Paper Study Graph
Each point is a paper I have read, placed near papers on close topics. Hover over a point to see the paper, and click it to open the paper.
Unlearning & Continual Learning
Continual Unlearning
- Which Concepts to Forget and How to Refuse? Decomposing Concepts for Continual Unlearning in Large Vision-Language Models
CVPR 2026
Breaks forget targets into visual and textual concepts, then routes them to refusal experts.
- The Forgetting-Retention Dilemma: Certified Unlearning Theory in Continual Learning
ICML 2026
Frames unlearning during continual learning as one excess-risk objective with certified guarantees.
- Adaptive Localization of Knowledge Negation for Continual LLM Unlearning
ICML 2025 · Poster
Masks task-vector gradients and adapts unlearning strength to each new request.
- On Large Language Model Continual Unlearning
ICLR 2025
Keeps an orthogonal LoRA per unlearning request and loads it with an OOD detector at inference.
- Locality-Aware Continual Unlearning for Diffusion Models
arXiv 2025
Removes concepts from text-to-image diffusion one after another while protecting nearby concepts.
Machine Unlearning
- Distillation Robustifies Unlearning
NeurIPS 2025 · Spotlight
Shows unlearned knowledge comes back easily, and distills into a noised copy to make forgetting stick.
- Balancing Forget Quality and Model Utility: A Reverse KL-Divergence Knowledge Distillation Approach for Better Unlearning in LLMs
NAACL 2025
Uses reverse-KL distillation to remove only the forget-target part of the token distribution.
- UnDIAL: Self-Distillation with Adjusted Logits for Robust Unlearning in Large Language Models
NAACL 2025
Self-distills with adjusted logits to lower target tokens without unstable gradient ascent.
Continual Learning
- Orthogonal Subspace Learning for Language Model Continual Learning
EMNLP 2023
Learns each new task in a LoRA subspace orthogonal to earlier tasks.
- InfLoRA: Interference-Free Low-Rank Adaptation for Continual Learning
CVPR 2024
Designs a LoRA subspace so new tasks do not interfere with old ones.
- Select and Distill: Selective Dual-Teacher Knowledge Transfer for Continual Learning on Vision-Language Models
ECCV 2024
Picks between the original and the latest model as teacher to keep vision-language models from forgetting.
Efficient Adaptation
PEFT
- DoRA: Weight-Decomposed Low-Rank Adaptation
ICML 2024 · Oral
Splits weights into magnitude and direction, and adapts the direction with LoRA.
- DiaBlo: Diagonal Blocks Are Sufficient For Finetuning
ICLR 2026
Fine-tunes only the diagonal blocks of weight matrices instead of low-rank adapters.
Reliability
Confidence Calibration
- A Stitch in Time Saves Nine: A Train-Time Regularizing Loss for Improved Neural Network Calibration
CVPR 2022
Adds a train-time loss that aligns confidence with accuracy across all classes.
- On Calibration of Modern Neural Networks
ICML 2017
Shows modern networks are overconfident and fixes it with temperature scaling.
- Measuring Calibration in Deep Learning
CVPR Workshops 2019
Examines the weaknesses of ECE and proposes better calibration metrics.
Foundation Models
CV
- Deep Residual Learning for Image Recognition
CVPR 2016 · Best Paper
Residual connections let very deep networks train without degrading.
- Rich Feature Hierarchies for Accurate Object Detection and Semantic Segmentation
CVPR 2014
Classifies region proposals with CNN features for object detection.
- You Only Look Once: Unified, Real-Time Object Detection
CVPR 2016
Treats detection as a single regression problem for real-time object detection.
- An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale
ICLR 2021 · Oral
Splits images into patches and classifies them with a plain Transformer.
- Generative Adversarial Networks
NeurIPS 2014
Trains a generator against a discriminator to learn the data distribution.
- NeRF: Representing Scenes as Neural Radiance Fields for View Synthesis
ECCV 2020 · Oral, Best Paper HM
Represents a scene as a neural radiance field to render new viewpoints.
NLP
- Attention Is All You Need
NeurIPS 2017
Replaces recurrence with self-attention for sequence modeling.
- BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding
NAACL 2019
Pre-trains a bidirectional Transformer encoder with masked language modeling.