Vision-language models describe images in fluent text, but they often name objects, attributes, or relations that are not in the picture. The most popular fixes act at decoding time, while the model is generating text. They do extra work at certain tokens, such as re-scoring the token against a noised image, a cropped view, or a summary of the context so far. Every one of these fixes needs a rule for when to do that extra work, and the rules are hand-written. HALC and Summary-Guided Decoding intervene only on nouns and adjectives. OPERA rolls back generation when an attention pattern crosses a fixed threshold. These are static if-then rules of the kind that intrusion detectors used before learning loops were added. They fire too often where they are not needed and stay silent on prepositions and verbs, where relational hallucinations live.
We propose replacing these hand-written triggers with a small rule whose few parameters are fitted by gradient descent. The design is modeled on CAPTAIN, an intrusion detector that turned its hand-written alarm rules into learnable trust values, spread rates, and thresholds. The new ingredient is propagation. Each generated token gets a "suspicion" score from its own weak grounding in the image. That suspicion then flows forward along the model's attention to later tokens, and an intervention fires only when accumulated suspicion crosses a learned threshold. The bet is that hallucinations in long captions are inherited from earlier ungrounded tokens, much as one bad event contaminates later events in an intrusion graph. If so, a gate that tracks this inheritance should catch cascaded errors and fire less often on safe tokens, so expensive corrections run on a small fraction of tokens.
Core Claim
Hallucinations in long vision-language outputs often arise because the model starts conditioning on its own earlier, ungrounded text instead of the image. A gate that explicitly propagates suspicion along attention edges from earlier generated tokens, with a few dozen fitted parameters, should do two things better than either fixed part-of-speech rules or per-token detectors. It should have a better trade-off between intervention rate and hallucination reduction on long captions, and it should transfer better across prompt phrasings and model families. If that holds, a standard corrective step such as visual contrastive decoding (VCD) can be applied selectively, approaching its full-strength hallucination reduction at a fraction of its latency.
Two lines of evidence motivate this claim. Why LVLMs Are More Prone to Hallucinations in Longer Responses shows that long-response hallucination comes from attention drifting from image tokens to the model's own generated tokens, a tension between coherence and completeness, rather than from independent per-token errors. Evaluating Object Hallucination in Large Vision-Language Models shows that co-occurrence statistics such as table-chair and laptop-mouse dominate when the visual signal is weak. Together these describe a dependency structure: one hallucinated "table" licenses a hallucinated "chair."
Two papers from the rule-rigidity literature supply the method. Incorporating Gradients to Rules turned hand-written propagation rules over host provenance graphs into three learnable quantities: initial trust per entity, propagation rate per edge, and an alarm threshold. Fitting them on benign traces cut false alarms by 11.49x compared with the static rule system MORSE, while staying lightweight. Its online follow-up, Toward Online, Adaptive Provenance-Based Intrusion Detection, supports the idea that such parameters can be refit cheaply. Neural-symbolic monitoring and adaptation adds the lesson that adapting the right internal node of a rule beats adapting its leaves: 100% versus 73–76% accuracy on drifted traces. We take this as a constraint, explained below.
None of these rule-learning systems has been applied to token generation. In the other direction, no hallucination decoder learns its triggering rule.
Research Program
The model generates a caption. For each new token $t$ the gate computes a suspicion score
$$s_t = \sigma\Big(a_{c(t)}\, u_t + \sum_{j<t} g_{c(j)\to c(t)}\, w_{tj}\, s_j - T\Big).$$
The symbols are as follows. The function $\sigma$ is the logistic function. The term $u_t$ is a local ungroundedness signal that comes free from the forward pass: one minus the share of middle-layer attention on image tokens, combined with the token's predictive entropy. The weight $w_{tj}$ is the middle-layer attention from token $t$ to earlier generated token $j$. The term $c(t)$ is a soft assignment of the token to one of $k \approx 8$ learned classes, produced by a tiny linear map on the middle-layer hidden state. The factor $a_c$ is a per-class initial-suspicion weight, $g_{c\to c'}$ is a per-class-pair propagation rate, and $T$ is the threshold.
When $s_t > 0.5$, the decoder applies a corrective step at that token only. The default corrective step is a VCD pass. As a secondary arm, the corrective step is a SHARP steering vector.
Following the RuleRunner finding, the learned classes are the "operator" the rule adapts. They replace the hand-written noun/adjective list rather than adding detail at individual tokens. The whole gate has well under a thousand parameters, and the rates $g$ can be read directly. We expect, for example, that object classes will pass suspicion strongly to later object and relation classes.
Training follows CAPTAIN's recipe of calibrating against benign behavior, extended with positive examples. We generate captions with LLaVA-1.5 on COCO training images. Tokens are labeled as hallucinated or grounded by the matching procedure used for CHAIR (Caption Hallucination Assessment with Image Relevance, which checks generated object words against the image's annotated objects). We then minimize a weighted loss with two parts: false triggers on grounded tokens, and misses on hallucinated tokens and on the tokens immediately following them. A penalty on the expected trigger rate keeps interventions within a fixed budget. Labels for relational hallucination are sparse, so relation-bearing tokens enter only through the propagation term during training. Relational behavior is then evaluated, not supervised, on MERLIM's plausible-but-false relation set.
The research then asks three questions:
- Does propagation help beyond per-token signals?
- Does a few-parameter rule transfer across prompts and models better than a high-dimensional probe?
- How much of full-strength VCD's hallucination reduction survives when it runs on a small fraction of tokens?
Baselines
The naive reference is the unmodified model with greedy and beam decoding. The strongest always-on competitor is Visual Contrastive Decoding, which contrasts logits from the clean image and a noised image at every token. It reports up to +7.4 F1 on POPE MSCOCO random and raises LLaVA-1.5's MME total from 565 to 604. We run it at every token with the authors' plausibility constraint. The comparison asks whether our gate, using the same VCD step, matches its CHAIR scores while intervening on, say, under 30% of tokens. That would be a clear latency win, given the 13.97s per sample SHARP measured for VCD.
The strongest learned-signal competitor is SHARP itself. It applies a linear steering vector at middle layers, where faithful and hallucinated states are linearly separable, and reports LLaVA-v1.5 CHAIRs dropping from 52.8 to 34.8 at 7.04s per sample. Beating it on CHAIRs at equal or lower latency would be a strong result. Matching it while reporting fewer interventions would be a modest one.
The strongest static-rule competitors are Summary-Guided Decoding and HALC. Summary-Guided Decoding reports a 16.5% CHAIRs improvement on LLaVA-1.5. HALC reports reducing MiniGPT-4 CHAIRs from 30.87% to 17.80%. To make the comparison fair, we also run "HALC/SGD-trigger + VCD": the same corrective step, gated by their noun/adjective rule. Any difference then comes from the gate alone. We also include OPERA, with its threshold and window tuned by grid search on our training captions. OPERA reports about a 35% CHAIR improvement over DoLa.
Two ablations isolate the key ingredient. The first sets every $g$ to zero, giving a learned per-token rule with no propagation. The second replaces the whole gate with a logistic probe on the full middle-layer hidden state, trained on identical labels. This is a direct application of SHARP's separability finding, and it is the most dangerous competitor.
All gated variants share the same labeled captions, the same trigger-rate budget, and the same hyperparameter search budget. Each is evaluated as a curve of hallucination rate against intervention rate, not at a single operating point.
Why not just…?
Why not just run VCD (or RITUAL) at every token? That is the ceiling we aim to approach, not beat, on hallucination. The case against it is cost and fluency. Always-on contrast roughly doubles decoding compute. Summary-Guided Decoding reports that contrastive decoding degrades text quality, and RITUAL reports diminishing returns from additional transformed passes; RITUAL lifts LLaVA-1.5 POPE accuracy from 84.13% to 88.87%. If our gated curve sits on the always-on VCD point only at intervention rates near 100%, the gate is pointless, and the experiment will show it.
Why not just train a linear probe on hidden states as the trigger? This is a real threat. The hidden state at token $t$ already mixes in earlier context, so it may encode inherited suspicion implicitly. SHARP's separability result and HalTrapper's 5–12% AUROC gains over the PoScore baseline both suggest probes are strong. We address this directly with the probe ablation. The propagation gate must win on at least one of three tests to justify itself:
- Detection of hallucinations occurring after an earlier hallucination in the same caption.
- Transfer to MERLIM prompt paraphrases, where models vary by up to 26.5% in MERLIM.
- Transfer of the fitted gate from LLaVA-1.5 to MiniGPT-4 with only $T$ refit.
If the probe ties on all three, we will report that propagation adds only interpretability.
Why not just grid-search the thresholds of existing hand rules? Tuned OPERA and the "POS-trigger + VCD" baseline are exactly this alternative. The CAPTAIN result suggests that tuning a threshold without learning where suspicion flows leaves most false alarms in place. The hallucination-side test is MERLIM's relation subset, where models drop 11.63% on plausible false relations. A noun/adjective trigger cannot fire there by construction, while the learned classes can. Tuned static rules could still match us on object-only CHAIR, and that would narrow the claim to relational and long-caption settings.
Why not just post-correct with Woodpecker? Woodpecker raises MiniGPT-4 POPE accuracy from 54.67% to 85.33%. However, it chains several external models and mis-corrects about 6.2% of the time. It is complementary rather than competing: our gate's suspicion scores could tell Woodpecker which spans to verify. We include it only as a reference on CHAIR.
Proposed Experiments
The main test uses LLaVA-1.5, which has the densest set of reported baselines, generating long COCO captions on the standard 500-image CHAIR split. We report CHAIRs, CHAIRi, object recall, and wall-clock time per sample.
The decisive measurements are three:
- Intervention-rate curve. The curve of CHAIRs against intervention rate for propagation gate, zero-propagation gate, probe gate, and part-of-speech gate, all using the same VCD step.
- Hallucinations by caption position. CHAIRi split into hallucinations occurring before and after the first hallucinated object in each caption. The propagation hypothesis predicts its advantage concentrates in the "after" bucket.
- Transfer. MERLIM relation accuracy and the variance across MERLIM prompt paraphrases, plus cross-model transfer to MiniGPT-4.
MME existence, count, and position, and POPE adversarial, serve as sanity checks that short-answer performance does not drop. POPE answers are a single token, so propagation cannot help there, and no gain is expected.
The proposal is falsified in two cases. The first is if the "after" bucket shows no gain over the zero-propagation ablation. The second is if the probe gate dominates the propagation gate on both cascade detection and paraphrase transfer.
References
- Why LVLMs Are More Prone to Hallucinations in Longer Responses: The Role of Context
- Evaluating Object Hallucination in Large Vision-Language Models
- Incorporating Gradients to Rules: Towards Lightweight, Adaptive Provenance-based Intrusion Detection
- Incorporating Gradients to Rules: Toward Online, Adaptive Provenance-Based Intrusion Detection
- Neural-symbolic monitoring and adaptation
- Mitigating Object Hallucinations in Large Vision-Language Models through Visual Contrastive Decoding
- SHARP: Steering Hallucination in LVLMs via Representation Engineering
- Mitigating Hallucinations in LVLMs via Summary-Guided Decoding
- HALC: Object Hallucination Reduction via Adaptive Focal-Contrast Decoding
- OPERA: Alleviating Hallucination in Multi-Modal Large Language Models via Over-Trust Penalty and Retrospection-Allocation
- RITUAL: Random Image Transformations as a Universal Anti-hallucination Lever in LVLMs
- Behind the Magic, MERLIM: Multi-modal Evaluation Benchmark for Large Image-Language Models
- Woodpecker: hallucination correction for multimodal large language models