Papers with feed-forward layers
LM-Debugger: An Interactive Tool for Inspection and Intervention in Transformer-Based Language Models (2022.emnlp-demos)
Copied to clipboard
| Challenge: | Transformer-based language models (LMs) are opaque and unexplained, causing problems for endusers and developers who wish to debug or fix their behaviour. |
| Approach: | They propose an interactive debugger tool for transformer-based LMs that provides a fine-grained interpretation of the model's internal prediction process and a powerful framework for intervening in LM behavior. |
| Outcome: | The proposed tool provides a fine-grained interpretation of the model's internal prediction construction process, and a powerful framework for intervening in LM behavior. |
LV-BERT: Exploiting Layer Variety for BERT (2021.findings-acl)
Copied to clipboard
| Challenge: | Modern pre-trained language models are mostly built upon stereotyped development sets . LV-BERT model obtained by our method outperforms BERT on various downstream tasks . |
| Approach: | They propose to exploit layer variety from the layer type set and the layer order to improve pre-trained models. |
| Outcome: | The proposed model outperforms BERT and its variants on various downstream tasks. |
Attention Under Attack: Analog Noise Effects and Mechanistic Vulnerabilities in Transformer Models (2026.acl-short)
Copied to clipboard
| Challenge: | Analog in-memory computing (AIMC) provides substantial efficiency gains for transformer inference but introduces hardware-induced noise that can distort attention behavior. |
| Approach: | They present the first fine-grained analysis of analog vulnerability in pretrained transformers . query (Q), key (K), and value (V) projections are most sensitive components . |
| Outcome: | The proposed analysis shows that query (Q), key (K), and value (V) projections are the most sensitive components . |
From Language Modeling to Instruction Following: Understanding the Behavior Shift in LLMs after Instruction Tuning (2024.naacl-long)
Copied to clipboard
| Challenge: | Large Language Models (LLMs) have achieved remarkable success in aligning with user intentions. |
| Approach: | They develop local and global explanation methods and a feed-forward-based method for input-output attribution to investigate the impact of instruction tuning on user intentions. |
| Outcome: | The proposed method compares explanations from pre-trained and instruction-tuned models . it empowers LLMs to recognize the instruction parts of user prompts, it encourages response generation . |
ModularMoE: Fast LLM Customization with Parameter-Sharing Mixture-of-Experts for Low-Resource Settings (2026.findings-acl)
Copied to clipboard
Jiaxing Liu, Qi Qi, Haifeng Sun, Dunjun Li, Zirui Zhuang, Bo He, Xiang Yang, Cong Liu, Jianxin Liao, Jingyu Wang
| Challenge: | Large Language Models impose significant computational and storage burdens on personal devices . existing customization approaches incur excessive computational costs or lead to suboptimal performance . |
| Approach: | They propose a training framework that converts pre-trained LLMs into parameter-sharing MoE models for lightweight deployment. |
| Outcome: | The proposed training framework outperforms state-of-the-art training frameworks at the same sparsity level while delivering up to 2.71 inference speedup. |
Neuron Specialization: Leveraging Intrinsic Task Modularity for Multilingual Machine Translation (2024.emnlp-main)
Copied to clipboard
| Challenge: | Language-specific modeling methods that focus on heuristics for allocation of capacity and lack knowledge transfer capabilities are often prone to interference due to conflicting optimization demands. |
| Approach: | They propose a method that identifies specialized neurons to modularize feed-forward layers and updates them through sparse networks to avoid interference under multilingual translation. |
| Outcome: | The proposed approach achieves consistent performance gains over strong baselines with additional analyses showing reduced interference and increased knowledge transfer. |
An Investigation of Neuron Activation as a Unified Lens to Explain Chain-of-Thought Eliciting Arithmetic Reasoning of LLMs (2024.acl-long)
Copied to clipboard
| Challenge: | Prior work has focused on ablating components in the CoT prompt, but the reason why these components are important to LLM reasoning is not explored. |
| Approach: | They investigate "neuron activation" as a lens to provide a unified explanation to previous work . they propose an approach to automatically identify neurons that imply arithmetic reasoning . |
| Outcome: | The proposed approach can explain the importance of components in a CoT prompt . it also automatically identifies neurons that imply arithmetic reasoning . |
F-MALLOC: Feed-forward Memory Allocation for Continual Learning in Neural Machine Translation (2024.naacl-long)
Copied to clipboard
| Challenge: | Existing approaches to address Catastrophic Forgetting (CF) have been developed to avoid forgetting and maintain system extensibility. |
| Approach: | They propose a method to reduce Catastrophic Forgetting (CF) by decomposing feed-forward layers into discrete memory cells and ensuring robust extendability. |
| Outcome: | The proposed method achieves higher BLEU scores and almost zero forgetting while maintaining robust extendability. |
Transformer Feed-Forward Layers Are Key-Value Memories (2021.emnlp-main)
Copied to clipboard
| Challenge: | Using feed-forward layers, we show that the learned patterns are human-interpretable, and that lower layers tend to capture shallow patterns, while upper layers learn more semantic ones. |
| Approach: | They propose that feed-forward layers in transformer-based language models operate as key-value memories where each key correlates with textual patterns in the training examples and each value induces a distribution over the output vocabulary. |
| Outcome: | The proposed model is based on key-value memories with a key-level correlation with the training examples and a distribution over the output vocabulary. |
Enhancing Chain-of-Thought Reasoning via Neuron Activation Differential Analysis (2025.emnlp-main)
Copied to clipboard
| Challenge: | Existing studies focus on optimizing external components of CoT, but lack internal explanations for the quality of the model's outputs. |
| Approach: | They propose an efficient method to identify reasoning-critical neurons by analyzing their activation patterns under reasoning chains of varying quality. |
| Outcome: | The proposed method shows that neurons in the feed-forward layers are critical in the generation of high-quality reasoning chains. |
Why LLMs Hallucinate on Structured Knowledge: A Mechanistic Analysis of Reasoning over Linearized Representations (2026.acl-long)
Copied to clipboard
Shanghao Li, Jinda Han, Yibo Wang, Yuanjie Zhu, Zihe Song, Langzhou He, Kenan Kamel A Alghythee, Philip S. Yu
| Challenge: | Existing literature primarily addresses this problem through external interventions such as retrieval augmentation and prompt engineering at the input or output level. |
| Approach: | They find that LLMs can still produce hallucinated outputs when using structured external knowledge. |
| Outcome: | The proposed models fail to ground the provided knowledge, causing the model to revert to parametric memory. |