Research map / Continual learning / Class-incremental

Hierarchical Semantic Tree Anchoring for CLIP-Based Class-Incremental Learning

NeurIPS 2026 · Continual learning › Class-incremental · approach: Exemplar-efficient class-incremental learning · 0 citations (OpenAlex)

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.

Proposes HASTEN, a CLIP-based class-incremental learner that uses a GPT-5-generated semantic tree to embed image and text features in hyperbolic space with entailment cones, and null-space gradient projection on a shared hyperbolic mapper to reduce forgetting.

Problem

Class-incremental learning with CLIP suffers from feature drift and catastrophic forgetting because Euclidean space fails to capture the hierarchical structure of real-world classes.

Approach

Supervise CLIP adapters with an LLM-generated is-a hierarchy embedded in Lorentz hyperbolic space via entailment-cone and margin losses, and restrict updates of the shared hyperbolic mapper to the approximate null space of prior-task features.

Key results

• HASTEN achieves the best last-stage accuracy on 8 of 9 benchmarks in the B0 and half-base settings, e.g., 89.03 vs 86.51 (PROOF) on Cars B0 Inc10 last-stage accuracy.
• On CIFAR100 B0 Inc10, MG-CLIP is best (82.78 last-stage vs HASTEN 82.45), the only exception reported.
• On UCF B0 Inc10, HASTEN reaches 92.61 last-stage accuracy vs 91.10 for PROOF.

Surprising findings

• MG-CLIP collapses on several settings with many incremental stages (e.g., ObjectNet B100 Inc20 and SUN B150 Inc30 last-stage accuracies around 12-35%).

Limitations

• Relies on a GPT-generated semantic tree, which may be noisy, incomplete, or biased when LLM domain coverage is limited.
• Does not beat MG-CLIP on CIFAR100 B0 Inc10 in last-stage accuracy.

Experimental setup

Datasets: CIFAR100 CUB200 ObjectNet ImageNet-R FGVCAircraft StanfordCars Food101 SUN397 UCF101

Benchmarks: CIFAR100 CUB200 ObjectNet ImageNet-R FGVCAircraft StanfordCars Food101 SUN397 UCF101

Metrics: average incremental accuracy last-stage accuracy

Baselines: Finetune SimpleCIL ZS-CLIP L2P DualPrompt CODA-Prompt PROOF RAPF MG-CLIP

Builds on: CLIP SimpleCIL PROOF RAPF MG-CLIP

Closest papers by meaning

• Anchor-based Robust Finetuning of Vision-Language Models (CVPR 2024 · similarity 0.87)
• Enhancing Visual Continual Learning with Language-Guided Supervision (CVPR 2024 · similarity 0.87)
• SmartCLIP: Modular Vision-language Alignment with Identification Guarantees (CVPR 2025 · similarity 0.86)
• Understanding Fine-tuning CLIP for Open-vocabulary Semantic Segmentation in Hyperbolic Space (CVPR 2025 · similarity 0.86)
• Tree of Attributes Prompt Learning for Vision-Language Models (ICLR 2025 · similarity 0.86)
• SuperCLIP: CLIP with Simple Classification Supervision (NeurIPS 2025 · similarity 0.86)