Research map / Large language models / Instruction and fine-tuning
LoFiT: Localized Fine-tuning on LLM Representations
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 LoFiT, a two-step localized fine-tuning method that selects a sparse set of attention heads via learned scaling factors and then tunes offset bias vectors on those heads, matching LoRA/RED/ReFT with 20x-200x fewer parameters.
Problem
Representation intervention methods localize behavior to attention heads but are learning-free and weak; PEFT methods modify representations uniformly or treat module choice as a hyperparameter without localization. The paper asks whether localization can make representation fine-tuning both effecti
Approach
Learn per-head scaling factors with L1 regularization to score and select the top-K attention heads, then fine-tune only bias offset vectors at those heads, combining interpretability-based localization with PEFT-style training.
Key results
- On Llama 2-13B, LoFiT averages 74.6 across TruthfulQA MC1/MC2, MQuAKE, and CLUTRR versus 55.1 for RepE and 48.3 for ITI.
- LoFiT matches LoRA on average across Gemma-7B and Llama 2-13B evaluation sets (75.9 vs 77.1 on Gemma-7B; 75.6 vs 75.1 on Llama 2-13B) while using 20x-200x fewer learned parameters.
- Under the low-data setting, LoFiT with 500 or fewer examples outperforms ITI and RepE by large margins, and head-selection ablations show task-specific heads beat different-task heads on TruthfulQA and MQuAKE.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use LoFiT instead of representation-intervention baselines | Gemma-7B · TruthfulQA | MC1 accuracy: 38.5 -> 60.5 | Helped |
| Use LoFiT instead of representation-intervention baselines | Llama 2-7B · MQuAKE | EM: 37.9 -> 73.4 | Helped |
| Select heads using LoFiT scaling-factor norms rather than random sampling (ablation) | LoFiT bias tuning, average across models and tasks · TruthfulQA, MQuAKE, and CLUTRR | Average accuracy: 69.4 -> 73.4 | Helped |
| Select heads using LoFiT scaling-factor norms rather than bias-vector norms (ablation) | LoFiT bias tuning, average across models and tasks · TruthfulQA, MQuAKE, and CLUTRR | Average accuracy: 69.9 -> 73.4 | Helped |
| Select heads using LoFiT scaling-factor norms rather than ITI head selection (ablation) | LoFiT bias tuning, average across models and tasks · TruthfulQA, MQuAKE, and CLUTRR | Average accuracy: 71.9 -> 73.4 | Helped |
| Use task-specific rather than different-task heads (ablation) | LoFiT bias tuning · TruthfulQA and MQuAKE | Accuracy | Helped |
| Use different-task heads instead of task-specific heads (ablation) | LoFiT bias tuning · CLUTRR | EM | No clear effect |
| Increase the selected-head proportion below 10% (ablation) | LoFiT bias tuning on MQuAKE and CLUTRR · MQuAKE and CLUTRR | Test accuracy | Helped |
| Increase the selected-head proportion to 10%-20% or more (ablation) | LoFiT bias tuning on MQuAKE and CLUTRR · MQuAKE and CLUTRR | Test accuracy | No clear effect |
| Use LoFiT rather than LoRA or RED in the low-data setting (n ≤ 100) | Llama 2-7B · CLUTRR and MQuAKE | Test accuracy | Helped |
| Use LoFiT rather than LoRA on commonsense and math benchmarks | Llama 2-7B and Llama 2-13B · SIQA, ARC-c, BoolQ, and SVAMP | Average accuracy: 62.0 -> 60.7; 70.4 -> 69.1 | Hurt |
| Fine-tune with LoFiT rather than leave the base model unfine-tuned | Llama 2-7B-Chat · TriviaQA | Accuracy: 76.5 -> 76.7 | Helped |
Surprising findings
- Tuning biases on heads selected for a different task can perform worse than random head selection on TruthfulQA (e.g., Llama 2-13B).
- Selecting heads by norm of learned scaling factors outperforms selecting by norm of the offset vectors themselves.
Limitations
- Evaluated only on English truthfulness and reasoning tasks with short contexts and responses.
- Only autoregressive transformer models up to 13B parameters were tested.
- The two-step procedure fits extra parameters in the head-selection stage, and the method is more costly than learning-free interventions.
Experimental setup
Datasets: TruthfulQA CLUTRR MQuAKE SIQA ARC-c BoolQ SVAMP TriviaQA Natural Questions MMLU
Benchmarks: TruthfulQA MQuAKE CLUTRR SIQA ARC-c BoolQ SVAMP TriviaQA Natural Questions MMLU
Metrics: MC1 MC2 exact match accuracy GPT-4 truthfulness and informativeness (True x Info)
Baselines: 0-shot ITI RepE LoRA RED ReFT Random head selection Probe-layers Bias-based selection
Builds on: Inference-time Intervention Representation Engineering RED LoRA ReFT direct preference optimization
Closest papers by meaning
- Beyond LoRA vs. Full Fine-Tuning: Gradient-Guided Optimizer Routing for LLM Adaptation (NeurIPS 2026 · similarity 0.89)
- Supervised Knowledge Makes Large Language Models Better In-context Learners (ICLR 2024 · similarity 0.89)
- Loquetier: A Virtualized Multi-LoRA Framework for Unified LLM Fine-tuning and Serving (NeurIPS 2025 · similarity 0.89)
- DELIFT: Data Efficient Language model Instruction Fine-Tuning (ICLR 2025 · similarity 0.89)
- Joint Localization and Activation Editing for Low-Resource Fine-Tuning (ICML 2025 · similarity 0.89)
- LongLoRA: Efficient Fine-tuning of Long-Context Large Language Models (ICLR 2024 · similarity 0.89)
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.