Large language models (LLMs) perform inference through hidden representations that evolve across layers and generated tokens, forming complex trajectories in hidden representation space. We study the geometric dynamics of these trajectories for label-free model self-evaluation problem: does the hidden-state tensor already contain model response correctness signals, and where do those signals reside? Our analysis is organized along three intrinsic dimensions of inference-time computation: layer depth, token span, and representation space. Across seven benchmarks spanning mathematical reasoning, commonsense reasoning, and multilingual understanding, and two primary instruction-tuned LLMs, we find that correctness separability is highly localized, consistently concentrating in late-layer representations and limited token regions. Motivated by this, we propose Norm Drop Rate (NDR), a minimal diagnostic metric based on final-layer norm contraction. NDR achieves competitive AUROC across tested models and the best AUPR on Qwen2-7B-Instruct, while remaining the simplest in design and most robust when controlled for response length or changing representation bases. These results suggest that hidden-state trajectory analysis should focus less on increasingly complex diagnoses but more on identifying simple, localized signals.

Hallucinations remain a persistent challenge for vision–language models (VLMs), which often describe nonexistent objects or fabricate facts. Existing detection methods typically operate after text generation, making intervention both costly and untimely. We investigate whether hallucination risk can instead be predicted before any token is generated by probing a model’s internal representations in a single forward pass. Across a diverse set of vision–language tasks and eight modern VLMs, including Llama-3.2-Vision, Gemma-3, Phi-4-VL, and Qwen2.5-VL, we examine three families of internal representations: (i) visual-only features without multimodal fusion, (ii) vision token representations within the text decoder, and (iii) query-token representations that integrate visual and textual information before generation. Probes trained on these representations achieve strong hallucination-detection performance without decoding, reaching up to 0.93 AUROC on Gemma-3-12B, Phi-4-VL 5.6B, and Molmo 7B. Late query-token states are the most predictive for most models, while visual or mid layer features dominate in a few architectures (e.g., ∼0.79 AUROC for Qwen2.5-VL-7B using visual-only features). These results demonstrate that (1) hallucination risk is detectable pre-generation, (2) the most informative layer and modality vary across architectures, and (3) lightweight probes has the potential to enable early abstention, selective routing, and adaptive decoding to improve both safety and efficiency.