Visual Hallucination and Grounding

When the model sees what is not there

A VLM hallucinates when it states something the image does not support: an object that is absent, a wrong count, a misread number, or a confident description of a blurry region. One cause is the language model's strong prior: a kitchen photo "usually" has a fridge, so the model may mention one. Another is resolution: small text is downsampled away before the model ever sees it, yet the model still produces a plausible-looking value. Counting, spatial relations (left of, behind) and reading charts are known weak spots. Mitigations work in layers. Send higher resolution or crops for detail questions. Ask for grounded output, such as bounding boxes or quoted text, which some models support and which you can check. Allow and encourage "not visible" as an answer. Verify with deterministic tools where possible: OCR for text, a detector for counts, arithmetic for totals.

Defence in depth for multimodal apps

Safe multimodal systems layer controls: clean inputs, isolate untrusted content, verify outputs and keep a human in the loop for high-stakes cases.

Give the model a way out

Prompts like "If the value is not clearly readable, answer null" usually reduce invented values in extraction tasks. Without an allowed escape, the model tends to fill the field anyway.

त्वरित जाँच: A VLM reports the wrong amount from a small, blurry total on a receipt. What is the best first fix?

• Raise the temperature
• Ask the model to be more confident
• Send a higher-resolution crop of the total region and allow a 'not readable' answer
• Switch to a text-only LLM