BLens: Contrastive Captioning of Binary Functions using Ensemble Embedding
summary
BLens builds an ensemble encoder that fuses PalmTree, Dexter and CLAP representations into patch embeddings, aligns them to tokenized names via contrastive alignment, and uses a specialized LORD decoder with MLM-style masking and flexible autoregression. The project is open sourced with artifact badges. Empirically it yields F1 0.79 vs 0.70 (cross-binary), 0.46 vs 0.29 (cross-project), and 0.32 vs 0.19 (reduced shared components), achieving high precision (reported 92%) and substantially outperforming prior methods.
AI-written summary · about
chapters (11)
- 0:00BLens introduction
- 1:55Function name prediction challenges
- 3:20Baseline limitations
- 4:24Prediction as captioning
- 5:49Pre-training with ensemble embedding
- 8:00Fine-tuning with LORD decoder
- 11:14Evaluation settings
- 13:25Performance vs baselines
- 14:07Ablation study
- 14:49Case studies
- 15:54Summary
related talks
26mMaximizing Multimodal: Exploring the search frontier of text-to-image models
16mAre Neuro-Inspired Multi-Modal Vision-Language Models Resilient to Membership...
13mInside Multimodal Models
32mEverything you wanted to know about VLMs but were afraid to ask
37mFlexibly Fair Representation Learning by Disentanglement
52mLanguage-guided Adaptation
01h 04mFrom Pixels to Words: Exploring Multi-Modal AI with Python
14mABC: Towards a Universal Code Styler through Model Merging