ProFit: Leveraging High-Value Signals in SFT via Probability-Guided Token Selection (2026.findings-acl)
Copied to clipboard
| Challenge: | Traditional fine-tuning ignores one-to-many nature of language, leading to overfitting . authors propose a method to fine- tune LLMs by leveraging tokens. |
| Approach: | They propose a method to fine-tune Large Language Models by leveraging tokens to mask low-probability tokens. |
| Outcome: | The proposed method outperforms baselines on general reasoning and mathematical benchmarks. |
Similar Papers
Select Before Use: On the Importance of Reference Model Selection in Preference Alignment (2026.acl-long)
Copied to clipboard
| Challenge: | Supervised Fine-Tuning (SFT) is used as the initialization and reference model for subsequent preference alignment. |
| Approach: | They propose to use RewardRank to estimate initial implicit alignment between reference model and preference objective to ensure LLMs generate safe, helpful, and instruction-aligned content. |
| Outcome: | Empirical evidence shows that using the selected model as reference can gain up to 67.6% relative increase on length-controlled win rate compared to baselines. |
Rethinking Data Selection at Scale: Random Selection is Almost All You Need (2025.findings-emnlp)
Copied to clipboard
| Challenge: | Existing data selection techniques are designed for small data pools, a study finds . filtering data by token length is an efficient method for improving results . |
| Approach: | They use self-scoring methods that do not rely on external help to perform fine-tuning . they also find that filtering data by token length offers a stable and efficient method . |
| Outcome: | The proposed methods outperform random selection on large datasets on large data pools. |
Massive Supervised Fine-tuning Experiments Reveal How Data, Layer, and Training Factors Shape LLM Alignment Quality (2025.emnlp-main)
Copied to clipboard
| Challenge: | Recent advances in large language models (LLMs) have greatly improved natural language understanding and generation. |
| Approach: | They train a wide range of base models on a variety of datasets including code generation, mathematical reasoning, and general-domain tasks. |
| Outcome: | The results show that training–task synergies persist across all models while others vary substantially, emphasizing the importance of model-specific strategies. |
Intuitive Fine-Tuning: Towards Simplifying Alignment into a Single Process (2025.acl-long)
Copied to clipboard
| Challenge: | Supervised Fine-Tuning (SFT) and Preference Optimization (PO) are key processes for aligning Language Models with human preferences post pre-training. |
| Approach: | They propose to combine Supervised Fine-Tuning and Preference Optimization (PO) with two sub-processes defined at token level within the Markov Decision Process (MDP) |
| Outcome: | The proposed process performs comparably or even superiorly to SFT and some typical PO methods across several tasks, particularly those requires generation, reasoning, and fact-following abilities. |
The Emperor’s New Reasoning: Format Imitation Overshadows Genuine Mathematical Understanding in SFT (2025.emnlp-main)
Copied to clipboard
| Challenge: | Recent advances in large language models (LLMs) have yielded impressive gains on mathematical reasoning benchmarks via supervised fine-tuning (SFT). |
| Approach: | They investigate the mechanisms behind SFT improvements in small-scale large language models by examining four key questions: (1) Are performance gains primarily due to format alignment rather than reasoning? (2) Can high-quality supervision encourage genuine reasoning? (4) Are format alignment gains consistent across model sizes and architectures? |
| Outcome: | The proposed models outperform the proprietary models on OlympiadBench and Omni-Math, but lack the brittleness of the models under perturbations to test their reasoning abilities. |
Prefix Text as a Yarn: Eliciting Non-English Alignment in Foundation Language Model (2024.findings-acl)
Copied to clipboard
| Challenge: | supervised fine-tuning (SFT) has been a straightforward approach for tailoring the output of foundation large language models (LLMs) to specific preferences. |
| Approach: | They propose a training-free alignment method that uses minimal prior tokens to bridge the foundation LLM and the SFT LLM. |
| Outcome: | The proposed method achieves comparable performance without training on machine translation and part-of-speech tagging across seven languages. |
Mitigating Training Imbalance in LLM Fine-Tuning via Selective Parameter Merging (2024.emnlp-main)
Copied to clipboard
| Challenge: | Existing studies suggest that the order of training samples can affect model performance, but this is not the case. |
| Approach: | They propose to merge supervised fine-tuning models with different data orders to mitigate this imbalance by parameter merging. |
| Outcome: | The proposed method outperforms the weighted-average method on five datasets. |
SED-SFT: Selectively Encouraging Diversity in Supervised Fine-Tuning (2026.acl-short)
Copied to clipboard
| Challenge: | Existing studies have proposed a new approach to optimize for SFT followed by RL . existing studies have suggested a method to optimize SFT for large language models . |
| Approach: | They propose a framework that encourages diversity based on token exploration space. |
| Outcome: | Experiments show that SED-SFT significantly improves generation diversity with a negligible computational overhead increase over CE loss. |
Reinforcement Learning with Supervised Alignment (2025.findings-emnlp)
Copied to clipboard
| Challenge: | Supervised fine-tuning (SFT) is a widely used method for adapting Large Language Models to specific tasks. |
| Approach: | They propose a method that uses supervised fine-tuning to train a reward model for reinforcement learning. |
| Outcome: | The proposed method outperforms existing methods on in-domain benchmarks but surpasses them 50 times on out-of-domain and cross-task evaluations. |
Disperse-Then-Merge: Pushing the Limits of Instruction Tuning via Alignment Tax Reduction (2024.findings-acl)
Copied to clipboard
| Challenge: | Pre-trained language models may not follow human instructions and produce toxic, hallucinated, or biased content. |
| Approach: | They propose a disperse-then-merge framework that dispersers instruction-following data into portions and trains multiple sub-models using different data portions. |
| Outcome: | The proposed framework outperforms data curation and training regularization on standard knowledge and reasoning benchmarks. |