates or fine-tuning [2601.05866].

## 5. Key Experimental Findings and Quantitative Performance

Tasked with token-level classification of citation indices for TREC NeuCLIR 2024 (15-document long-context RAG), FACTUM demonstrates robust and statistically significant improvements relative to prior baselines (ReDeEP ECS+PKS, uncertainty-based scores):

| Model            | Baseline           | AUC         | FACTUM (AUC)   | Relative Improvement |
|------------------|--------------------|-------------|---------------|---------------------|
| Llama-3.2-3B     | ECS+PKS (ReDeEP)   | ∼ 0.61      | 0.715–0.737   | +37.5%              |
| Llama-3.2-3B     | CAS+PFS            | ∼ 0.67      | 0.715–0.737   | +17–21%             |
| Llama-3.1-8B     | ECS+PKS            | 0.54–0.61   | 0.69–0.74     | +13–28%             |
| Llama-3.1-8B     | CAS+PFS            | ∼ 0.68      | 0.69–0.74     | +1.5–7%             |

All improvements hold under FDR-adjusted significance tests ($p<0.05$). Precision, Recall, and F1 metrics also increase, confirming balanced reduction in both hallucinated and missed citations [2601.05866].

## 6. Integration and Best Practices for RAG Systems

FACTUM provides a unified workflow for scalable, model-internal citation auditing:

- Integrate via forward hooks in production RAG systems to compute PFS, CAS, BAS, PAS at generation.
- Use distilled features in a lightweight detector (Logistic Regression/EBM).
- Flag potentially hallucinated citations ($P(y_i=1)<\tau$) for downstream “guardrail” interventions: warning users, triggering re-retrieval, or blocking low-trust attributions.
- For small models ($\sim$3B), encourage holistic pathway engagement. For large models ($\sim$8B), use component pruning and focus on orthogonality/force rather than just pathway agreement.
- Re-validate and re-select components post-fine-tuning, as attention/activation structure may shift.

These steps ensure robust, explainable detection of trust breakdowns in citation, addressing nuanced failure modes beyond mere “parametric over-reliance” and supporting systematic verifiability in production LLM-driven knowledge workflows.

## 7. Broader Implications for Trustworthy and Attested Citation

FACTUM fundamentally reframes citation hallucination as a complex, evolving dynamic across multiple neural mechanisms, rather than a single-pathway or scalar-confidence issue. By utilizing mechanistic signals—parametric force, context attention, attentional synthesis, and pathway alignment—Citation Trustworthiness becomes directly auditable and model-interpretable.

In the broader context of the FACTUM research program, the approach complements generative frameworks (such as RAEL and Intralign [2504.14856]) and fact-verification-centric pipelines (e.g., VeriFact-CoT [2509.05741]) by providing orthogonal, actionable evidence tied to the model's computational substrate. This mechanistic transparency is essential for deploying RAG systems in regulated/critical domains, enabling not just trust, but *attestable* provenance and error detection in AI-assisted citation generation.

Source: https://www.emergentmind.com/topics/factum-framework-for-attesting-citation-trustworthiness-via-underlying-mechanisms