Budgeting Reader Attention to Catch LLM Hallucinations

Language models state false things fluently, and the last line of defense is usually a person reading the answer. That person has a limited supply of attention. A neuroimaging study found that catching an AI hallucination costs a measurable burst of attentional effort, and that once a plausible falsehood slips past, the brain responds exactly as it does to true text. There is no residual "gut feeling" to fall back on (How do Humans Process AI-generated Hallucination Contents: a Neuroimaging Study). Long, dense answers therefore do two kinds of harm. They contain more claims that could be wrong, and they drain the vigilance a reader needs to catch those errors. This proposal treats the reader's attention as a fixed budget. The system decides, claim by claim, where to spend it. Claims the evidence supports well are compressed. Claims that are weakly supported are coarsened or dropped. The few claims that are both valuable and uncertain are isolated, with their evidence, where the reader will actually check them.

The quantity we optimize and measure is not how many false claims the model emits. It is how many false claims the reader ends up believing, per minute of reading, set against how much true information the reader takes away. No existing method optimizes this end-to-end quantity. Hallucination work stops at the text, and cognitive-load work stops at comprehension and satisfaction. Classic spoken-dialogue research already showed that how much information to present at once should be planned against the listener's memory limits (Optimising Information Presentation for Spoken Dialogue Systems). We bring that planning view to the problem of human verification of model output.

Core Claim

Given a draft answer broken into atomic claims, each with an estimated probability of being supported, a presentation policy chooses one action per claim under a reading-effort budget $B$. The four actions are omit, coarsen, state compactly, or flag-with-evidence. The policy maximizes

$$U = T_{\text{acc}} - \lambda \, F_{\text{acc}} \quad \text{subject to} \quad E \le B,$$

where $T_{\text{acc}}$ is the number of true claims the reader accepts, $F_{\text{acc}}$ is the number of false claims the reader accepts, $\lambda$ is how much worse a believed falsehood is than a missed truth, and $E$ is reading effort, measured as reading time or a claim-weighted length proxy.

The central hypothesis is that the probability a reader rejects a false claim is not fixed. It falls as claim density and position in the answer grow, and it rises sharply when a claim is visually and structurally isolated with a checkable source. If that holds, then under an equal effort budget, allocating attention beats both simply shortening the answer and simply removing unsupported claims.

Relation to Prior Work

Three papers motivate the gap. The neuroimaging study measured that human hallucination detection is effortful and all-or-nothing. It used word-by-word lab presentation of image descriptions and reported 84.29% behavioral verification accuracy on AMBER. It did not vary how output is presented, and we do exactly that. Verbosity ≠ Veracity showed that models pad their answers when they are uncertain, with verbose answers 27.61% less accurate on Qasper. That means length itself carries a signal about where errors are. We reuse the insight as a feature, but we act on the reader side rather than routing to a bigger model as its cascade method does. LikeBench and VibeCheck document a "helpfulness trap" and an "overexplanation trap" in frontier models. This is evidence that scale alone does not fix length, though neither paper measures what readers come to believe.

Two papers supply methods. Answer Only as Precisely as Justified introduces calibrated claim-level "specificity ladders," which replace an unsupported fine-grained claim with a coarser one or omit it. We reuse its extraction, backoff, and support-scoring pipeline as our omit and coarsen actions. The difference is that it optimizes the text, whereas we optimize the reader's resulting beliefs under an effort budget and add a flag action it lacks. Long-form factuality in large language models supplies LongFact and SAFE, a search-based claim checker, which gives us claim-level ground truth.

Two papers constrain the design. ConciseRL shows that naive token-length penalties hurt hard problems and that brevity must be gated on correctness. This rules out a pure length objective, which is why our objective counts claims and beliefs rather than tokens. Be Helpful but Don't Talk too Much found that penalizing listener effort, measured as surprisal under a simulated user, stopped its agent from gaming a helpfulness reward through repetition and hallucination. That is encouraging, but its effort proxy ignores verification difficulty, which is the thing we measure directly with people.

Research Program

The first stage is a measurement study, and it can falsify the proposal. We generate long-form answers with Llama-3.1 8B and 70B on LongFact and Biography prompts, the setting used by Improving Factuality with Explicit Working Memory. We label every atomic claim with SAFE and expert spot-checks. Because naturally occurring errors are unevenly placed, we also construct matched answers by swapping in controlled false claims of the entity, attribute, and relation types that the neuroimaging study found differ in difficulty. Participants read the answers in a normal scrolling interface and mark claims they would not trust. We record reading time and a standard subjective workload questionnaire. From this data we fit a per-claim model of $P(\text{reject} \mid \text{false}, \text{density}, \text{position}, \text{format}, \text{claim type})$ and the matching false-rejection rate for true claims.

The second stage builds the policy. Given support scores from the specificity-ladder pipeline and the fitted reader model, choosing actions under budget $B$ is a small knapsack problem. We solve it exactly per answer; no reinforcement learning is needed. We calibrate $\lambda$ and the support thresholds on a held-out split, as the specificity-control paper does. We also test whether early-token perplexity spikes, the signal found in the verbosity paper, improve the support estimates cheaply.

The third stage is a fresh human evaluation at three budgets, with participants and prompts disjoint from the fitting data. It includes a smaller replication with domain-knowledgeable readers, because the neuroimaging study's non-expert pool leaves open whether experts behave differently.

Baselines

The naive reference point is the unedited model draft shown as-is. Its reported text-level score is an Overcommitment-Aware Utility (OAU, the specificity-control paper's metric that rewards retained detail while penalizing unsupported claims) of 0.846 on LongFact, per Answer Only as Precisely as Justified. We measure its believed-falsehood rate directly. A second weak baseline is the same model prompted to be concise. It is cheap, and the verbosity paper suggests it is unreliable, but it defines the effort floor.

The strongest existing text-side competitor is calibrated specificity control from Answer Only as Precisely as Justified. It reports OAU 0.913 with 0.9865 support precision and 0.938 specificity retention on LongFact. We run it with the same drafts, the same support scorer, and the same calibration split as our policy, then truncate or pad to the same effort budget. Explicit working memory (Improving Factuality with Explicit Working Memory) serves as a strong generation-side competitor. It reports gains of 2 to 6 VeriScore F1 points over retrieval-augmented baselines, and we run it on the same prompts and backbone.

The strongest reader-side competitor is uncertainty highlighting: the unedited draft with low-support claims visually marked. We found no Lacuna paper reporting a result for this, so no reported result is found on Lacuna for highlighting. We also run specificity control combined with highlighting.

Two ablations isolate the key ingredient. One uses budget-constrained allocation without the flag action. The other uses the flag action without any budget. For human-side metrics none of these baselines has a reported number on Lacuna, so all are measured fresh under identical conditions. That means the same interface, randomized assignment within participants, and the same prompt set.

Beating the baselines means lower believed-falsehood rate at equal or higher accepted-true-claim count and equal reading time, compared with specificity control plus highlighting. The evidence must be a confidence interval on the paired difference that excludes zero. Beating them must also not reduce OAU below the specificity-control number at matched budget.

Why not just…?

Why not just highlight the uncertain claims? This is the most serious threat, and we treat it as one. If highlighting alone lets readers catch nearly all flagged falsehoods without extra fatigue, the budget and restructuring machinery adds little. The highlighting and flag-without-budget conditions test this directly. Our prediction, from the neuroimaging finding that vigilance is costly, is that highlighting inside a long dense answer helps less as the number of flags and the answer length grow. The measurement stage will show whether that interaction exists. If it does not, we will report that highlighting suffices and that the contribution is the believed-falsehood evaluation itself.

Why not just run specificity control and ship the cleaned text? Its support precision of 0.9865 leaves some unsupported claims. More importantly, it has no notion of which remaining claims readers will scrutinize, so it cannot trade detail for vigilance. If specificity control at matched length already drives believed falsehoods near zero, our policy's advantage vanishes. The head-to-head comparison at three budgets settles this.

Why not just make answers shorter? Prompted brevity is unreliable according to Verbosity ≠ Veracity. Blind length penalties cut necessary content according to ConciseRL. Shortening also removes true and false claims indiscriminately. The concise-prompt baseline and the effort-matched comparison measure how much of any gain is due to length alone.

Why not just generate more factual text in the first place? Better generation, such as explicit working memory, reduces errors but does not eliminate them, and residual errors are exactly the fluent ones readers miss. The two approaches are complementary, so we also apply our policy on top of working-memory outputs to check that the gains add up.

Decisive Checks

The first check is whether, in the measurement study, rejection of false claims actually declines with claim density or position and improves with isolation. If neither effect appears, the load-based premise fails, and the project reduces to specificity control plus an evaluation protocol. The second check is whether the budgeted policy lowers believed falsehoods relative to specificity control plus highlighting at matched reading time, on held-out participants and prompts. The third check is whether the fitted reader model, trained on non-experts, still predicts expert behavior well enough that the policy's ranking of conditions is preserved. Together these outcomes determine whether managing the reader's attention is a lever against hallucination separate from producing fewer errors.