Research map / Large language models / Instruction and fine-tuning
Gradient-free adaptation and merging
Adapt LLMs through hypernetworks, model merging, and search without full backpropagation.
24 accepted papers at ICML, NeurIPS, ICLR, CVPR and AAAI (2023–2026) · 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.
11 papers in 2023–24 and 13 in 2025–26 (at ICML, NeurIPS, ICLR, CVPR and AAAI); its share of the topic shrank, from 5.3% to 2.0%.
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
- 2023 · HyperTuning: Toward Adapting Large Language Models without Back-propagation (ICML 2023 · 9 citations)
- 2024 · LISA: Reasoning Segmentation via Large Language Model (CVPR 2024 · 381 citations)
- 2025 · API Pack: A Massive Multi-Programming Language Dataset for API Call Generation (ICLR 2025 · 1 citation)
- 2026 · Extra-Merge: Tracing the Rank-1 Subspace of Model Merging in Language Model Pre-Training (ICML 2026 · 0 citations)
Earliest papers here
- HyperTuning: Toward Adapting Large Language Models without Back-propagation (ICML 2023 · 9 citations)
- LISA: Reasoning Segmentation via Large Language Model (CVPR 2024 · 381 citations)
- Refusal in Language Models Is Mediated by a Single Direction (NeurIPS 2024 · 57 citations)
What has been tried
16 papers here report 199 measured changes (a component added, removed or swapped, compared with a baseline): helped 74%, no clear effect 4%, 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 |
|---|---|---|---|---|
| Apply LoRA to the SAM vision backbone | ReasonSeg val · gIoU | 44.4 -> 41.5 | Hurt | LISA: Reasoning Segmentation via Large Language Model (CVPR 2024 · 381 citations) |
| Ablate the refusal direction from activations at all layers and token positions | JailbreakBench harmful instructions · refusal rate | Hurt | Refusal in Language Models Is Mediated by a Single Direction (NeurIPS 2024 · 57 citations) | |
| Replace entity-type labels with generated type definitions | UniversalNER benchmark (43 datasets) · average F1 | 11.8% | Hurt | UniversalNER: Targeted Distillation from Large Language Models for Open Named Entity Recognition (ICLR 2024 · 25 citations) |
| Replace uniform training noise with Gaussian noise (α=5) | AlpacaEval, Evol-Instruct · ChatGPT win rate | 67.58 -> 65.09 | Hurt | NEFTune: Noisy Embeddings Improve Instruction Finetuning (ICLR 2024 · 14 citations) |
| Remove LoReFT orthogonality constraint and difference operation (DiReFT) | Commonsense reasoning (8-task Avg.) · accuracy | 80.2 -> 79.0; 83.3 -> 82.2; 81.8 -> 80.9; 86.6 -> 85.4 | Hurt | ReFT: Representation Finetuning for Language Models (NeurIPS 2024 · 10 citations) |
| Add orthogonal directions | TruthfulQA, component ablation · True*Info (%) | 29.6 -> 36.7 | Helped | Truth Forest: Toward Multi-Scale Truthfulness in Large Language Models through Intervention without Tuning (AAAI 2024 · 6 citations) |
| Apply ICV to all layers instead of only the last layer | ParaDetox · Toxicity (%) | 76.27 -> 34.77 | Helped | In-context Vectors: Making In Context Learning More Effective and Controllable Through Latent Space Steering (ICML 2024 · 4 citations) |
| Fine-tune CodeLlama-13B on 20,000 Python API Pack instances with the 3-shot template instead of the 0-shot template | Level 3 (unseen APIs and endpoints), 3-shot retrieved evaluation · API Call accuracy | 44.1% to 49.5% | Helped | API Pack: A Massive Multi-Programming Language Dataset for API Call Generation (ICLR 2025 · 1 citation) |
| Use schedule 2 instead of schedule 1 | UL2 pretraining · All Tasks Average | 26.7 -> 26.8 | Mixed | On the Inductive Bias of Stacking Towards Improving Reasoning (NeurIPS 2024 · 1 citation) |
| Add DPO fine-tuning | SymBench total · Average normalized score | 79.1 -> 86.4 | Helped | CodeSteer: Symbolic-Augmented Language Models via Code/Text Guidance (ICML 2025 · 0 citations) |
| add balance loss | BFCL · AVG accuracy | 72.3 -> 75.1 | Helped | MoRAgent: Parameter Efficient Agent Tuning with Mixture-of-Roles (ICML 2025 · 0 citations) |
| Increase MF rank from 1 to 8 | OLMo-7B · RMSE | 7.31 -> 7.12 | Helped | Demystifying Language Model Forgetting with Low-rank Example Associations (NeurIPS 2025 · 0 citations) |
What reviewers will expect
Datasets and baselines named in the 8 papers of this approach from 2025–2026 that have a full-text summary. Leaving out the common ones is a frequent review question.
Datasets
- ToolBench · 2 of 8 papers
- MMLU · 2 of 8 papers
Baselines
- GPT4 · 2 of 8 papers
Recent preprints
Loading recent arXiv preprints…
Who publishes on it
Institutions on the most papers of this approach since 2023 (of 24 papers; from OpenAlex and the papers' affiliations).
- Stanford University · 3 papers
- Anthropic · 3 papers
- Carnegie Mellon University · 2 papers
- University of Maryland, College Park · 2 papers
- Google · 2 papers
- University of California, Berkeley · 2 papers
- Microsoft · 2 papers
- University of Southern California · 2 papers
- Massachusetts Institute of Technology · 2 papers
- Google DeepMind · 2 papers
Use it
Download BibTeX of 21 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
- Calibration and knowledge in LLM tuning (85 papers)
- Instruction data selection (63 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).