Research map / Continual learning / Continual fine-tuning of LLMs
Updating Parametric Knowledge with Context Distillation Retains Post-Training Capabilities
Shankar Padmanabhan, Mustafa Omer Gul, Tanya Goyal
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 |
|---|---|---|---|
| use DiSC context distillation instead of standard fine-tuning (capability-preserving LR) | Qwen2.5-7B-Instruct · KUP | accuracy (%): 36.3 -> 44.18 | Improved |
| use DiSC instead of FT at unconstrained learning rate | Qwen2.5-7B-Instruct · KUP | MATH-Hard / IFEval: IFEval 65.83 -> 77.70; Math 24.47 -> 44.34 | Improved |
| use DiSC instead of CD base | post-trained LMs · KUP, Llama3.1-8B / Qwen2.5-7B / Qwen3-8B | accuracy (%): 35.80 -> 44.88 (Llama) | Improved |
| use DiSC instead of FT | post-trained LMs · BioASQ | accuracy (%): 73.08 -> 81.73 (Qwen2.5-7B) | Improved |
| add KL regularization to fine-tuning (ablation) | Qwen2.5-7B-Instruct · KUP | IFEval / MATH forgetting: IFEval 65.83 -> 69.78; Math 24.47 -> 27.57 | No clear effect |
| add TALR token-adaptive loss to fine-tuning (ablation) | Qwen2.5-7B-Instruct · KUP | MATH forgetting: 24.47 -> 26.51 | No clear effect |
| add LoRA to fine-tuning (ablation) | Qwen2.5-7B-Instruct · KUP | KUP accuracy and IFEval: KUP 40.9 -> 35.6; IFEval 65.83 -> 75.90 | Mixed |
| add on-policy rephrases to fine-tuning (ablation) | Qwen2.5-7B-Instruct and Llama3.1-8B · KUP | adaptation accuracy: 40.90 -> 42.90 (Qwen 7B) | Mixed |
| add replay data (GSM8K gold) to fine-tuning (ablation) | Qwen2.5-7B-Instruct · KUP | MATH: 24.47 -> 2.95 | Hurt |
| add replay data (AlpacaEval gold) to fine-tuning (ablation) | Qwen2.5-7B-Instruct · KUP | IFEval: 65.83 -> 48.44 | Hurt |
| use DiSC instead of FT as learning rate increases (ablation) | Qwen2.5-7B-Instruct · KUP | IFEval drop: -14.87 -> -8.4 | Improved |
| fine-tune on KUP (standard FT) | post-trained LMs · IFEval, Math, Code vs BBH, GPQA, MuSR | degradation | Hurt |
Closest papers by meaning
- Cooperative Knowledge Distillation: A Learner Agnostic Approach (AAAI 2024 · similarity 0.88)
- Distilling What Matters: Confidence-Aware Selective Distillation for Large Language Models (NeurIPS 2026 · similarity 0.88)
- TAID: Temporally Adaptive Interpolated Distillation for Efficient Knowledge Transfer in Language Models (ICLR 2025 · similarity 0.88)
- Mix-CPT: A Domain Adaptation Framework via Decoupling Knowledge Learning and Format Alignment (ICLR 2025 · similarity 0.88)
- Distilled Pretraining: A modern lens of Data, In-Context Learning and Test-Time Scaling (ICLR 2026 · similarity 0.87)
- Learning What to Remember: Test-Time Training via Context Distillation (NeurIPS 2026 · 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.
