LLM Reasoning Explorer Studio
Explore attribution results and feature interactions with our interactive visualization tools. Browse pre-computed examples or analyze your own text in real-time with powerful AI insights.
Build: 49b8995 (2026-02-27)
- Select a dataset from 10 available datasets (100 total examples, 10 per dataset)
- Choose a model to compare: Qwen3-4B, Qwen3-30B, or Mistral-7B
- Pick a scoring method: Perplexity or Semantic Similarity
- Set the feature level: Word, Sentence, or Paragraph
- Choose an attribution method: Shapley, Banzhaf, or Influence
- View results in the Text Interaction View (inline highlights) and Bar View (ranked interactions)
Interactions
Debug Scores
label | indices | score | total_logprob | avg_nll | ppl | target_len | error |
|---|---|---|---|---|---|---|---|
- Choose a dataset from the three sub-tabs:
- MIMIC-CXR (10 Samples) — chest X-rays across 10 pathology categories
- Dermoscopy ISIC (10 Samples) — skin-lesion dermoscopy across 8 diagnostic classes
- MS-COCO (5 Samples) — natural-image cross-modal benchmark
- Pick an example from the dropdown (each is an image + caption pair)
- Choose an attribution method: Influence (default, non-negative — clearer for clinicians) or Shapley (signed)
- Read the four panels side-by-side:
- Interactive Cross-Modal View — click any image patch or caption word to see its strongest cross-modal partners
- BiomedCLIP Cross-Modal Attribution — patch-level overlay + bar charts (cosine-similarity scoring)
- LLaVA-Med Attribution — log-prob and generation Shapley charts from the medical 7B VLM
- Compare Two Methods Side-by-Side — pick any two of the above to overlay their rankings
- Hover token chips and patches for exact attribution values; hover SVG arcs for pairwise interaction strength
10-sample MIMIC-CXR chest X-ray attribution benchmark (10 pathology categories).
Source: MIMIC-CXR-JPG — de-identified chest radiographs from Beth Israel Deaconess Medical Center.
Each example has a radiology report (impression = caption, findings = detail).
Select an example above to see the attribution analysis.
BiomedCLIP Cross-Modal Interaction View — click segments or words
How to use: Click any image region to see which caption words it connects to, or click a word to see which regions activate.
Green arrows = positive interaction. Red arrows = negative.
BiomedCLIP Cross-Modal Attribution
What it does: Uses BiomedCLIP — a CLIP model trained on 15 million biomedical figure-caption pairs — to jointly score image regions (via UnSAM segmentation) and caption tokens.
How to read: Green = positive Shapley value (contributes to alignment). Red = negative (hurts alignment).
LLaVA-Med Attribution (4×4 Patch Grid, P1–P16)
What it does: Uses LLaVA-Med — a 7B parameter medical VLM — evaluated over a uniform 4×4 patch grid (16 cells labeled P1–P16, row-major).
Two scoring approaches:
- Log-Prob: How removing a region affects confidence in the correct caption
- Generation: How removing a region changes what the model describes
Each method colors segments by its own Shapley values — green = positive, red = negative. Signs often differ between Log-Prob and Generation, so each has its own overlay.
Compare Two Methods Side-by-Side
Select two methods to compare their attribution overlays and Shapley value distributions on the same image.
10-sample ISIC-2019 dermoscopy attribution benchmark (8 diagnostic classes: MEL × 2, NV × 2, BCC, AK, BKL, DF, VASC, SCC).
Source: ISIC_2019_224 — dermoscopic skin-lesion images from the International Skin Imaging Collaboration.
Captions are synthesized from class labels (clinical descriptions of each diagnosis).
Select an example above to see the attribution analysis.
BiomedCLIP Cross-Modal Interaction View — click segments or words
How to use: Click any image region to see which caption words it connects to, or click a word to see which regions activate.
BiomedCLIP Cross-Modal Attribution
What it does: Uses BiomedCLIP to jointly score dermoscopic image regions (via UnSAM segmentation) and caption tokens.
How to read: Influence bars (default) show positive importance. Switch to Shapley above for signed values (green/red).
LLaVA-Med Attribution (4×4 Patch Grid, P1–P16)
What it does: Uses LLaVA-Med — a 7B parameter medical VLM — evaluated over a uniform 4×4 patch grid (16 cells labeled P1–P16, row-major).
Two scoring approaches:
- Log-Prob: How removing a region affects confidence in the caption
- Generation: How removing a region changes what the model describes
Compare Two Methods Side-by-Side
Select two methods to compare their attribution overlays and value distributions on the same image.
CLIP cross-modal attribution on MS-COCO natural images.
Click an image region or caption word below to explore which parts of the image and text are most strongly linked via CLIP's visual-language similarity score.
Interactive Cross-Modal View
Click a colored image region (left) to highlight the caption words it interacts with, or click a word (right) to highlight linked regions. Green = positive, red = negative.
Attribution Details
Cross-Modal Interaction Table
Image Region | Caption Word | Score |
|---|---|---|
Masked Image Browser
Browse ablation images: solo shows only the selected region (everything else inpainted away); removed shows the image with that region inpainted out.
Contributors — University of California, Berkeley
Stephen Tao · Loader Layer · stephen_tao@berkeley.eduYiting Gao · Attribution Layer · yg2025@berkeley.eduQingpeng Kong · Visualization Layer · qpkong@berkeley.edu
Advisor: Kannan Ramchandran · kannanr@berkeley.eduMentor: Landon Butler · landonb@berkeley.edu