Research map / Continual learning / Continual fine-tuning of LLMs
SiGMA: Sign-Guided Merging and Adaptation framework for Multimodal Continual Instruction Tuning
Keonhee Park, Gunhee Kim
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add sign-guided adaptive tuning (S-Tuning) to the expansion-merge baseline (ablation) | expansion-merge LoRA baseline · UCIT | Avg. All: 62.42 -> 64.86 | Improved |
| add sign-guided merging (S-Merge) to the baseline (ablation) | expansion-merge LoRA baseline · UCIT | Avg. All / Forgets: 62.42 -> 63.35 Avg. All; -5.98 -> -6.26 Forgets | Mixed |
| add S-Merge on top of S-Tuning (ablation) | SiGMA · UCIT | Avg. All: 64.86 -> 66.17 | Improved |
| add sign-based weight decoupling (SWD) when consolidating A and B (ablation) | S-Tuning · UCIT | Avg. All: 62.42 -> 64.86 | Improved |
| add SWD when consolidating full LoRA deltas (ablation) | S-Tuning · UCIT | Avg. All: 63.23 -> 64.15 | Improved |
| consolidate low-rank A and B separately instead of full deltas (with SWD) (ablation) | S-Tuning · UCIT | Avg. All: 64.15 -> 64.86 | Improved |
| exclude the top layer from S-Tuning (k=1) instead of tuning all layers (ablation) | S-Tuning · UCIT | Avg. All: 64.27 -> 66.17 | Improved |
| apply S-Tuning to more layers (k from 32 to 1) (ablation) | S-Tuning · UCIT | Avg. All: 63.35 -> 66.17 | Improved |
| add cosine-distance amplification d_k in S-Merge (ablation) | S-Merge · UCIT | Avg. All: 62.69 -> 66.17 | Improved |
| add magnitude threshold lambda=0.8 to S-Merge (ablation) | S-Merge · UCIT | Avg. All / Forgets: 54.32 -> 66.17 Avg. All; -19.65 -> -4.81 Forgets | Improved |
| raise lambda from 0.8 to 1.0 (ablation) | S-Merge · UCIT | Avg. All: 66.17 -> 65.89 | No clear effect |
| SiGMA versus strongest baseline DISCO | SiGMA · UCIT | Avg. All: 62.91 -> 66.17 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- CL-MoE — CL-MoE: Enhancing Multimodal Large Language Model with Dual Momentum Mixture-of-Experts for Continual Visual Question Answering (CVPR 2025) · better on 2 of 2 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- CoIN: A Benchmark of Continual Instruction Tuning for Multimodel Large Language Models (NeurIPS 2024 · similarity 0.89)
- SEFE: Superficial and Essential Forgetting Eliminator for Multimodal Continual Instruction Tuning (ICML 2025 · similarity 0.89)
- Dynamic Mixture of Curriculum LoRA Experts for Continual Multimodal Instruction Tuning (ICML 2025 · similarity 0.88)
- Large Continual Instruction Assistant (ICML 2025 · similarity 0.88)
- Overcoming Catastrophic Forgetting in Visual Continual Learning with Reinforcement Fine-Tuning (NeurIPS 2026 · similarity 0.87)
- Adapt-$\infty$: Scalable Continual Multimodal Instruction Tuning via Dynamic Data Selection (ICLR 2025 · similarity 0.87)
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.
