Vision-Language Grounding and Spatial Hallucination Triage Checklist

Role: Lead Multimodal Evaluation Engineer specializing in vision-language alignment, visual grounding, and hallucination triage.

Context

• Multimodal Foundation Model: {{multimodal_model_version}}
• Visual Input Description: {{input_image_context}}
• System Prompt Architecture: {{system_prompt_scaffold}}
• Observed Grounding Failure: {{spatial_grounding_failure}}
• Interleaved Query Sequence: {{interleaved_prompt_sequence}}
• Sampling & Attention Hyperparameters: {{attention_temperature_setting}}

Task

Construct an end-to-end debugging and remediation checklist to diagnose spatial reasoning failures, eliminate hallucinations, and enforce strict visual grounding in the multimodal reasoning workflow.

Method

1. Analyze {{spatial_grounding_failure}} against {{input_image_context}} to classify the error type (e.g., existence hallucination, relational misbinding, attribute misattribution, or coordinate hallucination).
2. Review {{system_prompt_scaffold}} for ambiguity regarding visual verification protocols, chain-of-thought grounding mandates, or coordinate formatting.
3. Dissect {{interleaved_prompt_sequence}} to determine if prompt positioning causes recency bias or visual attention degradation over long context windows.
4. Assess the impact of {{attention_temperature_setting}} on token sampling divergence during visual grounding tasks.
5. Devise explicit visual anchoring prompt patterns (e.g., bounding-box pre-generation, visual-crop verification, or step-back visual summarization).
6. Formulate negative reasoning constraints to force the model to explicitly cite visual coordinates or state absence when confidence is low.
7. Compile a structured debugging checklist covering prompt scaffolding, spatial coordinate verification, and iterative visual prompting.
8. Establish deterministic evaluation metrics to benchmark hallucination rates across transformed and perturbed test images.

Constraints

• MUST format all outputs as a prioritized, operational checklist with clear completion criteria.
• MUST NOT require model fine-tuning; all solutions must rely on prompt engineering, structured reasoning, and inference parameter tuning.
• MUST include explicit chain-of-thought grounding patterns tailored to {{multimodal_model_version}}.
• Every checklist item MUST provide concrete criteria for verifying visual grounding against {{input_image_context}}.

Output format

• Hallucination Mechanism Diagnostic Summary (max 175 words)
• Checklist Tier 1: Visual Prompt Scaffolding & System Constraint Fixes (5-6 actionable items)
• Checklist Tier 2: Spatial Decomposition & Grounded Reasoning Protocols (5-7 actionable items)
• Checklist Tier 3: Inference Parameter & Context Order Tuning (4-5 actionable items)
• Quality Assurance & Perturbation Testing Suite (markdown checklist with 4 distinct verification scenarios)

Self-review

• Are the diagnostic steps specifically tailored to the grounding failure outlined in {{spatial_grounding_failure}}?
• Does the checklist include explicit spatial reasoning techniques (e.g., coordinates, region-of-interest anchoring)?
• Are all variable references syntactically matched and fully contextualized throughout the body?