Papers with model retraining

21 papers
Thesis Proposal: Efficient KV Cache Reuse for Multi-Document Retrieval-Augmented Generation (2026.eacl-srw)

Copied to clipboard

Challenge: Retrieval-Augmented Generation (RAG) systems face efficiency bottlenecks in prefill due to attention mechanism, and traditional KV cache only accelerates decoding.
Approach: They propose a multi-document KV cache reuse framework for multi-doc RAG workloads . they propose to resolve position and context misalignment while eliminating document-specific quadratic complexity in prefill.
Outcome: The proposed framework solves position and context misalignment issues while eliminating document-specific quadratic complexity in prefill.
SCAIR: Schema-Conditioned Agentic Iterative Reasoning for Enterprise Knowledge Graphs (2026.acl-industry)

Copied to clipboard

Challenge: Existing agentic approaches for Knowledge Graph-based Retrieval-Augmented Generation fail to generalize to real-world enterprise Knowledge graphs (KGs) dense, schema-driven, and operationally constrained, requiring a training-free framework.
Approach: They propose a training-free framework that integrates structured planning with controlled iterative reasoning by injecting schema-conditioned structural priors and enforcing schemas during multi-hop reasoning.
Outcome: The proposed framework significantly improves on a real-world enterprise-oriented benchmark constructed from a Configuration Management DataBase (CMDB).
Self-Regulated Sample Diversity in Large Language Models (2024.findings-naacl)

Copied to clipboard

Challenge: Existing methods that require expensive setups or maintain static values during inference are inflexible and require expensive training.
Approach: They propose a self-regulating approach that adjusts sample diversity parameters dynamically based on the input prompt.
Outcome: The proposed method significantly improves the quality of responses generically without model retraining or fine-tuning.
TopicGPT: A Prompt-based Topic Modeling Framework (2024.naacl-long)

Copied to clipboard

Challenge: TopicGPT uses large language models to uncover latent topics in text . topic models represent topics as bags of words that require "reading the tea leaves" topic models also offer limited control over formatting and specificity of topics .
Approach: TopicGPT uses large language models to uncover latent topics in text . authors propose a prompt-based framework that produces topics that align better with human categorizations .
Outcome: TopicGPT produces topics that align better with human categorizations compared to competing methods.
RelayAttention for Efficient Large Language Model Serving with Long System Prompts (2024.acl-long)

Copied to clipboard

Challenge: a long system prompt causes throughput/latency bottlenecks as the cost of generating the next token increases w.r.t the sequence length.
Approach: They propose an attention algorithm that reads hidden state from DRAM exactly once for a batch of input tokens.
Outcome: The proposed algorithm reduces the need for redundant memory accesses in existing algorithms.
How Credible Is an Answer From Retrieval-Augmented LLMs? Investigation and Evaluation With Multi-Hop QA (2025.coling-main)

Copied to clipboard

Challenge: Retrieval-augmented large language models (RaLLMs) are reshaping knowledge acquisition, offering long-form, knowledge-grounded answers through advanced reasoning and generation capabilities.
Approach: They propose a benchmarking system to evaluate RaLLMs' correctness and Groundedness to determine their reliability in multi-hop question-answering tasks.
Outcome: The proposed model-based evaluation pipeline for multi-hop question-answering tasks reveals that the model generates inaccuracies when dealing with flawed or partial knowledge.
RecGPT: A Foundation Model for Sequential Recommendation (2025.emnlp-main)

Copied to clipboard

Challenge: Existing approaches fail in cold-start and cross-domain scenarios where new users or items lack sufficient interaction history.
Approach: They propose a foundation model for sequential recommendation that achieves genuine zero-shot generalization capabilities by deriving item representations exclusively from textual features.
Outcome: The proposed model achieves zero-shot generalization capabilities in cold-start and cross-domain scenarios.
Mitigating Shortcut Learning via Smart Data Augmentation based on Large Language Model (2025.coling-main)

Copied to clipboard

Challenge: Existing methods to improve shortcut learning performance are limited by manual definition of shortcuts and inherent confirmation bias during model training.
Approach: They propose a method of Smart Data Augmentation based on Large Language Models to identify shortcuts and generate their anti-shortcut counterparts.
Outcome: The proposed method shows an improvement of 5.61% across various natural language processing tasks.
FairSteer: Inference Time Debiasing for LLMs with Dynamic Activation Steering (2025.findings-acl)

Copied to clipboard

Challenge: Existing prompt-based debiasing methods exhibit instability due to sensitivity to prompt changes . fine-tuning-based techniques incur substantial computational overhead and catastrophic forgetting .
Approach: They propose a debiasing framework that encodes fairness-related features into separable directions in the hidden activation space.
Outcome: The proposed framework performs inference-time debiasing without requiring retraining or prompt design . it detects bias signatures in activations and then computes debiased steering vectors . the proposed framework is available to download in the u.s.
HeteroSpec: Leveraging Contextual Heterogeneity for Efficient Speculative Decoding (2026.acl-long)

Copied to clipboard

Challenge: Autoregressive decoding limits the inference throughput of Large Language Models due to its sequential dependency.
Approach: They propose a framework that allocates verification effort in proportion to candidate uncertainty.
Outcome: Speculative decoding achieves an average speedup over state-of-the-art methods . a small subset of high-confidence predictions accounts for most successful verifications .
Finding the Sweet Spot: Preference Data Construction for Scaling Preference Optimization (2025.acl-long)

Copied to clipboard

Challenge: Large language models generate unintended outputs due to their unsupervised nature.
Approach: They propose a method to construct preference pairs of selected and rejected LLMs by repeated random sampling to improve alignment performance.
Outcome: The proposed method improves performance as the sample size increases.
Towards Teachable Reasoning Systems: Using a Dynamic Memory of User Feedback for Continual System Improvement (2022.emnlp-main)

Copied to clipboard

Challenge: Using simulated feedback, our system (called TeachMe) continually improves with time, and without model retraining.
Approach: They propose to augment a QA model with a dynamic memory of user feedback, containing user-supplied corrections toerroneous model beliefs that users identify during interaction.
Outcome: The proposed system improves with time and without model retraining, and with real users, by 15% on a hidden test set after teaching.
Where Am I From? Identifying Origin of LLM-generated Content (2024.emnlp-main)

Copied to clipboard

Challenge: Generative models have produced high-quality content, but they pose security risks . a new framework for deep learning systems enables the tracing of AI-generated content back to its source .
Approach: They propose a digital forensics framework that embeds a secret watermark into the generated output and a "depth watermark" this watermark strengthens the link between content and generator, enabling accurate tracing while maintaining the quality of the generated content.
Outcome: The proposed framework ensures accurate tracing while maintaining quality of generated content.
Peering into the Mind of Language Models: An Approach for Attribution in Contextual Question Answering (2024.findings-acl)

Copied to clipboard

Challenge: Large language models (LLMs) are increasingly used for question answering . lack of explicit references or attributions hinders ability to verify accuracy of answers .
Approach: They propose a method for attribution in contextual question answering . they use hidden state representations of large language models to identify copied segments .
Outcome: The proposed method performs better than GPT-4 at identifying verbatim copied segments in LLM generations and attributing these segments to their source.
BeliefBank: Adding Memory to a Pre-Trained Language Model for a Systematic Notion of Belief (2021.emnlp-main)

Copied to clipboard

Challenge: Pretrained language models can produce inconsistent answers when probed, even after specialized training.
Approach: They propose to embed a pretrained language model in a broader system that includes an evolving, symbolic memory of beliefs that records but may modify the raw PTLM answers.
Outcome: The proposed architecture improves belief consistency in the overall system by revising beliefs that clash with others and generating queries using known beliefs as context.
Evaluating and Characterizing Human Rationales (2020.emnlp-main)

Copied to clipboard

Challenge: a new study examines how human rationales perform on automatic metrics . human-generated rationale evaluation is difficult because of its ambiguity .
Approach: They propose to use model-dependent baseline performance to evaluate rationale quality . they propose to also use "fidelity curves" to reveal properties such as irrelevance and redundancy .
Outcome: The proposed methods characterize rationale quality based on model retraining and using "fidelity curves" the proposed methods lead to actionable suggestions for evaluating and characterizing rationales .
FAME: Towards Factual Multi-Task Model Editing (2024.emnlp-main)

Copied to clipboard

Challenge: Large language models embed extensive knowledge and perform exceptionally well across tasks. outdated knowledge or factual errors within LLMs can lead to misleading or incorrect responses.
Approach: They propose to use a dataset to enhance the practicality of model editing to correct inaccurate information within LLMs.
Outcome: The proposed method performs excellently across tasks and scenarios, confirming its practicality.
FAEDKV: Infinite-Window Fourier Transform for Unbiased KV Cache Compression (2025.findings-emnlp)

Copied to clipboard

Challenge: Current compression strategies, including token eviction and learned projections, often lead to biased representations and may require costly model retraining.
Approach: They propose a training-free KV cache compression framework that equalizes the contribution of all tokens to the compressed representation.
Outcome: The proposed framework ensures unbiased information retention in the KV cache.
Rectifying the Emotional Flow: Aligning Priors and Dynamic Guidance for High-Arousal Text-to-Speech (2026.acl-long)

Copied to clipboard

Challenge: Existing systems suffer from linguistic collapse when pursuing high intensity or fail to meet target emotional levels.
Approach: They propose an inference framework that introduces a neutral prosody bias and a uniform Classifier-Free Guidance that distorts the acoustic manifold, leading to artifacts.
Outcome: The proposed framework achieves superior linguistic accuracy and expressiveness without model retraining.
Disentangled Information Bottleneck for Adversarial Text Defense (2025.emnlp-main)

Copied to clipboard

Challenge: Existing studies have proven that these deep models are super vulnerable to adversarial examples, which are slightly modified inputs.
Approach: They propose a novel text defense method that separates the robust and non-robust features with a disentangled two-line framework rather than the one-line compression network in IB.
Outcome: The proposed method outperforms six baselines on four datasets with accuracy improvements ranging from 3.8% to 20.7%.
VLA-Mark: A cross modal watermark for large vision-language alignment models (2025.emnlp-main)

Copied to clipboard

Challenge: Existing text watermarking methods disrupt visual-textual alignment, leaving semantic-critical concepts vulnerable.
Approach: They propose a vision-aligned framework that embeds detectable watermarks into outputs . they combine localized patch affinity, global semantic coherence, contextual attention patterns .
Outcome: The proposed framework shows lower PPL and higher BLEU than conventional methods with near-perfect detection (98.8% AUC).

What is GenGO?

GenGO is an NLP powered publication search system. It currenctly indexes 30k+ papers from ACL Anthology, and implements multi-aspect summarization, semantic search, and more!

Information

About
Limitations