Retrieval-augmented language models answer questions by reading documents pulled from the web. When those documents disagree, or when ten of them repeat the same unverified claim, today's systems still produce an answer. Their decision about whether to answer at all comes from the model's own confidence: how sure the model sounds, how peaked its token probabilities are. That confidence says little about the evidence behind it. A model reading a fluent fabricated article will answer with low uncertainty. A model reading five syndicated copies of one press release treats them as five confirmations.
We propose to make the abstain-or-answer decision depend on an explicit count of how much trustworthy, independent evidence supports the leading answer. Each retrieved source carries a trust record, kept as evidence counts rather than a single score, so that a source known to be bad differs from a source nobody knows anything about. Sources that copy or share an upstream origin are grouped and counted as one witness. The result is a per-query belief over candidate answers that separates three reasons to abstain: too little trusted evidence, trusted sources that contradict each other, and internal model doubt. We then set the abstention threshold with a finite-sample guarantee on error among answered queries. The two-sentence version: abstain when independent, credible witnesses are scarce or disagree, not merely when the model feels unsure; and do not let copies of one source masquerade as consensus.
Core Claim
Our claim is that the right selection score for retrieval-augmented answering is a fused evidential opinion over answer candidates. It is built from three ingredients: per-source trust opinions that keep "distrusted" apart from "unknown", fusion across provenance clusters rather than across individual documents, and the model's internal confidence. This score should beat internal-confidence selectors on the risk–coverage trade-off when misinformation is overrepresented or sources are unfamiliar. It should also beat scalar reliability voting, because scalar weights cannot tell ignorance from conflict and count duplicates as independent votes.
The last part is the piece we think is actually new. Section 2 of the literature shows the obvious combination is already partly occupied. Retrieval augmented generation with estimation of source reliability (RA-RAG) estimates scalar source reliability by cross-checking answers and aggregates per-source answers by weighted majority vote, but it always answers. Resolving Conflicting Evidence in Automated Fact-Checking shows that putting publisher credibility into the prompt beats filtering or reweighting documents, but it never abstains. Discounted Belief Fusion turns confident disagreement between modalities into uncertainty mass, but only for multimodal classification, with no source histories and no abstention guarantee. Bolting a Subjective Logic layer onto RA-RAG and thresholding its uncertainty would be an incremental extension of these. It would also likely be matched by a simple vote-margin threshold, as discussed below. The dependence-aware counting is what neither scalar voting nor standard cumulative fusion provides. Both assume each document is an independent witness, and that assumption is exactly what the "majority vote trap" documented by Ge et al. and the feedback loop described in Automated Evidence Collection for Fake News Detection exploit.
Research Program
Source opinions. Every source (a web domain, optionally refined by topic) holds a Beta reputation in the style of The Beta Reputation System, which models trust as counts of positive and negative evidence. We read this as a Subjective Logic opinion of belief, disbelief and uncertainty summing to one. Each source also has a base rate $a$ that sets the expected trust $E = b + a\,u$ for a source with no record. Prior counts come from two places. On CONFACT we use Media Bias/Fact Check ratings, converted to a modest number of pseudo-observations. Elsewhere we use RA-RAG-style cross-checking on a held-out calibration pool, which supplies observed counts. Topic-specific records address the open problem from Direction 1 that a domain can be reliable on technology and unreliable on medicine. When a topic has little data, its record shrinks toward the domain-level record.
Answer candidates. Following RA-RAG, the model generates one candidate answer per retrieved source. Candidates are grouped into equivalence classes with an entailment check. For claim verification there are exactly two classes, support and refute.
Provenance clusters. Retrieved documents are grouped before fusion. Two documents go in the same cluster if they are near-duplicates in text, share a cited upstream source or wire origin, or belong to a known common owner. This operationalizes the nondistributivity result of Trust Representation and Aggregation in a Distributed Agent System: witnesses relaying the same referrer are dependent, and aggregating them as independent manufactures reputation. Each cluster $c$ contributes evidence $e_{c,k} = \bar{t}_c\, s_{c,k}$ toward answer class $k$. Here $\bar{t}_c$ is the mean expected trust of the cluster's members, and $s_{c,k} \in [0,1]$ is the fraction of the cluster's members whose answer falls in class $k$. Adding copies therefore raises a cluster's evidence by at most one unit. Over $K$ answer classes, the fused opinion is a Dirichlet with parameters $\alpha_k = 1 + \sum_c e_{c,k}$. Its leftover ignorance is
$$u = \frac{K}{K + \sum_{c,k} e_{c,k}},$$
so $u$ shrinks only as trusted, independent clusters accumulate, regardless of how many raw documents were retrieved.
Conflict. Separately from ignorance, we compute a pairwise conflict between clusters, in the spirit of DBF. It is high when two well-trusted clusters back different classes. Following DBF, conflict moves belief mass into uncertainty, with the conflict sensitivity tuned on validation data. Tuning this sensitivity is a known weakness of DBF, and we report how sensitive our results are to it.
Selection and guarantee. The final selection score is a small logistic combination of four inputs: the fused expected probability of the top answer class, $u$, the conflict, and the model's internal confidence (the likelihood of its answer). The thresholds are set with the Selection with Guaranteed Risk procedure of Selective Classification for Deep Neural Networks, which uses a binomial bound to certify that error on answered queries stays below a target with high probability. Answer correctness is binary on all our benchmarks (exact match or verdict label), so the 0/1-loss assumption behind this procedure holds. At test time, every abstention is labeled by the term that drove it, which makes deferrals auditable.
Baselines
The naive reference point is vanilla retrieval-augmented generation over the top documents, forced to answer at full coverage. Ge et al. report it fails when misinformation is overrepresented, but give no number. We pair it with the softmax-response or token-likelihood selector of Geifman and El-Yaniv, run on identical retrieval.
The strongest internal-confidence competitor is RLSR from Aligning Language Models with Selective Prediction, which trains verbalized confidence with a ranking-based reward. On HotPotQA-Modified it reports AURC (area under the risk–coverage curve, lower is better) of 0.44, against 0.56 for correctness-only alignment and 0.51 for calibration alignment. We also include ASPIRE from Adaptation with Self-Evaluation to Improve Selective Prediction in LLMs, a cheaper learned self-evaluation that reports CoQA AUROC rising from 74.61% to 80.25%. Both are the strongest representatives of selection driven purely by internal signals. We train them with the same base model, the same training queries and an equal tuning budget. Their scores may feed the same logistic selector as additional features, so that any gain we show comes from the source signals rather than from a better final combiner.
The strongest source-aware competitor is RA-RAG. No reported result found on Lacuna for RA-RAG's NQ or TriviaQA accuracy; the file reports only that it outperformed vanilla RAG and Self-RAG under noise and cut token use by 99.1%. We turn it into a selector by thresholding its weighted-vote margin. We give it the same reliability priors, the same per-source answer candidates and the same SGR calibration. On CONFACT we add Source Background Augmentation, which writes publisher background into the prompt. Its reported gain on the ModC split is 10% absolute F1 for LLaMA-3.1 and 5% accuracy for Mistral, both relative to vanilla RAG. We pair it with an explicit instruction to answer "insufficient evidence", plus its likelihood as a selection score.
Three ablations isolate the key ingredients. The first removes provenance clustering and fuses per document. The second replaces clustering with plain near-duplicate deletion. The third collapses each source opinion to its scalar expectation, which removes the distrusted-versus-unknown distinction. Beating a baseline means a lower AURC with a bootstrap confidence interval excluding zero, and higher coverage at the target risk, with the SGR guarantee violated no more often than its nominal rate across resampled splits.
Why not just…?
Why not just threshold the model's own confidence, or train it to verbalize confidence with RLSR? Internal confidence measures how sure the model is given its context, not whether that context deserves belief. Know When to Abstain shows that logit-based selectors degrade under distribution shift and that distance-aware selectors roughly halve selective risk there. A fabricated but fluent source is a shift that no internal signal sees. Still, RLSR was evaluated on HotPotQA-Modified, where evidence availability varies, so it partly learns evidence sufficiency. If RLSR matches our selector on the syndication and cold-start splits described below, then external source modeling adds nothing beyond what alignment learns, and we will report that.
Why not just threshold RA-RAG's weighted-vote margin? This is the most serious threat to the Subjective Logic machinery. A margin threshold combined with the total vote weight already approximates "how much trusted evidence agrees." We therefore also give this baseline the total weight as a second feature. If it matches the full method on all splits, the evidential formalism is decoration and the contribution reduces to the evaluation and the clustering. The test that settles it is the cold-start split. There, the scalar baseline must assign unfamiliar domains either a default weight (naive trust) or zero (wasted evidence). Our method keeps such domains at high uncertainty and defers instead.
Why not just deduplicate retrieved documents and vote? Near-duplicate deletion catches verbatim copies. It does not catch paraphrased syndication, rewritten press releases, or coordinated sites citing one origin, which is the collusion failure noted in Direction 3. The second ablation above tests this directly. If deduplication matches clustering even on paraphrased coordinated injections, we will adopt deduplication as the cheaper component and say so.
Why not just prompt with source background and let the model abstain? Ge et al. show this prompt-based approach is the best way to resolve conflicts on CONFACT. They also show it degrades as context grows from five to ten chunks, and it offers no risk guarantee. We test whether its abstentions remain reliable at ten or more documents, compared with our pre-generation summary of the evidence, which does not lengthen the prompt.
Proposed Experiments
CONFACT. We use its ModC and HumC splits (3,180 claims with publisher ratings) as the primary real-conflict benchmark, reporting AURC, F1 at matched coverage, and coverage at 10% verdict error. We additionally hold out 30% of publishers' ratings to create genuinely unknown sources.
Synthetic noise. On Natural Questions and TriviaQA we reuse RA-RAG's Beta-prior and adversarial-injection noise protocols. We add two stress tests absent from prior work. In the syndication test, one false document is replicated across $m$ fake domains as verbatim copies and as paraphrases, for $m$ from 1 to 10. In the cold-start test, honest sources appear with no history.
Other checks. HotPotQA-Modified provides direct comparability with RLSR's reported 0.44 AURC. PolitiFact checks whether learned source opinions track oracle truthfulness ratings, measured by rank correlation.
Decisive outcomes. The method succeeds if three things hold. Its AURC advantage over RA-RAG's margin selector should grow with $m$ in the syndication test, while the per-document ablation's advantage should not. It should beat RLSR on the syndication and cold-start splits without losing on HotPotQA-Modified. And the SGR target risk should hold at its nominal confidence across 100 resampled calibration splits. If the advantage is flat in $m$, or the margin-plus-weight baseline ties everywhere, the core claim is wrong.
References
- retrieval augmented generation with estimation of source reliability
- Resolving Conflicting Evidence in Automated Fact-Checking: A Study on Retrieval-Augmented LLMs
- Multimodal Learning with Uncertainty Quantification based on Discounted Belief Fusion
- Automated Evidence Collection for Fake News Detection
- The Beta Reputation System
- Trust Representation and Aggregation in a Distributed Agent System
- Selective Classification for Deep Neural Networks
- Aligning Language Models with Selective Prediction
- Adaptation with Self-Evaluation to Improve Selective Prediction in LLMs
- Know When to Abstain: Optimal Selective Classification with Likelihood Ratios