Source-linked AI summary
UniProbe: A Learnable Token-Level Hallucination Detector for Large VLMs using Multi-Structural Internal Representations
Dvir Samuel, Guy Bar-Shalom, Fabrizio Frasca, Ethan Fetaya, Yftah Ziser, Gal Chechik, Haggai Maron
TL;DR
LVLMs often generate visually unsupported text, making token-level hallucination localization important for targeted intervention. UniProbe learns from frozen-model computational traces using multiple structural representations and improves hallucination detection across backbones, while reducing object hallucinations during decoding by up to 55% at 1.06× latency.
Problem
LVLM hallucinations make accurate token-level localization a central challenge for targeted intervention in trustworthy multimodal AI.
Method
UniProbe processes a frozen LVLM’s computational trace with interleaved graph, visual-geometry, and sequential modules, plus streaming rejection-resampling during generation.
Results
Token-level F1 improves by 4–6 points over the strongest trained baseline, while UniProbe consistently outperforms state-of-the-art detectors across three tasks and diverse backbones.
Takeaways & Limitations
Structured internal model traces provide a practical signal for hallucination detection and mitigation while keeping the underlying LVLM frozen.
Takeaways & Limitations
UniProbe requires hidden states and attention maps, applies only to accessible models or providers, and needs a separate readout for each backbone.
Abstract
from arXiv · showhide
Large Vision-Language Models (LVLMs) achieve impressive visual reasoning and dialogue capabilities, yet frequently hallucinate content unsupported by the visual input. Effective mitigation requires token-level localization, enabling targeted intervention without discarding the entire response. Existing detectors require expensive full-model fine-tuning, rely on external verifiers that ignore the model's generation process, or reduce internal signals to isolated features and hand-crafted statistics, discarding spatial, sequential, and relational structure. We introduce \textbf{UniProbe}, a lightweight, unified, learnable detector that models a frozen LVLM's heterogeneous computational trace from a single forward pass. UniProbe constructs a directed graph over image patches, query tokens, and generated tokens, with attention weights encoding their relations. It processes this trace with alternating structure-aware modules: a GNN for relational evidence, a ViT for 2-D visual geometry, and a GRU for response order. Interleaving them allows spatial, relational, and sequential evidence to interact throughout the detector. We further develop a streaming variant for hallucination-aware decoding, which detects and resamples hallucinated tokens during generation, and a self-adaptation strategy aligning the detector with the LVLM's own generations. Across diverse LVLM backbones, UniProbe achieves state-of-the-art token-level and object-hallucination detection. During decoding, it reduces object hallucinations by up to 55\% at $1.06\times$ the latency of standard generation.
1 Introduction
UniProbe is a lightweight detector that models a frozen LVLM’s heterogeneous computational trace to localize hallucinated tokens and support targeted mitigation. It extends this detector to streaming decoding and host-model adaptation, achieving strong performance across detection and mitigation tasks.
- Motivation: Token-level hallucination localization enables targeted intervention instead of discarding an entire response.The paper identifies precise affected tokens as necessary for effective mitigation.
- Method: UniProbe reads a frozen LVLM’s heterogeneous computational trace from a single forward pass and models image, query, and response tokens as a directed graph.Alternating GNN, ViT, and GRU modules capture relational evidence, 2-D visual geometry, and response order.
- Method: The online variant assesses each candidate token during decoding, rejecting flagged hallucinations and resampling alternatives before they reach the final response.This enables hallucination-aware decoding rather than only post hoc detection.
- Method: UniProbe adapts to responses sampled from the target LVLM, addressing train–test distribution shift caused by backbone-specific hallucination patterns.The adaptation aligns the detector with the host model’s generation dynamics.
- Results: Token-level F1 improves by 4–6 points over the strongest trained baseline, while POPE object-hallucination detection rises from 41.0 to 63.1 F1 with the backbone frozen.Evaluation covers token-level detection, object-hallucination detection, and real-time detection and mitigation across diverse open-source LVLM backbones.
2 Related Work
Prior hallucination detectors probe internal representations, structured computational traces, or external verification, while decoding-time methods intervene through attention, image contrasts, or learned steering. UniProbe instead uses a lightweight frozen-backbone detector that models the full heterogeneous multimodal trace for token-level localization and streaming rejection/resampling.
- Internal-representation detection: Recent detectors probe hidden states, steer activations, or model structured traces such as attention maps and activation tensors.These approaches extend beyond single-vector probes by treating computational traces as structured data.
- Internal-representation detection: Multimodal hallucination detection must represent heterogeneous traces spanning image patches, query tokens, and generated response tokens.This structure distinguishes multimodal computational traces from unimodal internal representations.
- LVLM hallucination detection: Existing LVLM approaches use external verification, full-model fine-tuning, or internal probing, with some methods predicting only a global hallucination label.HalLocalizer compares images and responses externally, whereas HALP probes backbone representations but does not localize hallucinations at the token level.
- LVLM hallucination detection: UniProbe keeps the backbone frozen and treats its computational trace as a structured representation for lightweight, learnable token-level hallucination localization.This distinguishes it from approaches relying on external verification or full-model fine-tuning.
- Decoding-time mitigation: Existing decoding-time mitigation methods penalize attention patterns, contrast original and distorted images, correct completed responses, or steer latent states toward truthfulness.OPERA, VCD, Woodpecker, Lookback Lens, and TruthPrInt represent complementary intervention strategies during or after decoding.
- Decoding-time mitigation: UniProbe extends detector-guided decoding to multimodal generation with a streaming detector over the full internal trace that rejects and resamples flagged tokens.Unlike attention-only guidance, this targeted intervention preserves the model’s output distribution while reducing hallucinations.
3 Notation and Problem Setup
The setup treats a frozen pre-trained LVLM as a backbone that processes an image and query before autoregressively generating a token sequence. UniProbe uses intermediate hidden states and attention maps to learn token-level probabilities identifying response tokens unsupported by the image or query.
- Backbone and representations: The frozen backbone receives an image I and text query x, then autoregressively generates a T-token response r.The model processes inputs and outputs as one sequence of n tokens.
- Backbone and representations: The internal sequence is partitioned into three disjoint groups: image patches, query tokens, and response tokens to verify.This partition defines the token groups used for token-level hallucination analysis.
- Backbone and representations: At intermediate layer ℓ, hidden states h_i ∈ R^d summarize computation at each token position, while attention maps encode pairwise token relations.The attention matrix A ∈ [0, 1]^(n×n) contains weights A_ij from position i to position j.
- Detection objective: The detector g_θ outputs T probabilities p ∈ [0, 1]^T, with p_i estimating whether response token r_i is unsupported by I or x.Training uses responses annotated with hallucinated spans.
4 Method
UniProbe is a lightweight detector that extracts a frozen LVLM’s internal computational trace in one forward pass and combines relational, spatial, and sequential structure for token-level hallucination prediction. It also supports causal streaming detection and self-adaptation to the target model’s own generations.
- Structure-aware detector: Alternating GNN, ViT, and BiGRU modules integrate cross-modal, 2-D spatial, and response-order evidence before a linear head predicts hallucination scores for each response token.Type-specific projections and attention-derived neighborhoods let the GNN distinguish visual and textual evidence, while the ViT and BiGRU provide within-modality spatial and sequential modeling.
- Computational-trace graph: UniProbe builds a directed attributed graph from one forward pass, representing image patches, query tokens, and response tokens with hidden states, attention, types, and patch coordinates.Image and query nodes are selected by response-to-node attention, while edges retain strongest attention links into each response token, including only earlier response tokens.
- Efficiency: Fixed node and edge budgets keep UniProbe lightweight, adding only modest overhead beyond the backbone computation it reads.
- Streaming detection: Replacing the bidirectional response GRU with a unidirectional one makes detection causal, so each token score depends only on the generated prefix.During decoding, tokens whose hallucination probability exceeds threshold τ are rejected and re-decoded; other tokens are accepted.
- Self-adaptation: Self-adaptation fine-tunes UniProbe on the target LVLM’s automatically labeled free-form generations to reduce train–test distribution shift.Captions are generated on a subset of Objects365 images, and CHAIR marks object mentions hallucinated when they do not match any ground-truth object.
5 Experiments
Experiments evaluate UniProbe for token-level, object-level, and streaming hallucination detection across diverse LVLM backbones and benchmarks. UniProbe consistently improves detection and mitigation while avoiding full-model fine-tuning and adding modest latency.
- Token-level detection: 63.2/52.9 F1M/F1IoU on GLM-4V and 61.7/51.2 on Qwen-3-VL surpass HaloDet’s 59.1/49.8 and 55.4/46.7.UniProbe updates no backbone weights, whereas HaloDet fine-tunes every backbone weight.
- Token-level detection: 1.15× latency for full-response scoring is below HaloDet’s 1.30×, HalLocalizer’s 1.21×, and Whitehead et al.’s 1.23× relative to vanilla generation.This yields the strongest accuracy–efficiency trade-off among the trained baselines.
- Object hallucination detection: 92.3 F1 on self-generated COCO captions and 63.1 F1 / 90.0 AUC on POPE exceed the strongest prior detector’s 82.0 and 41.0 / 75.0.Prior detectors plateau at 69–75 AUC on POPE, while UniProbe structures frozen internal evidence as a computational trace.
- Streaming detection and hallucination-aware decoding: 63.8 F1obj after self-adaptation, versus 32.6 before adaptation, is achieved while retaining 63.5/53.0 F1M/F1IoU with IF = 100.The streaming detector is evaluated on GLM-4.1V’s freely generated captions with CHAIR-based object-hallucination labels.
- Streaming detection and hallucination-aware decoding: 55% reduction in CHAIRi/CHAIRs, from 18.0/37.2 to 8.2/16.6, occurs at 1.06× vanilla latency with self-adaptation.The same reject-and-resample guardrail is applied to every detector, isolating detection quality.
6 Ablation Study
Ablations show that UniProbe’s gains depend on preserving heterogeneous trace structure, reading from a mid-network layer, and selecting a decoding threshold that balances hallucination reduction with content preservation.
- Architecture and components: Flat MLP and Transformer probes trail UniProbe by at least 30.0 F1M, while the best individual module, GNN, reaches 56.9 F1M versus 63.2 for the full model.Removing the relational graph causes the largest drop (−10.1), followed by removing response recurrence (−4.6).
- Read-out layer: UniProbe’s detection peaks at layer 20 on GLM-4.1V, while LLaVA-1.5 and InternVL2 peak at layer 14 of 32.The peak layer is selected per backbone using validation data.
- Decoding threshold: At τ = 0.40, CHAIRi decreases by 36% but 12% of correct mentions are removed, illustrating the threshold’s trade-off.The default τ = 0.70 is the most aggressive setting reported to reduce hallucinations without removing correctly generated content.
7 Conclusion
UniProbe is a lightweight learnable detector that models heterogeneous internal representations from a frozen LVLM. Across backbones and benchmarks, it improves hallucination detection, while its streaming variant reduces object hallucinations during decoding with little latency and preserved response quality.
- UniProbe models a frozen LVLM’s internal attention, visual, and response representations with a lightweight learnable detector.
- Across multiple backbones and benchmarks, UniProbe improves token-level and object-level hallucination detection while keeping the underlying model frozen.
- The streaming variant reduces object hallucinations during decoding while preserving response quality and adding little latency.
- Structured internal model traces provide a practical signal for hallucination detection and mitigation.
Limitations · Appendix
UniProbe remains limited by its dependence on internal model signals and backbone-specific training. Cross-backbone transfer is left for future work.
- Limitations: UniProbe requires access to hidden states and attention maps.This restricts applicability to open-source models or closed-source providers that directly expose these signals.
- Limitations: The detector applies only to open-source models or closed-source models whose providers expose the required internal signals.
- Limitations: A separate readout must be trained for each backbone because internal representations differ across models.
- Appendix: Cross-backbone transfer remains an open direction for future work.
A Implementation Details
UniProbe extracts internal signals from one selected layer of each frozen LVLM backbone and uses a compact, shared architecture across backbones. Its implementation uses two alternating structure-aware blocks with approximately 16M trainable parameters.
- Backbone configuration: Layer 20 is selected for GLM-4.1V and InstructBLIP-7B, while layer 14 is selected for LLaVA-1.5 and InternVL2.The layer is chosen per backbone using validation F1 over a coarse layer sweep.
- Detector architecture: UniProbe uses hidden width h = 256 and L = 2 alternating blocks across all backbones.The configuration contains approximately 16M trainable parameters.
- Detector architecture: Each block combines one message-passing GNN layer, a single-layer bidirectional GRU over response tokens, and one ViT encoder layer over image nodes.The ViT uses 4 attention heads.
B User Study
The study evaluates whether mitigated captions are preferred by people, complementing automatic object-hallucination rates. In 2,000 blinded judgments, UniProbe’s captions were preferred most often, matching the automatic CHAIR ranking.
- Motivation: The study complements automatic object-hallucination rates with human preferences for captions mitigated by UniProbe, PAS, and HaloDet.Automatic rates assess removal of ungrounded objects but not whether captions read as better image descriptions.
- Protocol: The protocol comprised 100 randomized image-caption comparisons judged by 100 Mechanical Turk raters, totaling 2,000 judgments.Each comparison displayed the three unlabeled mitigated captions in random order.
- Instructions: Raters selected exactly one caption based on factual accuracy and completeness, without favoring captions solely for length or fluency.Participants were instructed to choose the caption that best described the visible image content.
- Results: 55% of judgments preferred UniProbe’s mitigated caption, versus 30% for PAS and 15% for HaloDet.Human preference mirrored the automatic CHAIR ranking, with UniProbe’s captions judged most faithful by a wide margin.
C Additional Qualitative Results … Streaming Read-out
Additional examples show UniProbe grounding hallucinated object, attribute, and object-presence errors across three LVLM backbones. Ablations further examine trace budgets, head averaging, and streaming read-out, with streaming matching bidirectional detection on model-generated responses.
- C Additional Qualitative Results: UniProbe recovers objects that the backbone verbally denies but whose presence is encoded in the model’s internal trace.The qualitative presence cases involve POPE-style questions and direct recovery from the backbone’s internal state.
- D Additional Ablation Results: Additional ablation studies complement the results reported in the main paper.The supplementary ablations cover trace budgets, head aggregation, and streaming read-out.
- Budget: UniProbe bounds its trace to at most 200 image nodes and 64 query patches, retaining patches ranked by response attention.It also bounds the number of edges drawn by each response token from every modality.
- Budget: F1M remains essentially flat at 62.1–63.0 when image-node caps vary from 100 to 600 and per-token image edges from 8 to 64.The sweep reports no trend, indicating that a small set of highly attended patches carries the relevant visual evidence.
- Head Averaging Ablation: Mean attention over heads is the most accurate and lightest representation, while retaining every head adds L×H features per edge and a gating layer yet slightly lowers detection.The comparison uses an identical evaluation protocol and considers head readings from a single mid-network layer or across all layers.
- Streaming Read-out: 63.2→61.2 F1M is the teacher-forced MHALO cost of a unidirectional response GRU, while streaming and bidirectional detectors tie at CHAIR detection F1 32.6 vs. 32.2 on model generations.F1IoU is tied in teacher-forced MHALO, and the unidirectional design is intended for token-by-token deployment.