Small language models can be trained to spot hallucinations, meaning fluent statements that are not supported by the source or by the facts. The usual recipe is to have a large model write fake, subtly wrong answers and use them as negative examples. The best recent recipe orders these fakes from obvious to subtle and trains on them in that order. We argue that the next gain will not come from more fakes or subtler fakes. It will come from fakes that keep covering many different kinds of error as they get harder. When fakes are generated iteratively, with each round rewriting the fakes that fooled the detector, the loop drifts toward whichever kind of error happens to fool the detector most easily. Evolutionary search has a name for this drift: premature convergence. Evolutionary computation also has well-studied fixes for it, namely tournaments held only within a niche and balanced tie-breaking at the selection cutoff.
The proposal is to run hallucination synthesis as an evolutionary loop against the detector being trained. Each candidate fake has to pass an independent check that it really is false, so that "diverse" does not quietly become "mislabeled." Diversity is preserved in the detector's own internal representation, not in surface text. We expect the main payoff on the rare, structurally complex error types where current detectors are weakest. These include multi-hop and comparison errors, and the factual-inference category, where the base HAD detector scores 26.87 F1.
Core Claim
When hard negatives are generated adversarially over several rounds, the selection rule that decides which fakes become parents for the next round determines which failure modes the detector ever sees. Choosing fakes by a single difficulty score, as Teaching with Lies: Curriculum DPO on Synthetic Negatives for Hallucination Detection (HaluCheck) does, concentrates parents in a few niches. A different rule should do better. It holds tournaments only among fakes in the same niche, defines niches by clusters of the detector's intermediate-layer activations, and splits the survivor cutoff evenly across niches. Under a matched budget of generator calls, this rule should yield a detector with substantially higher worst-category F1 and equal or better average F1.
The claim is specifically about the generative feedback loop. A one-shot pool that is diversified after the fact cannot steer later generation, because nothing gets rewritten. Our experiments are built to test whether this distinction matters.
Relation to Prior Work
Two papers supply the mechanism. Speeding Up the NSGA-II with a Simple Tie-Breaking Rule studies the cutoff point in multi-objective evolutionary search. It proves that random tie-breaking at that cutoff starves rare objective signatures and causes exponential runtime, while spreading the cutoff evenly across groups restores polynomial runtime. Fitness Uniform Selection to Preserve Genetic Diversity shows that subsidizing sparsely populated fitness levels speeds up search on deceptive landscapes. These results come from combinatorial benchmarks, not language. They motivate our selection operator but do not guarantee it will work here.
Three papers motivate the gap by synthesizing hallucination data without any feedback-driven diversity control. HaluCheck scores fakes with the MiniCheck fact-checker and runs a curriculum version of direct preference optimization (DPO, a method that trains a model to prefer one response over another from paired examples). The curriculum runs along that single score, from easy to subtle fakes. HAD injects errors from an 11-category taxonomy into 90,172 GPT-4o-written samples using static prompts. PALE generates paired truthful and hallucinated responses with prompts, then separates them by Mahalanobis distance in an intermediate layer (layer 11 of a LLaMA-3.1 model).
We reuse PALE's finding that intermediate activations separate truth from hallucination, but we use those activations to define niches during generation rather than to detect after the fact. We reuse HAD's taxonomy as the menu of rewrite instructions for producing new fakes, not as a fixed quota.
Two papers constrain the design. Diversity-oriented Data Augmentation with Large Language Models (DoAug) maximizes embedding distance from the source, gaining 10.52% on average over 12 tasks. Its authors concede that pushing for distance invites hallucination and label drift. For a hallucination detector, this is fatal in a specific way: a "fake" that is actually true is a mislabeled negative. We therefore treat verified falsity as a hard constraint and never trade it against diversity. When Dynamic Data Selection Meets Data Augmentation makes the same point for images, where augmenting sparse samples helps only if they are verified clean. The fitness-distance balance literature, including Chaotic Wind Driven Optimization with Fitness Distance Balance Strategy, reports that Euclidean distance loses discriminative power in high dimensions. This finding is why we form niches by clustering a low-dimensional projection, not by raw distance.
Research Program
The loop starts from source documents and questions drawn from the training splits of HaluEval and FactCHD. The population is a set of candidate fakes. In each generation, an instruction-tuned generator applies mutation operators to parent fakes. A mutation is a targeted rewrite following one of the HAD error categories, such as swapping an entity, hardening a hedged claim into a certainty, or composing two facts incorrectly. There is also a recombination operator that merges the errors of two parents into one answer.
Every child must pass a falsity check before it can enter the population. Two independent checks must agree that the child contradicts or is unsupported by the source: a natural-language-inference contradiction check, and MiniCheck applied to the edited span compared with the original. We audit a random 200 children per generation by hand. We report the auditors' false-negative-label rate as a first-class result, not a footnote.
Fitness is the current detector's probability of calling the fake faithful, which measures how well the fake fools it. Niches are defined freshly each generation. We take intermediate-layer activations of the current detector on each fake, project them to about 32 dimensions, and cluster them with k-means. Parents are chosen by binary tournaments held only within a niche, so a niche full of easy-to-fool fakes cannot monopolize reproduction. Survivors are ranked on two objectives, fitness and distance to the nearest already-kept fake in the projected space. When the cutoff falls inside a rank, the remaining slots are dealt evenly across niches in round-robin, which is the NSGA-II balanced rule transplanted.
After each generation, the surviving fakes are paired with their verified-faithful counterparts. The detector, Llama-3.2-3B as in HaluCheck, takes one round of DPO on these pairs. The generations themselves form the curriculum, since later generations are by construction harder for the detector. We run eight generations at a fixed total of generator calls.
Baselines
The naive reference point is the Llama-3.2-3B backbone prompted zero-shot as a detector. Next is the same backbone fine-tuned once on an equal number of uniformly sampled, statically prompted fakes. This measures what data volume alone buys.
The strongest existing method is HaluCheck's MiniCheck-ordered curriculum DPO. It is the closest prior system on exactly this question and reports average F1 of 75.6% for HaluCheck 3B and 0.759 F1 on MedHallu, against 0.737 for GPT-4o. We re-run it on the same source documents, with the same generator, the same number of generator calls (counting all calls, including rejected ones), the same DPO hyperparameter search budget, and the same three seeds. To beat it, our method must match or exceed its average F1 within seed noise while raising worst-category F1 on HADTest by a margin larger than three seed standard deviations. We also report HAD-14B, a larger supervised detector with retrieval, as a ceiling reference: 89.10% binary accuracy on HADTest and 92.37% on HaluEval-QA. We do not claim to beat it at 3B.
Three ablations isolate the key ingredient. The first is an adversarial loop with plain elitist top-k survivor selection, which is identical in every respect except niching and balanced tie-breaking. The second replaces activation-defined niches with HAD category labels. The third uses sentence-embedding niches in the style of DoAug instead of detector activations.
Why not just…?
Why not just stratify a large one-shot pool by HAD category and run HaluCheck's curriculum within each stratum? This is the most serious threat, because balanced tie-breaking over fixed categories is close to stratified sampling. If the categories capture what matters, stratifying a static pool could match our results. Our position is that categories are assigned by the generator's prompt, not by what the detector actually confuses, and that a static pool never rewrites the fakes the detector has learned to catch. The category-niche ablation and a matched-budget stratified one-shot baseline test this directly. If either matches the full method on worst-category F1, we will drop the evolutionary loop and report stratification as the answer.
Why not just pick a diverse subset from one big pool, for example with greedy farthest-point selection in embedding space? This kind of diversity-after-the-fact selection is what DoAug and the dynamic data selection paper do. It cannot create fakes in niches the generator rarely visits on its own. We run it at a matched generator-call budget. We also measure how many activation clusters each method populates by generation eight. If coverage is equal and F1 is equal, the loop is not needed.
Why not just raise sampling temperature or ask for "diverse" errors? DoAug's ablations found that high-temperature sampling and diversity prompting performed significantly worse than distance-driven selection, and that high temperature degraded label consistency. Here, degraded labels mean true statements labeled as hallucinations. We include temperature 1.2 generation as a baseline and report its audited mislabel rate alongside F1.
Why not just use a bigger detector? In HaluCheck's comparison, Qwen-2.5 14B reached only 72.0% F1 against 75.6% for the curated 3B model, and TruthfulQA shows truthfulness does not reliably improve with scale. Scale is therefore not a substitute for training data. Still, we report the method applied to a 7B backbone to check that the gains survive scaling.
Proposed Experiments and Decisive Checks
We evaluate on HaluEval (average F1, comparable to HaluCheck's 75.6%), MedHallu (F1, against 0.759), and CovidQA zero-shot (HaluCheck reports +6.4% accuracy over the backbone). We also evaluate on HADTest per-category F1, with worst-category F1 and the factual-inference category (26.87 base, 52.17 with retrieval in HAD) as headline numbers. FactCHD supplies multi-hop, set-operation, and comparison subsets; TRUTH-TRIANGULATOR reports a FACTCLS score of 78.15% there, which we cite as context, not a matched comparison. On TruthfulQA we report AUROC against PALE's 73.20%.
The decisive check has three parts. First, the full method should beat the elitist adversarial loop on worst-category F1 while holding average F1. Second, it should beat both the category-stratified and farthest-point one-shot baselines on worst-category F1 under an equal generator-call budget. Third, its audited mislabel rate should be no higher than HaluCheck's pipeline.
We will also track the share of the population held by the largest niche across generations. If our diagnosis is right, that share should climb under elitist selection and stay flat under niche-preserving selection. The worst-category F1 gap should follow this share. If the share climbs under both selection rules, or the F1 gap does not track it, the premature-convergence story is wrong regardless of headline F1.
References
- Teaching with Lies: Curriculum DPO on Synthetic Negatives for Hallucination Detection
- Speeding Up the NSGA-II with a Simple Tie-Breaking Rule
- Fitness uniform selection to preserve genetic diversity
- HAD: HAllucination Detection Language Models Based on a Comprehensive Hallucination Taxonomy
- Bolster Hallucination Detection via Prompt-Guided Data Augmentation
- Diversity-oriented Data Augmentation with Large Language Models
- When Dynamic Data Selection Meets Data Augmentation: Achieving Enhanced Training Acceleration
- Chaotic Wind Driven Optimization with Fitness Distance Balance Strategy
- TruthfulQA: Measuring How Models Mimic Human Falsehoods