| Challenge: | Existing activation intervention methods involve an activation probe to detect undesirable generation, triggering the activation modification to steer subsequent generation. |
| Approach: | They propose a probe-free intervention method FLORAIN for all attention heads in a specific activation layer. |
| Outcome: | The proposed method outperforms baseline methods in enhancing model truthfulness and quality across generation and multiple-choice tasks. |
Similar Papers
Task-driven Layerwise Additive Activation Intervention (2025.naacl-short)
Copied to clipboard
| Challenge: | Existing approaches to task adaptation rely heavily on heuristic rules or prompt inputs. |
| Approach: | They propose a layer-wise additive activation intervention framework that steers the LMs’ generation process by identifying and manipulating the activations. |
| Outcome: | The proposed framework improves the accuracy of pretrained LMs and competing baselines on various datasets, demonstrating improvements in the accuracy and sample efficiency of the proposed framework. |
Distributional Surgery for Language Model Activations (2025.findings-emnlp)
Copied to clipboard
| Challenge: | Language models can produce undesirable outputs including harmful or toxic outputs. |
| Approach: | They propose a method to detect undesirable content using activations . they propose layerwise distributional steering policies that transform the attention heads . |
| Outcome: | The proposed method outperforms baselines in reducing undesirable output generation. |
Gradient-guided Attention Map Editing: Towards Efficient Contextual Hallucination Mitigation (2025.findings-naacl)
Copied to clipboard
| Challenge: | Large Language Models (LLMs) often experience “contextual hallucination” where they prioritize self-generated content over input context, leading to a disregard for pertinent details. |
| Approach: | They propose a method that dynamically adjusts attention maps to enhance contextual relevance by using a trained classifier to identify attention maps likely to induce hallucinations. |
| Outcome: | The proposed approach reduces hallucinations across open-source models on summarization and open-book QA tasks. |
AROMA: Autonomous Rank-one Matrix Adaptation (2025.emnlp-main)
Copied to clipboard
| Challenge: | Low-rank adaptation (LoRA) and adaptive low-rank adaption (AdaLoRa) are effective for large language models but are expensive as model sizes escalate into hundreds of billions of parameters. |
| Approach: | They propose a framework that automatically builds up rank-one components with very few trainable parameters that gradually diminish to zero. |
| Outcome: | The proposed framework significantly reduces parameters compared to LoRA and AdaLoRA while maintaining subspace independence. |
GrAInS: Gradient-based Attribution for Inference-Time Steering of LLMs and VLMs (2026.acl-long)
Copied to clipboard
| Challenge: | Existing methods for fine-tuning large language models often ignore token-level causal influence and underutilize model logits. |
| Approach: | They propose a novel approach that uses a gradient-based approach to identify influential tokens and construct directional steering vectors based on their contribution to preferred over dispreferred outputs. |
| Outcome: | The proposed approach outperforms fine-tuning and prior steering methods on both LLM and VLM tasks without degrading fluency or general capabilities. |
Enhanced Language Model Truthfulness with Learnable Intervention and Uncertainty Expression (2024.findings-acl)
Copied to clipboard
| Challenge: | Large language models (LLMs) generate long-form and coherent text, yet they often hallucinate facts, which undermines their reliability. |
| Approach: | They propose a Learnable Intervention method for Truthfulness Optimization that automatically identifies the optimal intervention intensity tailored to each query context. |
| Outcome: | Experiments on multiple LLMs and question-answering datasets show that LITO improves truthfulness while preserving task accuracy. |
Astra: Activation-Space Tail-Eigenvector Low-Rank Adaptation of Large Language Models (2026.findings-acl)
Copied to clipboard
| Challenge: | Existing methods for fine-tuning pre-trained models are limited due to suboptimal activation subspaces. |
| Approach: | They propose a method that leverages tail eigenvectors of model output activations to construct low-rank adapters. |
| Outcome: | The proposed method outperforms existing methods across 16 benchmarks and surpasses full fine-tuning in certain scenarios. |
IMPACT: Importance-Aware Activation Space Reconstruction (2026.acl-long)
Copied to clipboard
| Challenge: | Large language models (LLMs) achieve strong performance across domains but remain difficult to deploy in resource-constrained environments due to their massive size. |
| Approach: | They propose an importance-aware activation reconstruction framework that links compression to its effect on model performance. |
| Outcome: | Experiments show that IMPACT reduces model size by 55.4% while maintaining accuracy comparable to or better than state-of-the-art models. |
Steering into New Embedding Spaces: Analyzing Cross-Lingual Alignment Induced by Model Interventions in Multilingual Language Models (2025.acl-long)
Copied to clipboard
Anirudh Sundar, Sinead Williamson, Katherine Metcalf, Barry-John Theobald, Skyler Seto, Masha Fedzechkina
| Challenge: | Large language models (LLMs) exhibit impressive performance on a variety of tasks from text summarization to zero-shot common-sense reasoning. |
| Approach: | They propose to manipulate the embedding space of mLLMs by manipulating its activations to steer generation into the desired direction. |
| Outcome: | The proposed model interventions improves alignment of cross-lingual representations in multilingual large language models with up to 2x improvements in top-1 accuracy on cross-linguistic retrieval tasks. |
Scalable Efficient Training of Large Language Models with Low-dimensional Projected Attention (2024.emnlp-main)
Copied to clipboard
| Challenge: | Existing studies have found that low-rank pre-training often compromises effectiveness. |
| Approach: | They propose to apply low-dimensional module only to the attention layer to improve both effectiveness and efficiency. |
| Outcome: | The proposed model saves 12.4% time while improving test perplexity and on downstream tasks compared with vanilla Transformer. |