Research map / Large language models / Instruction and fine-tuning
Calibration and knowledge in LLM tuning
Fine-tune LLMs for calibration and multilingual ability using gradient-based parameter selection.
85 accepted papers at ICML, NeurIPS, ICLR, CVPR and AAAI (2021–2026) · an approach in Instruction and fine-tuning, keyword cluster “fine tuning · lora · adaptation”
State History Tried Checklist Preprints Who Export
Where it stands
Not enough benchmark data Fewer than four papers report the same benchmark, so there is no record to follow.
34 papers in 2023–24 and 50 in 2025–26 (at ICML, NeurIPS, ICLR, CVPR and AAAI); its share of the topic shrank, from 16.3% to 7.7%.
How this is decided
A benchmark is tracked when at least four papers of this approach report their own method on the same dataset and metric in at least two different years. Its record is the best number reported up to 2024 compared with the best reported in 2025–26: still improving means at least 1% better, flat means at most 0.2% better with at least two papers from 2025–26 trying. Numbers on a different scale from the rest (a fraction among percentages) and far outliers are left out. Results come from different papers with different setups, so they are signals, not head-to-head comparisons. The share of papers is counted at ICML, NeurIPS, ICLR, CVPR and AAAI.
How it got here
Most cited paper of each year
- 2021 · Process for Adapting Language Models to Society (PALMS) with Values-Targeted Datasets (NeurIPS 2021 · 37 citations)
- 2023 · GPT4Tools: Teaching Large Language Model to Use Tools via Self-instruction (NeurIPS 2023 · 32 citations)
- 2024 · WizardCoder: Empowering Code Large Language Models with Evol-Instruct (ICLR 2024 · 81 citations)
- 2025 · Calibrating Large Language Models with Sample Consistency (AAAI 2025 · 19 citations)
- 2026 · RAG-Enhanced Collaborative LLM Agents for Drug Discovery (AAAI 2026 · 3 citations)
Earliest papers here
- Process for Adapting Language Models to Society (PALMS) with Values-Targeted Datasets (NeurIPS 2021 · 37 citations)
- GPT4Tools: Teaching Large Language Model to Use Tools via Self-instruction (NeurIPS 2023 · 32 citations)
- Memory-Efficient Fine-Tuning of Compressed Large Language Models via sub-4-bit Integer Quantization (NeurIPS 2023 · 27 citations)
What has been tried
61 papers here report 694 measured changes (a component added, removed or swapped, compared with a baseline): helped 71%, no clear effect 7%, hurt 19%, mixed 3%. Papers mostly publish changes that helped, so "helped" is over-represented.
Ablations in recent papers
| What was changed | Measured on | Result | Effect | Paper |
|---|---|---|---|---|
| Use GPT-4 instead of GPT-3.5 to generate evolved instructions | HumanEval · pass@1 | 59.8 -> 62.2 | Helped | WizardCoder: Empowering Code Large Language Models with Evol-Instruct (ICLR 2024 · 81 citations) |
| Add layout-aware pre-training | DocVQA test · ANLS | +1.49% | Helped | LayoutLLM: Layout Instruction Tuning with Large Language Models for Document Understanding (CVPR 2024 · 60 citations) |
| Add pseudo-label CoT tasks | Unseen dataset · Rouge | 55.46 -> 59.20 | Helped | EcomGPT: Instruction-Tuning Large Language Models with Chain-of-Task Tasks for E-commerce (AAAI 2024 · 28 citations) |
| use dynamic batch loading instead of the original RedPajama domain proportions | Downstream tasks · downstream performance | Helped | Sheared LLaMA: Accelerating Language Model Pre-training via Structured Pruning (ICLR 2024 · 20 citations) | |
| Add short-context RoPE readjustment | Proof-pile 4k · perplexity | 4.16 -> 3.71 | Helped | LongRoPE: Extending LLM Context Window Beyond 2 Million Tokens (ICML 2024 · 15 citations) |
| Add reverse-order output comparison data | WMT22 Zh⇒En · BLEU | 23.08 -> 23.41 | Helped | Teaching Large Language Models to Translate with Comparison (AAAI 2024 · 13 citations) |
| Add the hallucinatory comparator | KNIGHT-Judge · F1 Score | 62.05 -> 65.77 | Helped | Improving Factuality in Large Language Models via Decoding-Time Hallucinatory and Truthful Comparators (AAAI 2025 · 6 citations) |
| add dynamic merging of task-specific knowledge to the shared model | GLUE (8 discriminative tasks) · average normalized score | 67.80 -> 96.14 | Helped | Twin-Merging: Dynamic Integration of Modular Expertise in Model Merging (NeurIPS 2024 · 6 citations) |
| Add continuous pre-training | MedKQ&A · GPT-4 win rate | 61% -> 68% | Helped | PediatricsGPT: Large Language Models as Chinese Medical Assistants for Pediatric Applications (NeurIPS 2024 · 5 citations) |
| Use TNCSI_SP instead of TNCSI as the regression target | NAID test set · NDCG | 0.865 -> 0.901 | Helped | From Words to Worth: Newborn Article Impact Prediction with LLM (AAAI 2025 · 4 citations) |
| Add Stage 1 schema alignment training | Schema Alignment · relative score | 66.81 ± 5.61 -> 69.71 ± 6.32 | Helped | Instruction Tuning Large Language Models to Understand Electronic Health Records (NeurIPS 2024 · 4 citations) |
| Add logits-based English Thinking supervision | XQUAD zh · F1 | 23.41 -> 24.47 | Helped | Enhancing Non-English Capabilities of English-Centric Large Language Models Through Deep Supervision Fine-Tuning (AAAI 2025 · 3 citations) |
What reviewers will expect
Datasets and baselines named in the 30 papers of this approach from 2025–2026 that have a full-text summary. Leaving out the common ones is a frequent review question.
Datasets
- GSM8K · 5 of 30 papers
- Winogrande · 2 of 30 papers
- MetaMathQA · 2 of 30 papers
- HumanEval · 2 of 30 papers
- AdvBench · 2 of 30 papers
- FLORES-200 · 2 of 30 papers
- Alpaca · 2 of 30 papers
- MATH · 2 of 30 papers
Baselines
- GPT-4o · 6 of 30 papers
- GPT-3.5 · 5 of 30 papers
- GPT-4 · 4 of 30 papers
- SFT · 3 of 30 papers
- Llama 3.1 70B · 3 of 30 papers
- Llama-3.1-8B · 3 of 30 papers
- Llama 3 70B · 2 of 30 papers
- Llama 3 8B · 2 of 30 papers
Recent preprints
Loading recent arXiv preprints…
Who publishes on it
Institutions on the most papers of this approach since 2023 (of 84 papers; from OpenAlex and the papers' affiliations).
- Tsinghua University · 9 papers
- Microsoft · 5 papers
- Peking University · 5 papers
- Tencent · 4 papers
- Shanghai Artificial Intelligence Laboratory · 4 papers
- Alibaba Group · 3 papers
- Chinese Academy of Sciences · 3 papers
- Princeton University · 3 papers
- University of Illinois, Urbana Champaign · 3 papers
- Fudan University · 3 papers
Use it
Download BibTeX of 25 key papersOpen the topic in the interactive map
Get alerts. Describe your idea in scime atlas and save it (free scime account): you get a weekly email when new preprints land close to it.
Related approaches
Same keyword cluster
- Instruction data selection (63 papers)
- Gradient-free adaptation and merging (24 papers)
Elsewhere in the topic
- Efficient LLM fine-tuning systems (219 papers)
- Fine-tuning generalization and delta compression (159 papers)
- Fine-tuning for reasoning data (158 papers)
- Safety alignment after fine-tuning (120 papers)
- Low-rank adaptation variants (43 papers)
Benchmark numbers, ablations and baseline numbers are extracted by AI from each paper's full text and can be wrong; the paper is the authority. Data: results index of 2026-11-02, citation counts from OpenAlex (2026-10-09).