Papers with LLM fine-tuning
Text2Sql: Pure Fine-Tuning and Pure Knowledge Distillation (2025.naacl-industry)
Copied to clipboard
| Challenge: | Text2Sql is a task that translates natural language questions and database schemas into SQL queries. |
| Approach: | They employ pure fine-tuning strategy to reduce redundancy by using only 53% of the baseline prompt length to fine- tune the model. |
| Outcome: | The model outperforms the baseline model by 8.2% and 8.6% in Test-suite accuracy (TS) and exact-set-match accuracy (EM) under the most refined Spider dev set of prompts, the model achieves 73.5% and 75.4%, respectively, approaching state-of-the-art (SOTA) levels. |
Acceleration of Backpropagation in Linear Layers of Transformer Models Based on Gradient Structure (2026.eacl-srw)
Copied to clipboard
| Challenge: | a sparsity-exploiting backward pass is a memory-efficient way to accelerate LLM fine-tuning. |
| Approach: | They propose a method that exploits padding-induced gradient sparsity to accelerate backward computation. |
| Outcome: | The proposed method achieves a backward pass speedup of 2.15x on GLUE and 1.99x on reasoning benchmarks while maintaining memory usage identical to the regular PyTorch fine-tuning. |
AdaRefiner: Refining Decisions of Language Models with Adaptive Feedback (2024.findings-naacl)
Copied to clipboard
| Challenge: | Large Language Models (LLMs) have demonstrated significant success across various domains, but their application in complex decision-making tasks often necessitates intricate prompt engineering or fine-tuning. |
| Approach: | They propose a lightweight Adapter Language Model (LM) which automatically refines task comprehension based on feedback from RL agents. |
| Outcome: | The proposed framework enhances synergy between LLMs and RL feedback while maintaining generalization abilities and enhancing decision-making capabilities in downstream tasks. |
SYMDIREC: A Neuro-Symbolic Divide-Retrieve-Conquer Framework for Enhanced RTL Synthesis and Summarization (2026.eacl-industry)
Copied to clipboard
Prashanth Vijayaraghavan, Apoorva Nitsure, Luyao Shi, Charles Mackin, Ashutosh Jadhav, David Beymer, Ehsan Degan, Vandana Mukherjee
| Challenge: | Existing prompting and retrieval-augmented generation methods lack symbolic planning . rigid HDL syntax, limited supervision, and weak alignment with natural language hinder RTL synthesis and summarization. |
| Approach: | SYMDIREC decomposes RTL tasks into symbolic subgoals and assembles verified outputs . a neuro-symbolic framework supports both Verilog and VHDL without LLM fine-tuning . |
| Outcome: | SYMDIREC achieves higher Pass@1 rates for synthesis and 15–20% ROUGE-L improvements for summarization over prompting and RAG . synthesis, summarizing require preserving strict HDL syntax, modular structure, and precise functional semantics, authors show . |
Exploring Forgetting in Large Language Model Pre-Training (2025.acl-long)
Copied to clipboard
| Challenge: | Existing research on task-level forgetting in LLMs has focused on pretraining . but, there is limited attention to finer-grained forgetting during training . |
| Approach: | They investigated the existence and measurement of forgetting in pre-training . they examined low-cost, straightforward methods to mitigate forgetting during the pre- training phase . |
| Outcome: | The proposed methods could be used to mitigate forgetting during the pre-training phase and offer insights into the dynamics of forgetting. |
Unsupervised Text Style Transfer for Controllable Intensity (2026.findings-eacl)
Copied to clipboard
| Challenge: | Unsupervised Text Style Transfer (UTST) aims to transfer the stylistic properties of a given text without parallel text pairs. |
| Approach: | They propose a SFT-then-PPO paradigm to fine-tune an LLM with parallel data and reward functions for distinguishing stylistic intensity in hierarchical levels. |
| Outcome: | The proposed system can transfer stylistic properties without parallel text pairs even for adjacent levels of intensity. |
TRACT: Regression-Aware Fine-tuning Meets Chain-of-Thought Reasoning for LLM-as-a-Judge (2025.acl-long)
Copied to clipboard
| Challenge: | Existing methods for fine-tuning LLMs use cross-entropy (CE) loss . existing methods neglect the numeric nature of score prediction . |
| Approach: | They propose a method that fine-tunes large language models (LLMs) for automated text evaluation, assigning a score to the input based on scoring rubrics. |
| Outcome: | The proposed model outperforms existing methods in four LLM-as-a-judge datasets and two LLMs. |
Hi-ZFO: Hierarchical Zeroth- and First-Order LLM Fine-Tuning via Importance-Guided Tensor Selection (2026.findings-acl)
Copied to clipboard
| Challenge: | generative tasks require a high degree of exploratory capacity, but zeroth-order methods suffer from slow convergence . generative task-specific methods tend to converge toward local minima, causing noise and inefficient estimation . |
| Approach: | They propose a framework that synergizes FO precision with exploratory capability of ZO estimation. |
| Outcome: | The proposed framework synergizes precision of FO gradients with exploratory capability of ZO estimation. |
IRPO: Implicit Policy Regularized Preference Optimization (2026.findings-eacl)
Copied to clipboard
| Challenge: | Recent DPOs introduce additional hyperparameters, reducing feasibility for LLM fine-tuning. |
| Approach: | They propose an algorithm that regularizes the reward against a reference policy without extra hyperparameters to address suboptimal outcomes. |
| Outcome: | The proposed algorithm outperforms baseline algorithms with the same hyperparameter complexity while maintaining training simplicity. |
FRAG: A Flexible Modular Framework for Retrieval-Augmented Generation based on Knowledge Graphs (2025.findings-acl)
Copied to clipboard
| Challenge: | Existing methods for retrieval-augmented generation struggle with a trade-off between flexibility and retrieval quality. |
| Approach: | They propose a flexible modular KG-RAG framework that uses query text instead of KGs . they propose to use query text to infer the structural information of reasoning paths . |
| Outcome: | The proposed method achieves state-of-the-art performance with high efficiency and low resource consumption. |
A Study of Implicit Ranking Unfairness in Large Language Models (2024.findings-emnlp)
Copied to clipboard
| Challenge: | Large language models (LLMs) have demonstrated superior ability to serve as ranking models, but they will exhibit discriminatory ranking behaviors based on users’ sensitive attributes (gender). |
| Approach: | They propose an evaluation method to investigate the severity of implicit ranking unfairness and a pair-wise regression method to conduct fair-aware data augmentation for LLM fine-tuning. |
| Outcome: | The proposed method outperforms existing methods in ranking fairness, achieving this with only a small reduction in accuracy. |
On Bilingual Lexicon Induction with Large Language Models (2023.emnlp-main)
Copied to clipboard
| Challenge: | Existing approaches to induction bilingual lexicons still require cross-lingual word representations . a recent study shows that few-shot prompting with in-context examples from nearest neighbours achieves the best performance . |
| Approach: | They examine whether it is possible to prompt and fine-tune multilingual LLMs for BLI . they experiment with 18 open-source text-to-text mLLMs of different sizes . |
| Outcome: | The proposed approach is compared with existing approaches on two standard BLI benchmarks covering a range of typologically diverse languages. |
Structure Trumps Size: Rethinking Data Quality for LLM Reasoning (2025.findings-emnlp)
Copied to clipboard
| Challenge: | Existing methods for fine-tuning Large Language Models rely on heuristic strategies and lack systematic, quantitative frameworks for evaluating data quality. |
| Approach: | They propose a multi-dimensional quantitative framework for reasoning data management . they rigorously evaluate and optimize datasets along six orthogonal dimensions . |
| Outcome: | The proposed framework rigorously evaluates and optimizes datasets along six orthogonal dimensions. |
LoRaDA: Low-Rank Direct Attention Adaptation for Efficient LLM Fine-tuning (2025.findings-emnlp)
Copied to clipboard
| Challenge: | Recent advances in parameter-efficient fine-tuning techniques allow for adjustments to only a minor fraction of the parameters of language models. |
| Approach: | They propose a low-rank direct attention adapted method for efficient LLM fine-tuning . they propose LMAM, which can bring negative attention to self-attention modules . |
| Outcome: | The proposed method outperforms the full fine-tuning method by 2.1% on GLUE benchmark. |
HyQE: Ranking Contexts with Hypothetical Query Embeddings (2024.findings-emnlp)
Copied to clipboard
| Challenge: | Existing approaches to rank contexts rely on similarity between contexts and queries, but these methods are limited by the number of candidate contexts. |
| Approach: | They propose a scalable ranking framework that combines embedding similarity and large language models without fine-tuning. |
| Outcome: | The proposed framework improves the performance across multiple benchmarks. |
G-LoRA: Global-Local Decoupled Low-Rank Adaptation (2026.findings-acl)
Copied to clipboard
| Challenge: | Low-Rank Adaptation (LoRA) improves the fine-tuning efficiency and performance of large language models. |
| Approach: | They propose a low-rank adaptive approach that decomposes update matrix into global and local adapters and assigns them to local and global adapters. |
| Outcome: | The proposed method achieves up to 2.7% accuracy improvement over LoRA and its variants on commonsense reasoning, mathematical reasoning, and code generation. |
Steering LLM Thinking with Budget Guidance (2026.findings-acl)
Copied to clipboard
| Challenge: | Existing budget control methods for large language models are inadequate for long reasoning . budget guidance can be used to control reasoning length without fine-tuning . |
| Approach: | They propose a budget guidance method that models a Gamma distribution over remaining thinking length during next-token generation and uses it to guide generation in a soft, token-level manner. |
| Outcome: | The proposed method achieves up to 26% accuracy gain on the MATH-500 benchmark compared to baseline methods while maintaining competitive accuracy with only 63% of the thinking tokens used by the full-thinking model. |