Research map / Vision-language and multimodal / Human and embodied modeling
Automated Detection of Visual Attribute Reliance with a Self-Reflective Agent
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces SAIA, a self-reflective agent that iteratively hypothesizes, tests, and self-evaluates visual attribute reliances in vision models, evaluated on a 130-model benchmark and on CLIP and YOLOv8.
Problem
Detecting which visual attributes unintentionally drive a vision model's predictions (spurious or non-robust reliance) without predefined concept sets or ground-truth labels.
Approach
An LLM agent runs a scientific hypothesis-testing loop, then self-evaluates its conclusion by generating high- and low-expected images with a text-to-image model and checking whether model scores separate; if not, it reflects and starts a new round.
Key results
- Predictiveness scores rise steadily over self-reflection rounds and exceed MAIA-style, MILAN, and multi-agent non-reflective baselines for both alpha=0.9 and alpha=0.5.
- For compositional AND-logic models, SAIA identifies both attributes in 80% of cases versus 0% for MAIA; for OR-logic models, SAIA recovers both in 20% of cases.
- Agent identifies previously undocumented dependencies in CLIP ViT (scientist recognized via laboratory settings) and YOLOv8 (biker detection favoring 45-degree and rear views).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add self-reflection rounds to reliance detection | SAIA on the controlled attribute-reliance benchmark · Controlled benchmark, strong and weak reliance settings | predictiveness score | Helped |
| use self-reflection instead of one-shot Milan-style detection | SAIA on the controlled attribute-reliance benchmark · Controlled benchmark | predictiveness score | Helped |
| use self-reflection instead of non-reflective hypothesis testing | SAIA on the controlled attribute-reliance benchmark · Controlled benchmark | predictiveness score | Helped |
| use self-reflection instead of repeated sampling with a non-reflective multi-agent ensemble | SAIA on the controlled attribute-reliance benchmark · Controlled benchmark | predictiveness score | Helped |
| increase reliance strength from α=0.5 to α=0.9 | SAIA · Controlled benchmark | predictiveness score | Helped |
| add self-reflection for counterfactual demographic reliance detection | SAIA · Counterfactual demographic reliances | predictiveness score | Helped |
| add self-reflection for AND-logic compositional reliance detection | SAIA · Synthetic compositional models, AND logic | both-dependency recovery rate: 80% | Helped |
| add self-reflection for OR-logic compositional reliance detection | SAIA · Synthetic compositional models, OR logic | both-dependency recovery rate: 20% | Helped |
| identify both dependencies in AND-logic cases rather than use non-reflective hypothesis testing | SAIA versus MAIA · Synthetic compositional models, AND logic | both-dependency recovery rate | Helped |
| identify both dependencies in OR-logic cases rather than use non-reflective hypothesis testing | SAIA versus MAIA · Synthetic compositional models, OR logic | both-dependency recovery rate | Helped |
Surprising findings
- SAIA recovers both attributes in only 20% of OR-logic compositional models, and MAIA recovers none, showing the method struggles when only one attribute must be found.
Limitations
- Performance depends on quality of generated images, limiting robustness for rare or out-of-distribution concepts.
- Backbone LLM priors and biases may constrain hypothesis space and induce confirmation bias, reducing performance on counterfactual demographic models.
- Struggles to recover multiple attributes in OR-logic settings because the search terminates after finding one attribute.
Experimental setup
Datasets: ImageNet
Benchmarks: Controlled benchmark of 130 models (18 reliance categories) CLIP ViT-L/14 and YOLOv8-m reliance discovery
Metrics: predictiveness score 2AFC preference rate
Baselines: MAIA-style agent MILAN Multi-agent ensemble Ground-truth descriptions
Builds on: MAIA MILAN Reflexion SELF-Refine
Closest papers by meaning
- Data-efficient Large Vision Models through Sequential Autoregression (ICML 2024 · similarity 0.87)
- Benchmarking the Attribution Quality of Vision Models (NeurIPS 2024 · similarity 0.86)
- UniBench: Visual Reasoning Requires Rethinking Vision-Language Beyond Scaling (NeurIPS 2024 · similarity 0.86)
- What's in the Image? A Deep-Dive into the Vision of Vision Language Models (CVPR 2025 · similarity 0.85)
- Language Model as Visual Explainer (NeurIPS 2024 · similarity 0.85)
- Black Sheep in the Herd: Playing with Spuriously Correlated Attributes for Vision-Language Recognition (ICLR 2025 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.