Papers with Gemma
Exploring Concept Depth: How Large Language Models Acquire Knowledge and Concept at Different Layers? (2025.coling-main)
Copied to clipboard
Mingyu Jin, Qinkai Yu, Jingyuan Huang, Qingcheng Zeng, Zhenting Wang, Wenyue Hua, Haiyan Zhao, Kai Mei, Yanda Meng, Kaize Ding, Fan Yang, Mengnan Du, Yongfeng Zhang
| Challenge: | Large language models have shown remarkable performances across a wide range of tasks, but mechanisms by which they encode tasks of varying complexity remain poorly understood. |
| Approach: | They propose to explore the possibility that LLMs process concepts in different layers . they propose to categorize concepts based on their level of abstraction . |
| Outcome: | The proposed model can process complex concepts in shallow layers, the authors show . the proposed model could be used to prob complex tasks in shallow ones . |
Measuring and Mitigating Shortcut Reliance in Language Models with Probe-Based Representation Entanglement (2026.acl-srw)
Copied to clipboard
| Challenge: | Shortcut learning remains a major obstacle to robust NLP systems. |
| Approach: | They propose to fine-tune Gemma 3 1B Instruct and Llama 3.2 1B on two synthetic sentiment shortcuts in SST-2 and one natural shortcut in MNLI based on lexical overlap. |
| Outcome: | The proposed model improves on two synthetic sentiment shortcuts and one natural shortcut in MNLI with a 99% shortcut ratio, while Gemma drops from 91.8% to 60.2%. |
MEGAVERSE: Benchmarking Large Language Models Across Languages, Modalities, Models and Tasks (2024.naacl-long)
Copied to clipboard
Sanchit Ahuja, Divyanshu Aggarwal, Varun Gumma, Ishaan Watts, Ashutosh Sathe, Millicent Ochieng, Rishav Hada, Prachi Jain, Mohamed Ahmed, Kalika Bali, Sunayana Sitaram
| Challenge: | Several new LLMs have been introduced necessitating their evaluation on non-English languages. |
| Approach: | They perform a thorough evaluation of the non-English capabilities of SoTA LLMs by comparing them on the same set of multilingual datasets. |
| Outcome: | The proposed model outperforms models on multilingual datasets on 22 languages including low-resource African languages. |
MoLA: MoE LoRA with Layer-wise Expert Allocation (2025.findings-naacl)
Copied to clipboard
Chongyang Gao, Kezhen Chen, Jinmeng Rao, Ruibo Liu, Baochen Sun, Yawen Zhang, Daiyi Peng, Xiaoyuan Guo, Vs Subrahmanian
| Challenge: | Recent efforts to integrate low-rank adaptation (LoRA) with the Mixture-of-Experts (MoE) have achieved performance comparable to full-parameter fine-tuning by tuning much fewer parameters. |
| Approach: | They propose a parameter-efficient MoE method for low-rank adaptation with the Mixture-of-Experts (MoE) they use layers of LoRA experts to allocate more LoRA expert to middle layers . |
| Outcome: | The proposed method outperforms baseline models on six well-known NLP and commonsense QA benchmarks on LLAMA-2, Mistral, and Gemma. |
Punctuations and Predicates in Language Models (2026.findings-eacl)
Copied to clipboard
| Challenge: | Recent work has shown that LLMs perform tasks in ways that diverge significantly from human reasoning. |
| Approach: | They examine the computational importance of punctuation tokens in large language models . they use zeroing and layer-swapping techniques to examine their necessity and sufficiency . |
| Outcome: | The proposed model differs in GPT-2, DeepSeek, and Gemma in that punctuation is necessary and sufficient in multiple layers . the findings offer new insight into the internal mechanisms of punctuations in LLMs and have implications for interpretability and model analysis. |
SylloBio-NLI: Evaluating Large Language Models on Biomedical Syllogistic Reasoning (2025.naacl-long)
Copied to clipboard
| Challenge: | Existing models are far from achieving the robustness and consistency required for safe biomedical NLI applications. |
| Approach: | They propose a framework that leverages external ontologies to instantiate diverse syllogistic arguments for biomedical NLI by identifying valid conclusions and extracting supporting evidence. |
| Outcome: | The proposed framework evaluates large language models on identifying valid conclusions and extracting supporting evidence across 28 syllogistic schemes instantiated with human genome pathways. |
Instruction Tuning on Public Government and Cultural Data for Low-Resource Language: a Case Study in Kazakh (2025.acl-long)
Copied to clipboard
| Challenge: | Instruction tuning in low-resource languages remains underexplored due to limited text data, particularly in government and cultural domains. |
| Approach: | They propose to open-source a large-scale instruction-following dataset covering key institutional and cultural knowledge relevant to Kazakhstan. |
| Outcome: | The proposed dataset improves LLMs’ understanding of procedural, legal, and structural governance topics. |
RLHF Can Speak Many Languages: Unlocking Multilingual Preference Optimization for LLMs (2024.emnlp-main)
Copied to clipboard
| Challenge: | Preference optimization is a widely adopted post-training technique to align large language models with human preferences. |
| Approach: | They propose a method for generating multilingual feedback data to balance data coverage. |
| Outcome: | The proposed method achieves 54.4% win-rate against current state-of-the-art multilingual LLM in its parameter class and 69.5% win- rate or higher against widely used models like Gemma, Mistral and Llama 3. |
TUBA: Cross-Lingual Transferability of Backdoor Attacks in LLMs with Instruction Tuning (2025.findings-acl)
Copied to clipboard
Xuanli He, Jun Wang, Qiongkai Xu, Pasquale Minervini, Pontus Stenetorp, Benjamin I. P. Rubinstein, Trevor Cohn
| Challenge: | Despite the increasing support for multilingual capabilities, the impact of backdoor attacks on LLMs remains under-explored. |
| Approach: | They propose to use poisoned instructiontuning data to attack multilingual LLMs . their results show that more powerful models show increased susceptibility to transferable cross-lingual backdoor attacks . |
| Outcome: | The proposed attack is effective in models like BLOOM and GPT-4o with high success rates in more than 7 out of 12 languages. |
Text-to-TrajVis: Enabling Trajectory Data Visualizations from Natural Language Questions (2026.findings-acl)
Copied to clipboard
| Challenge: | Existing datasets for this task are limited and there is no suitable one available. |
| Approach: | They propose a new visualization language called Trajectory Visualization Language (TVL) to facilitate querying trajectory data and generating visualizations. |
| Outcome: | The proposed language can be used to query and generate trajectory data and generate visualizations with large language models. |
“You are Beautiful, Body Image Stereotypes are Ugly!” BIStereo: A Benchmark to Measure Body Image Stereotypes in Language Models (2025.findings-acl)
Copied to clipboard
| Challenge: | BIStereo is a suite of language models that uncover body image stereotypes in language models. |
| Approach: | They propose a metric, TriSentBias, that captures the biased preferences of LMs towards a certain body type over others. |
| Outcome: | The proposed metric captures biased preferences of LMs towards a certain body type over others. |
Towards Understanding the Robustness of Sparse Autoencoders (2026.findings-acl)
Copied to clipboard
| Challenge: | Large Language Models (LLMs) are vulnerable to optimization-based jailbreak attacks that exploit internal gradient structure. |
| Approach: | They propose to integrate pretrained Sparse Autoencoders into transformer residual streams at inference time without modifying model weights or blocking gradients. |
| Outcome: | The proposed model reduces jailbreak success rate by 5x compared to baseline models . compared with models with weak white-box attacks, the proposed model is more robust . |
MathAgent: Adversarial Evolution of Constraint Graphs for Mathematical Reasoning Data Synthesis (2026.findings-acl)
Copied to clipboard
| Challenge: | Current approaches to synthesising high-quality mathematical reasoning data without human priors suffer from mode collapse and limited logical complexity. |
| Approach: | They propose a hierarchical synthesis framework that formulates data synthesis as an unsupervised optimization problem over a constraint graph followed by semantic instantiation rather than a direct text generation task. |
| Outcome: | The proposed framework outperforms widely-used datasets on eight mathematical benchmarks. |
Structured Semantic Information Helps Retrieve Better Examples for In-Context Learning Applied to Few-Shot Relation Extraction (2026.acl-long)
Copied to clipboard
| Challenge: | Existing approaches to relation extraction require many training examples per relation, resulting in low results. |
| Approach: | They propose a strategy where new examples are selected based on their similarity to the provided 1-shot example. |
| Outcome: | The proposed strategy outperforms other methods on FS-TACRED and FS - FewRel subsets and achieves state-of-the-art performance on both datasets. |
Take Out Your Calculators: Estimating the Real Difficulty of Question Items with LLM Student Simulations (2026.findings-acl)
Copied to clipboard
| Challenge: | Standardized math assessments require expensive human pilot studies to establish the difficulty of test items. |
| Approach: | They propose to use large language models to model difficulty of multiple-choice math questions for real-world students. |
| Outcome: | The proposed model predicts difficulty of multiple-choice math questions for students . correlations between model and real-world difficulty are high, the authors show . |
A Head to Predict and a Head to Question: Pre-trained Uncertainty Quantification Heads for Hallucination Detection in LLM Outputs (2025.emnlp-main)
Copied to clipboard
Artem Shelmanov, Ekaterina Fadeeva, Akim Tsvigun, Ivan Tsvigun, Zhuohan Xie, Igor Kiselev, Nico Daheim, Caiqi Zhang, Artem Vazhentsev, Mrinmaya Sachan, Preslav Nakov, Timothy Baldwin
| Challenge: | Uncertainty quantification (UQ) is a framework for assessing the reliability of model outputs. |
| Approach: | They introduce pre-trained UQ heads for LLMs that are highly robust and generalized to languages they were not explicitly trained on. |
| Outcome: | The pre-trained heads significantly improve their ability to capture uncertainty compared to unsupervised methods. |
SafetyALFRED: Evaluating Safety-Conscious Planning of Vision Language Models (2026.findings-acl)
Copied to clipboard
Josue Torres-Fonseca, Naihao Deng, Yinpei Dai, Shane Storks, Yichi Zhang, Rada Mihalcea, Casey Kennington, Joyce Chai
| Challenge: | Existing safety evaluations focus on hazard recognition through disembodied question answering (QA) settings, but lack a critical gap in evaluating an agent. |
| Approach: | They evaluate multimodal large language models with six categories of kitchen hazards . they propose a safety-based approach that prioritizes multi-step corrective actions . |
| Outcome: | The proposed model can recognize hazards in QA settings, but average mitigation success rates are low . the proposed model is based on the embodied agent benchmark ALFRED . |