A deployed language model assistant cannot afford to fact-check every sentence it writes. Checking every claim against retrieved evidence roughly multiplies the cost of each answer, a cost that Chain-of-Verification and EVER both name as their main limitation. The natural fix is to check only the claims the system thinks are riskiest. This proposal argues that this fix repeats a known failure from security and public-health surveillance. If you only inspect where you already expect trouble, your records will show trouble only there. The error detector then learns from this lopsided record and becomes permanently blind to confident, fluent fabrications that nobody ever checks.
We propose treating claim verification in a deployed assistant as an ongoing audit program run under a fixed budget. The system splits each answer into atomic claims (single checkable assertions). It groups those claims into families, and it chooses which ones to send to an expensive retrieval-based checker using a bandit rule. A bandit rule is a method that balances testing where errors are known to be common against testing where the error rate is still uncertain. The same uncertainty estimate that tells the system where to look also decides which unchecked claims get hedged before release. The goal is a system that catches more hallucinations per verification call over a stream of queries and that does not develop blind spots when the query mix shifts.
Core Claim
Selective verification is a learning problem, and the feedback it learns from is censored: a claim's truth is observed only if the claim is chosen for checking. Under this censoring, a greedy selector that checks the claims its risk model rates as most likely false will under-sample claim families where the model is confident but wrong. The risk model never corrects itself on those families. An optimistic selector keeps an upper confidence bound on each family's hallucination rate, so families that have gone unchecked for a long time become attractive to check. We claim this selector will find more hallucinations per check over a query stream, recover faster after a shift toward rare entities, and yield less biased estimates of the true hallucination rate. We also claim the same upper bound resolves the apparent conflict between optimistic exploration and pessimistic verification. A high upper bound on "this claim is false" is a reason to check it when budget remains, and a reason to hedge or drop it when budget has run out.
Concretely, for claim family $c$ with $n_c$ past checks and an observed error fraction $\hat p_c$, the index is
$$U_c = \hat p_c + \beta \sqrt{\frac{\log t}{n_c + 1}},$$
where $t$ is the number of checks made so far across the stream and $\beta$ is a tuned exploration weight. In the full system, $\hat p_c$ is replaced by a contextual prediction that also uses the claim's own features. Claims are checked in decreasing order of $U_c$ until the per-answer budget is spent. Unchecked claims whose $U_c$ exceeds a release threshold $\tau$ are flagged or removed.
Positioning Against Prior Work
The security and surveillance line supplies the method and the motivating evidence. Eva allocated COVID-19 tests at Greek borders with an optimistic Gittins-index bandit over traveler types. It found 1.85x more infections than random testing, and it showed that population-level risk signals predicted traveler infection at chance (AUROC about 0.5). Strategic APT monitoring combined a sequence model's "malevolence" score with a UCB bonus for time since last audit. It reached full detection at timestamp 39, against 54 for greedy monitoring. BOBA showed that splitting a huge candidate space into embedding clusters and running UCB1 over the clusters cuts expensive oracle calls while recovering over 90% of top hits; it also found that ε-greedy degrades as the number of clusters grows. We reuse BOBA's cluster-as-arm structure and the APT monitor's split between a learned risk score and a staleness bonus. What we change is the object: the thing being audited is an LLM's own claims, not travelers or processes.
The hallucination line motivates the gap and constrains the design. FActScore shows that precision falls sharply for rare entities and later in a generation. That makes risk heterogeneous and structured, so grouping claims into families should carry signal. Tang et al. document that models state tentative findings as settled facts, and HaluEval reports that ChatGPT detects hallucinations in summaries with only 58.53% accuracy. Both are reasons to distrust the model's confidence as the sole selection signal. Chain-of-Verification and EVER verify every claim or every sentence. EVER explicitly leaves "selective verification based on model confidence" to future work, and our strongest simple baseline is exactly that suggestion.
Three papers come closest. VoG uses a bandit, but only to choose how much context to supply when revising a knowledge-graph reasoning plan after an error is found. It does not decide what to check. Bayesian semantic entropy on a budget and guided semantic exploration adapt how many samples to draw for a single short question. Their allocation happens within a prompt, and their feedback is not censored across a stream. We allocate across claims and across queries, and we learn from the outcomes. Discriminative verification contributes two things: a cheap scorer, which we use as a context feature, and its pessimistic scoring of small answer clusters, which our release rule mirrors.
Research Program
The system has four parts. The first is a claim decomposer, borrowed from FActScore, that splits each response into atomic claims. The second is a family assignment that clusters claim embeddings with K-means. Each claim is also tagged with its relation type and its position in the response. The third is a contextual risk model: Bayesian logistic regression over features available at test time. These features are the minimum token log-probability of the claim, a one- or two-sample semantic-entropy estimate following the budgeted Bayesian method, a cheap discriminative verifier score, position, and family identity. We run it as either LinUCB or Thompson sampling, where Thompson sampling means choosing actions by drawing from the model's uncertainty over error rates. The fourth is the expensive oracle, retrieval-plus-LM checking against Wikipedia as in FActScore, called at most $B$ times per response. Checked claims that fail are corrected EVER-style. Unchecked claims are released or hedged by the threshold rule above.
The analysis centers on a measurable version of confirmation bias. We will compare the risk model's calibration error on high-confidence claim families, measured against exhaustive ground truth, under greedy selection and under optimistic selection. The hypothesis fails if greedy selection produces no widening blind spot.
Baselines
The naive reference points are no verification, which is the raw model output, and uniform random selection of $B$ claims per response. Random selection is the analogue of Eva's random-testing comparison. It gives unbiased audit data but low yield, and any method that cannot beat it on hallucinations caught per check has no reason to exist.
The strongest competitor for the budgeted question is confidence-thresholded selective verification. It checks the $B$ claims with the lowest model confidence, and it is the policy EVER proposes. A stronger variant is greedy-by-learned-risk, which uses our own risk model trained online but picks by predicted error probability with no exploration bonus. We found no reported result on Lacuna for confidence-based selective verification on FActScore, so we will run it ourselves. Both greedy baselines get the same features, the same oracle, the same per-response budget, and the same number of tuning runs for their thresholds as our method gets for $\beta$ and $\tau$. We also include ε-greedy with a tuned ε, because it is the cheapest form of exploration.
The ceiling is exhaustive verification. EVER reports that external retrieval raises GPT-3.5 Turbo's FActScore on biographies from 71.8% to 94.5%. Chain-of-Verification (Factor+Revise) reports raising Llama-65B from 55.9 to 71.4. We will rerun exhaustive EVER-style checking on our own base models so the comparison is on identical data.
The key ablation holds everything fixed and removes only the exploration bonus. Here, greedy uses the posterior mean and the full method uses the upper bound. A second ablation trains the risk model on exhaustive labels, which shows how much of greedy's loss comes from censored feedback. Beating the field means two things. At a budget of 30% of claims, the method must catch more hallucinations over the stream than both greedy variants and tuned ε-greedy, by more than the spread across five seeds. The released text's FActScore must also come within three points of exhaustive verification.
Why not just…?
Why not just verify the lowest-confidence claims? This is the most serious alternative. If the model's confidence were well calibrated, greedy selection would be close to optimal, and exploration would only waste budget. The evidence against it is indirect. Tang et al. and HaluEval show confident fabrication, and in the APT and border settings greedy or population-signal monitoring lagged exploratory monitoring. None of this has been measured for LLM claims, however. The exploration-bonus ablation settles the question directly. If greedy matches us, we report that confident fabrications are rare enough to ignore at this budget.
Why not just add ε-greedy random audits? Random audits do fix the censoring. The open question is whether untargeted exploration spends too much budget on already well-understood families. BOBA reports ε-greedy degrading as the number of clusters grows, but that result is from chemical search, not text. We therefore sweep the number of claim families and compare yield curves. If tuned ε-greedy ties UCB, the contribution shrinks to the measurement of confirmation bias plus the release rule, and we will say so.
Why not just run the cheap discriminative verifier on everything? The discriminative verification paper shows scalar verifiers are orders of magnitude faster than generative ones: 1.66 seconds against more than 3,400 seconds for 32 solutions. If such a scorer were accurate on factual claims, budgeting would be unnecessary. That result is on AIME and GPQA reasoning, though, and HaluEval's low detection accuracies suggest cheap judges miss factual errors. We will report the cheap scorer alone, used as a filter, as a baseline. In our method it serves only as a context feature.
Proposed Experiments
The main testbed is FActScore biographies with Llama-2-7B and Llama-3.1-8B as generators. EVER reports Llama-2-7B's base FActScore at 36.8%, which ensures plenty of errors. We first verify every atomic claim offline. With every claim's label known, any selection policy can be replayed counterfactually on identical streams, which removes oracle noise from the comparisons. Streams are ordered to include a mid-stream shift from frequent to tail entities, the regime where FActScore shows precision collapsing. Transfer tests use the Chain-of-Verification Wikidata list task and HaluEval QA.
Metrics are cumulative hallucinations caught per oracle call; FActScore of the released text at budgets of 10%, 20%, 30%, and 50% of claims; recovery time after the distribution shift; calibration error of the risk model on high-confidence families; and bias of the stream-level hallucination-rate estimate against ground truth.
The decisive checks are three. First, greedy's calibration error on confident families must grow over the stream while the optimistic selector's does not. Second, after the shift, the optimistic selector must regain its yield in fewer queries than greedy. Third, the release rule must lower the rate of false claims that survive unhedged, compared with releasing all unchecked claims, at an equal hedge rate. If the first check fails, the confirmation-bias premise is wrong for LLM claims, and the proposal reduces to a cost study.
References
- Efficient and targeted COVID-19 border testing via reinforcement learning
- Strategic Monitoring for Efficient Detection of Simultaneous APT Attacks with Limited Resources
- Target-Aware Bandit Allocation for Scalable Surrogate Optimization in Chemical Space
- FActScore: Fine-grained Atomic Evaluation of Factual Precision in Long Form Text Generation
- Evaluating large language models on medical evidence summarization
- HaluEval: A Large-Scale Hallucination Evaluation Benchmark for Large Language Models
- Chain-of-Verification Reduces Hallucination in Large Language Models
- Ever: Mitigating Hallucination in Large Language Models through Real-Time Verification and Rectification
- VoG: Enhancing LLM Reasoning through Stepwise Verification on Knowledge Graphs
- Hallucination Detection on a Budget: Efficient Bayesian Estimation of Semantic Entropy
- Efficient Hallucination Detection: Adaptive Bayesian Estimation of Semantic Entropy with Guided Semantic Exploration
- Budget-aware Test-time Scaling via Discriminative Verification