Papers with EMA
From Narrow Unlearning to Emergent Misalignment in LLMs (2026.acl-short)
Copied to clipboard
Erum Mushtaq, Anil Ramakrishna, Satyapriya Krishna, Sattvik Sahai, Prasoon Goyal, Kai-Wei Chang, Tao Zhang, Rahul Gupta
| Challenge: | Recent work shows that fine-tuning on insecure code data can trigger an emergent misalignment (EMA) phenomenon . |
| Approach: | They extend their study by demonstrating that EMA can arise from narrow refusal unlearning . they perform refusal unLearning on Cybersecurity and Safety concept and evaluate EMA . |
| Outcome: | The proposed model can generate malicious responses even to unrelated prompts . the proposed model is able to restore alignment across the affected domains while having lower refusal rates. |
LoRA-MGPO: Mitigating Double Descent in Low-Rank Adaptation via Momentum-Guided Perturbation Optimization (2025.findings-emnlp)
Copied to clipboard
| Challenge: | Low-Rank Adaptation (LoRA) adapts large language models by training only a small fraction of parameters, but as the rank of the low-rank matrices increases, LoRA exhibits an unstable “double descent” phenomenon, which delays convergence and impairs generalization by causing instability due to the attraction to sharp local minima. |
| Approach: | They propose a framework that incorporates Momentum-Guided Perturbation Optimization (MGPO) MGPO stabilizes training dynamics by mitigating double descent phenomenon and guiding weight perturbations using momentum vectors from the optimizer’s state. |
| Outcome: | The proposed framework improves performance on natural language understanding benchmarks and shows that it improves convergence and generalization. |
Preference-Aware Memory Update for Long-Term LLM Agents (2026.findings-acl)
Copied to clipboard
| Challenge: | Existing methods for integrating long-term memory do not provide dynamic and personalized memory refinement. |
| Approach: | They propose a long-term memory update mechanism that enables dynamic and personalized memory refinement. |
| Outcome: | The proposed mechanism improves the performance of LLM-based agents in five tasks. |
LLMs on a Budget? Say HOLA (2025.emnlp-industry)
Copied to clipboard
Zohaib Hasan Siddiqui, Jiechao Gao, Ebad Shabbir, Mohammad Anas Azeez, Rafiq Ali, Gautam Siddharth Kashyap, Usman Naseem
| Challenge: | Current solutions such as quantization, pruning, and Retrieval-Augmented Generation (RAG) offer only partial optimizations and often sacrifice accuracy, speed, or generality. |
| Approach: | They propose an end-to-end optimization framework for efficient LLM deployment . it leverages Hierarchical Speculative Decoding (HSD) for faster inference without quality loss. |
| Outcome: | HOLA delivers +17.6% EMA on GSM8K, +10.5% MCA on ARC, and reduced latency and memory on edge devices like Jetson Nano. |
Reviving Iterative Refinement in Diffusion-based NER with an Initializer-Restorer Approach (2026.acl-short)
Copied to clipboard
| Challenge: | Named Entity Recognition (NER) is a fundamental task in Information Extraction. |
| Approach: | They propose a generative paradigm for Named Entity Recognition by modeling NER as a boundary diffusion process. |
| Outcome: | The proposed model performs better than baseline on ACE2004, GENIA, and CleanCoNLL. |
BV-Blend: Uncertainty-Weighted Historical Baselines for Stable Critic-Free RL with Verifiable Rewards (2026.findings-acl)
Copied to clipboard
| Challenge: | Critic-free reinforcement learning with verifiable rewards (RLVR) is a practical paradigm for aligning Large Language Models. |
| Approach: | They propose a framework that stabilizes advantage estimation by combining prompt-local on-policy statistics with semantic-cluster-conditioned historical moments. |
| Outcome: | Experiments show that RLVR improves training stability and performance compared to critic-based methods . compared with other approaches, RL VR improves in cold-start regimes with binary verifiers . |
MAFiD: Moving Average Equipped Fusion-in-Decoder for Question Answering over Tabular and Textual Data (2023.findings-eacl)
Copied to clipboard
| Challenge: | Experimental results show that Transformer-based questions have a "long" hybrid sequence over tabular and textual elements, causing long-range reasoning problems. |
| Approach: | They propose a moving average-equipped fusion-in-decoder to handle long-range reasoning problems . they use FiD and EMA to combine different levels of reasoning . |
| Outcome: | Experimental results show that the proposed model increases exact matching and F1 by 1.1 and 1.7 on the blind test set. |
EMA: An Episodic Memory Agent for Efficient and Selective Memory (2026.findings-acl)
Copied to clipboard
| Challenge: | Existing memory-augmented methods often incorporate full dialog histories without filtering, resulting in information redundancy and inference latency. |
| Approach: | They propose a framework that abstracts conversational context into Episodic Memory Units (EMUs) they propose EMA, MemDecider and a filtering decision module to reduce noise and improve overall performance. |
| Outcome: | The proposed framework reduces token consumption by 11.48% while improving performance on two widely-used benchmarks. |
AGGC: Adaptive Group Gradient Clipping for Stabilizing Large Language Model Training (2026.findings-acl)
Copied to clipboard
| Challenge: | Adaptive group-wise gradient clipping (AGGC) is a new approach to stabilize training of Large Language Models. |
| Approach: | They propose a method to stabilize gradient clipping by partitioning parameters into groups based on functional types and a time-dependent scheduling mechanism to balance exploration and convergence. |
| Outcome: | The proposed algorithm outperforms standard LoRA and achieves 72.93% accuracy . it can be integrated into existing pipelines with negligible overhead. |
Noisy Pair Corrector for Dense Retrieval (2023.findings-emnlp)
Copied to clipboard
| Challenge: | Existing dense retrieval models assume that query-document pairs are exactly matched, resulting in mismatched-pair noise. |
| Approach: | They propose a novel approach to train an effective model with mismatched-pair noise. |
| Outcome: | The proposed model performs well on natural question and triviaQA, code-search benchmarks and SO-DS. |
LENS: LLM-Enabled Narrative Synthesis for Mental Health by Aligning Multimodal Sensing with Language Models (2026.acl-long)
Copied to clipboard
Wenxuan Xu, Arvind Pillai, Subigya Nepal, Amanda C. Collins, Daniel M Mackin, Michael V. Heinz, Tess Z Griffin, Nicholas C. Jacobson, Andrew Campbell
| Challenge: | Current LLMs cannot natively ingest long-duration sensor streams and paired sensor–text datasets are scarce. |
| Approach: | They propose a framework that aligns multimodal sensing data with language models to generate clinically grounded mental-health narratives. |
| Outcome: | The proposed framework outperforms baselines on NLP metrics and task-specific measures of symptom severity and clinically meaningful narratives. |
Emergent Abilities of Large Language Models under Continued Pre-training for Language Adaptation (2025.acl-long)
Copied to clipboard
| Challenge: | Existing large language models are notoriously English-centric, and their performance has been reported to drop significantly in lessresourced languages. |
| Approach: | They propose a language-agnostic benchmark for in-context learning that reveals catastrophic forgetting early on CPT when English is not included. |
| Outcome: | The proposed method does not impact validation perplexity but is critical for emergence of downstream capabilities in the target language. |
DPLoRA: A Dual-Pruning Framework based on ILP Optimization and Progressive Pruning for Parameter-Efficient LoRA Fine-Tuning (2026.findings-acl)
Copied to clipboard
| Challenge: | Large language models (LLMs) require computational resources for fine-tuning. |
| Approach: | They propose a framework that optimizes rank allocation via two stages . they propose an initial pruning stage and a progressive pruning stage . |
| Outcome: | The proposed framework outperforms existing PEFT baselines on GLUE and instruction-following tasks while reducing training time and trainable parameters by over 80%. |