ROT: Rotating Hidden States towards Contextual Vectors for Hallucination Mitigation in LVLMs

AI summary

Overview

Research area: Multimodal machine learning / Large Vision-Language Models (LVLMs), specifically training-free inference-time interventions for object hallucination mitigation.

Technical level: Advanced. The paper assumes familiarity with Transformer internals, self-attention, residual streams, hidden-state geometry, and cosine similarity in high-dimensional vector spaces.

Scope: The paper proposes ROT, a layer-specific, training-free framework that detects semantically deviating hidden states in the middle layers of an LVLM and applies a norm-preserving rotation to steer them back toward the plane spanned by textual and visual context vectors, followed by representational smoothing in later layers.

What This Paper Is About

Large Vision-Language Models often describe objects that are not actually in the image, a failure known as object hallucination. Existing training-free fixes mostly manipulate attention weights or filter visual tokens, which the authors argue only indirectly controls the semantics that reach the final predictive layers. This paper instead intervenes directly on the hidden state vectors after self-attention and residual addition, showing that hallucinated tokens are geometrically "off-plane" relative to both textual and visual contexts, and correcting that deviation by rotation.

Key Contributions

1. A new intervention target. The paper shifts hallucination mitigation from attention weights to the hidden state vectors extracted after self-attention and residual addition (before layer normalization), addressing semantic deviation directly in hidden space.

2. A hidden-state-based detection mechanism. It defines a contextual deviation indicator E = 2 - S_T - S_V (where S_T and S_V are cosine similarities of the hidden state with the textual context C_T and visual context C_V), observes that E peaks in the middle layers, and uses a similarity-based threshold to trigger intervention there.

3. A layer-specific mitigation strategy. Deviating hidden states in the middle layers undergo a norm-preserving rotation toward the plane spanned by C_T and C_V via spherical linear interpolation, while subsequent layers apply a weighted-interpolation smoothing mechanism to keep the corrected trajectory stable.

4. Broad empirical validation. The method is tested on LLaVA-1.5-7B, LLaVA-1.5-13B, Qwen2-VL-7B, and Qwen2.5-VL-32B across CHAIR, POPE, and MME, with ablations isolating the rotation and smoothing components.

Main Findings

• Hallucinated tokens deviate from both contexts. Under the common language-prior hypothesis, hallucinated tokens should align more strongly with textual context. Instead, averaged over 3,000 CHAIR evaluation instances decoded with LLaVA-1.5-7B, hallucinated tokens show consistently lower cosine similarity to both the textual context C_T and the visual context C_V than normal tokens.

• Deviation concentrates in middle layers. The indicator E stays relatively low in early and final layers and peaks in intermediate layers, which the authors describe as a critical multimodal fusion zone. For LLaVA-1.5-7B, layers 6 through 26 satisfied the sensitivity criterion and became the active intervention zone.

• CHAIR results on LLaVA-1.5-7B. ROT reduces the sentence-level hallucination metric C_S from 54.4 (greedy) to 33.0 and the instance-level metric C_I from 15.7 to 6.7. Compared with the second-best method reported, LVLMs-Saliency (35.6 / 8.2), ROT improves C_S by 2.6 points and C_I by 1.5 points.

• POPE results on LLaVA-1.5-7B. Accuracy rises from 85.5 (greedy) to 86.6 and F1 from 85.9 to 86.3, outperforming VCD, OPERA, TAME, and the other listed baselines.

• MME results. ROT achieves the highest reported total score of 506.66, compared with 440.67 for greedy decoding and 498.33 for LVLMs-Saliency.

• Generalization across architectures and scales. Across LLaVA-1.5-7B, LLaVA-1.5-13B, Qwen2-VL-7B, and Qwen2.5-VL-32B, ROT yields an average relative reduction of 26.4% in C_S and 42.7% in C_I, and Accuracy lifts of up to 2.0 absolute points on POPE (including its adversarial, popular, and random subsets). Individual CHAIR results: LLaVA-1.5-13B goes from 41.4 / 12.7 to 30.8 / 4.3; Qwen2-VL-7B from 24.2 / 8.0 to 21.4 / 4.9; Qwen2.5-VL-32B from 43.6 / 9.5 to 30.9 / 8.7.

• Both components are necessary (ablation). Removing rotation raises LLaVA-1.5-7B's C_I from 6.7 to 12.4 and C_S from 33.0 to 40.5. Removing smoothing raises C_I from 6.7 to 8.3 and C_S from 38.2. On Qwen2-VL-7B, removing rotation gives 23.4 / 6.6, removing smoothing gives 21.9 / 5.2, versus 21.4 / 4.9 for full ROT.

• Case study. For an image of a person skateboarding, the baseline hallucinated bench (S_T = 0.41, S_V = 0.24) and clock (S_T = 0.36, S_V = 0.20), whereas ROT generated the grounded token fence with higher similarities (S_T = 0.54, S_V = 0.36).

Methodology in Plain English

The authors first ran an analysis pass. At each layer, they took the attention weights the current token places on all previous tokens, split them into text-token weights and vision-token weights, normalized each group separately, and used them to compute two weighted averages of the previous layer's hidden states: a textual context vector C_T and a visual context vector C_V. They then measured the cosine similarity of the current token's post-residual hidden state h_k with each of these. Comparing normal versus hallucinated tokens showed that hallucinated ones are farther from both, so the authors built a deviation score E = 2 - S_T - S_V.

If E crosses a threshold τ in a middle layer, the method builds a target direction C_target = Norm((1 - α)·C_T + α·C_V), blending the two contexts with a balance factor α. The hidden state is then rotated toward that target using spherical linear interpolation with intensity β, and rescaled so its original norm is preserved: h_new = ||h_k|| · Norm(Slerp(h_k, C_target; β)). Keeping the norm constant means the direction is corrected without shrinking the signal before LayerNorm and the feed-forward network.

For all layers after the intervention zone, the authors stop rotating and instead smooth, updating each hidden state as h_k = γ·h_{k-1} + (1 - γ)·h_k with a smoothing factor γ, so the corrections propagate to the prediction head without destabilizing the vocabulary distribution. Three hyperparameters control everything (α, β, γ), and they were set per model: LLaVA-1.5-7B used α = 0.44, β = 0.062, γ = 0.51; LLaVA-1.5-13B used 0.45, 0.050, 0.55; Qwen2-VL-7B used 0.50, 0.048, 0.43; Qwen2.5-VL-32B used 0.46, 0.024, 0.41. All experiments ran on a single NVIDIA A800 (80GB) GPU. Results are measured with CHAIR (sentence-level C_S, instance-level C_I, lower is better), POPE (Accuracy and F1 on adversarial, popular, and random subsets), and MME (per-subtask scores with a maximum of 200, combining basic and pair-wise accuracy).

Why This Matters

Impact on research. The paper challenges the widespread "language prior" explanation of object hallucination by showing a counter-intuitive geometric signature: hallucinated tokens are less similar to textual context, not more. It reframes mitigation as a direction-correction problem in hidden space rather than a weight-allocation problem, and it does so without any training, which matters given the cost of retraining or RLHF-aligning large LVLMs.

Real-world applications:

• Assistive technology for blind and low-vision users, where hallucinated objects in captions can cause real harm.
• Medical or industrial image description and reporting systems, where ungrounded findings are unacceptable.
• Content moderation and media monitoring pipelines that rely on automatic captioning at scale.
• Robotics and autonomous systems that use vision-language descriptions as grounding signals for action.

Industry relevance. ROT requires no fine-tuning, no extra forward passes (unlike contrastive decoding, which the paper notes doubles inference latency), and three tunable scalars. That makes it attractive as a drop-in inference-time module for teams serving LVLMs who cannot afford retraining budgets.

Future Directions

• Richer grounding references. The grounded plane is built only from the local textual and visual contexts in the attention mechanism. Incorporating external structured knowledge or global scene graphs could help resolve harder hallucinations, as the authors note in the limitations.
• Automated configuration. The critical deviating layers and the trigger threshold τ are chosen from empirical observation of representative model families. Adaptive fine-tuning or automated search may be needed for ultra-large models with different depth-wise representation dynamics.
• Object-attribute and relational hallucinations. The paper's empirical analysis is framed around object hallucination; whether the same contextual-deviation signature holds for attributes or relationships is not reported.
• Comparison against logit-contrast approaches. The paper does not report latency or throughput measurements for ROT versus baselines, so its efficiency advantage relative to VCD or OPERA is asserted qualitatively rather than quantified.

Target Audience

This paper is best suited to researchers and engineers already working on LVLM inference, mechanistic interpretability of Transformer hidden states, or hallucination mitigation who want a training-free, geometry-based alternative to attention manipulation. It is also useful for practitioners deploying captioning or visual question answering systems who need a lightweight grounding intervention, and for graduate students studying the internal representation dynamics of multimodal models.

Note: the paper does not report the numeric value of the sensitivity threshold used to select layers 6–26, the numeric value of the trigger threshold τ, or any latency/FLOP overhead measurements.

Authors’ abstract

Large Vision-Language Models (LVLMs) frequently suffer from object hallucination. Existing training-free interventions primarily manipulate attention weights, which indirectly affect the deep semantics reaching the final predictive layers. In this work, we shift our focus to the hidden state vectors extracted after self-attention and residual addition. Empirical analysis reveals that hallucinated tokens do not simply over-rely on linguistic priors; instead, they exhibit an anomalous contextual deviation, showing significantly lower similarities to both textual and visual contexts in intermediate layers. Motivated by this, we propose ROT, a layer-specific, training-free framework. ROT dynamically detects semantic deviation in the middle layers and applies a norm-preserving rotation to steer the hidden states back toward the local multimodal context plane spanned by the contexts. For subsequent layers, a representational smoothing mechanism is introduced to stabilize the calibrated trajectory. Extensive experiments on multiple benchmarks demonstrate that ROT consistently reduces hallucinations across various model architectures and scales, offering an efficient, geometry-driven solution for grounded generation.