Neura Market
Look Twice: Training-Free Evidence Highlighting in MLLMs (April 2026) logo

Look Twice: Training-Free Evidence Highlighting in MLLMs (April 2026)

Free

Inference-time framework exploiting MLLM attention patterns to identify relevant visual regions and text, then re-conditions generation on highlighted evidence — consistent VQA improvements, no training required

FreeFree tier
Inputs: image
Type
Open Source
Visit Website GitHub (40 stars)

About Look Twice: Training-Free Evidence Highlighting in MLLMs (April 2026)

Look Twice (LoT) is a training-free, inference-time framework designed to improve how pretrained Multimodal Large Language Models (MLLMs) utilize multimodal evidence when answering knowledge-intensive questions. By exploiting the model's own attention patterns, LoT identifies which visual regions and retrieved textual elements are most relevant to a query, then re-conditions the generation on highlighted evidence using lightweight prompt-level markers. This approach yields consistent improvements across multiple knowledge-based VQA benchmarks, as well as on vision-centric and hallucination-oriented benchmarks, without any additional training or architectural modifications. The source code is planned for public release.

Key Features

Training-free inference-time framework for MLLMs
Exploits model attention patterns to estimate relevant visual regions and retrieved textual elements
Generates answers conditioned on highlighted evidence via lightweight prompt-level markers
Works with or without external textual context (e.g., retrieved passages)
No additional training or architectural modifications required
Consistent improvements on knowledge-based VQA, vision-centric, and hallucination-oriented benchmarks

Pros & Cons

Pros
  • Requires no training or fine-tuning of existing models
  • Improves accuracy on multiple VQA benchmarks over zero-shot baselines
  • Applicable to both multimodal (text+image) and pure visual QA settings
  • Lightweight and easy to integrate into any pretrained MLLM
Cons
  • May introduce slight inference-time overhead due to attention analysis and re-conditioning
  • Effectiveness depends on the quality of the underlying MLLM's attention patterns

Best For

Knowledge-based visual question answering (VQA)Vision-centric tasks without textual contextReducing hallucination in MLLM responses

FAQ

What is Look Twice?
Look Twice is a training-free inference-time framework that improves how Multimodal Large Language Models (MLLMs) use visual and textual evidence by highlighting relevant regions based on attention patterns, leading to more accurate answers.
Does Look Twice require additional training?
No, Look Twice is entirely training-free and does not modify the underlying model architecture. It works by analyzing attention patterns at inference time.
Is the source code available?
According to the paper, the source code will be publicly released. As of the publication date, it is planned but not yet available.