Thailand

Large language models are meticulously aligned to be both helpful and harmless. However, recent research points to a potential overkill which means models may refuse to answer benign queries. In this paper, we investigate the factors for overkill by exploring how models handle and determine the safety of queries. Our findings reveal the presence of shortcuts within models, leading to excessive attention to harmful words like &#39;kill&#39; and prompts emphasizing safety will exacerbate overkill. Based on these insights, we introduce Self-Contrastive Decoding (Self-CD), a training-free and model-agnostic strategy, to alleviate this phenomenon. We first extract such excessive attention by amplifying the difference in the model&#39;s output distributions when responding to system prompts that either include or omit an emphasis on safety. Then we determine the final next-token predictions by downplaying the excessive attention via contrastive decoding. Empirical results have indicated that our method has achieved an average reduction of the refusal rate by 20 % while having almost no impact on safety.

ACL 2024

Navigating the OverKill in Large Language Models

large language models; safety; overkill

## Welcome!
"I am excited to welcome you to this year’s edition of the Conference on Empirical Methods in Natural Language Processing! Importantly, it marks the 30th edition of EMNLP. With over 8,000 submissions, more than 3,000 accepted papers, and thousands of attendees, we have come a long way from that first
workshop, which had 14 accepted papers. As the field looks ahead, Suzhou is the fitting location for celebrating this milestone: rooted in a long literary tradition, yet modern and forward-looking, and home to a large share of the NLP community." 

*Message from the General Chair, Dirk Hovy*

EMNLP 2025

Celebrate 30 Years of EMNLP! 
EMNLP 2025 will be held in Suzhou, China from November 5th to November 9th, 2025.

Large language models are meticulously aligned to be both helpful and harmless. However, recent research points to a potential overkill which means models may refuse to answer benign queries. In this paper, we investigate the factors for overkill by exploring how models handle and determine the safety of queries. Our findings reveal the presence of shortcuts within models, leading to excessive attention to harmful words like 'kill' and prompts emphasizing safety will exacerbate overkill. Based on these insights, we introduce Self-Contrastive Decoding (Self-CD), a training-free and model-agnostic strategy, to alleviate this phenomenon. We first extract such excessive attention by amplifying the difference in the model's output distributions when responding to system prompts that either include or omit an emphasis on safety. Then we determine the final next-token predictions by downplaying the excessive attention via contrastive decoding. Empirical results have indicated that our method has achieved an average reduction of the refusal rate by 20 % while having almost no impact on safety.

poster

### Welcome!
The 62nd Annual Meeting of the Association for Computational Linguistics (ACL 2024) will take place in Bangkok, Thailand from August 11th to 16th, 2024. Our Virtual Poster Sessions will take place online Thursday, August 22, 2024.

You are required to register for this event. **Please register [here](https://2024.aclweb.org/registration). **

If you have already registered, please check your inbox for an email from Underline granting you access to ACL 2024 content.

Please register!

The 62nd Annual Meeting of the Association for Computational Linguistics (ACL 2024) will take place in Bangkok, Thailand from August 11th to 16th, 2024. More information will be announced soon.

Projecting intermediate representations onto the vocabulary is an increasingly popular interpretation tool for transformer-based LLMs, also known as the logit lens (Nostalgebraist). We propose a quantitative extension to this approach and define spectral filters on intermediate representations based on partitioning the singular vectors of the vocabulary embedding and unembedding matrices into bands. We find that the signals exchanged in the tail end of the spectrum, i.e. corresponding to the singular vectors with smallest singular values, are responsible for attention sinking (Xiao et al., 2023), of which we provide an explanation. We find that the negative log-likelihood of pretrained models can be kept low despite suppressing sizeable parts of the embedding spectrum in a layer-dependent way, as long as attention sinking is preserved. Finally, we discover that the representation of tokens that draw attention from many tokens have large projections on the tail end of the spectrum, and likely act as additional attention sinks.

Spectral Filters, Dark Signals, and Attention Sinks

Large language models, comprising billions of parameters and pre-trained on extensive web-scale corpora, have been claimed to acquire certain capabilities without having been specifically trained on them. These capabilities, referred to as "emergent abilities," have been a driving force in discussions regarding the potentials and risks of language models. A key challenge in evaluating emergent abilities is that they are confounded by model competencies that arise through alternative prompting techniques, including in-context learning, which is the ability of models to complete a task based on a few examples. We present a novel theory that explains emergent abilities, taking into account their potential confounding factors, and rigorously substantiate this theory through over 1000 experiments. Our findings suggest that purported emergent abilities are not truly emergent, but result from a combination of in-context learning, model memory, and linguistic knowledge. Our work is a foundational step in explaining language model performance, providing a template for their efficient use and clarifying the paradox of their ability to excel in some instances while faltering in others. Thus, we demonstrate that their capabilities should not be overestimated.

Are Emergent Abilities in Large Language Models just In-Context Learning?

Recent generalist vision-language models (VLMs) have demonstrated impressive reasoning capabilities across diverse multimodal tasks. However, these models still struggle with fine-grained object-level understanding and grounding. In terms of modeling, existing VLMs implicitly align text tokens with image patch tokens, which is ineffective for embedding alignment at the same granularity and inevitably introduces noisy spurious background features. Additionally, these models struggle when generalizing to unseen visual concepts and may not be reliable for domain-specific tasks without further fine-tuning. To address these limitations, we propose a novel method to prompt large language models with in-context visual object vectors, thereby enabling controllable object-level reasoning. This eliminates the necessity of fusing a lengthy array of image patch features and significantly speeds up training. Furthermore, we propose region-level retrieval using our object representations, facilitating rapid adaptation to new objects without additional training. Our experiments reveal that our method achieves competitive referring object classification and captioning performance, while also offering zero-shot generalization and robustness to visually challenging contexts.

OLIVE: Object Level In-Context Visual Embeddings

Large Multimodal Models (LMMs) excel in natural language and visual understanding but are challenged by exacting tasks such as Knowledge-based Visual Question Answering (KB-VQA) which involve the retrieval of relevant information from document collections to use in shaping answers to questions. We present an extensive training and evaluation framework, M2KR, for KB-VQA. M2KR contains a collection of vision and language tasks which we have incorporated into a single suite of benchmark tasks for training and evaluating general-purpose multi-modal retrievers. We use M2KR to develop PreFLMR, a pre-trained version of the recently developed Fine-grained Late-interaction Multi-modal Retriever (FLMR) approach to KB-VQA, and we report new state-of-the-art results across a range of tasks. We also present investigations into the scaling behaviors of PreFLMR intended to be useful in future developments in general-purpose multi-modal retrievers.

PreFLMR: Scaling Up Fine-Grained Late-Interaction Multi-modal Retrievers

This paper reveals a novel linear characteristic exclusive to transformer decoders, including models like GPT, LLaMA, OPT, BLOOM and others. We analyze embedding transformations between sequential layers, uncovering an almost perfect linear relationship (Procrustes similarity score of 0.99). However, linearity decreases when the residual component is removed, due to a consistently low transformer layer output norm. Our experiments show that pruning or linearly approximating some of the layers does not impact loss or model performance significantly. Moreover, we introduce a cosine-similarity-based regularization in our pretraining experiments on smaller models, aimed at reducing layer linearity. This regularization not only improves performance metrics on benchmarks like Tiny Stories and SuperGLUE but as well successfully decreases the linearity of the models. This study challenges the existing understanding of transformer architectures, suggesting that their operation may be more linear than previously assumed.

Your Transformer is Secretly Linear

This work explores whether language models encode meaningfully grounded representations of sounds of objects. We learn a linear probe that retrieves the correct text representation of an object given a snippet of audio related to that object, where the sound representation is given by a pretrained audio model. This probe is trained via a contrastive loss that pushes the language representations and sound representations of an object to be close to one another. After training, the probe is tested on its ability to generalize to objects that were not seen during training. Across different language models and audio models, we find that the probe generalization is above chance in many cases, indicating that despite being trained only on raw text, language models encode grounded knowledge of sounds for some objects.

What Do Language Models Hear? Probing for Auditory Representations in Language Models

Recent work shows that in-context learning and optimization of in-context examples (ICE) can significantly improve the accuracy of large language models (LLMs) on a wide range of tasks, leading to an apparent consensus that ICE optimization is crucial for better performance. However, most of these studies assume a fixed or no instruction provided in the prompt. We challenge this consensus by investigating the necessity of optimizing ICE when task-specific instructions are provided and find that there are many tasks for which it yields diminishing returns. In particular, using a diverse set of tasks and a systematically created instruction set with gradually added details, we find that as the prompt instruction becomes more detailed, the returns on ICE optimization diminish. To characterize this behavior, we introduce a task-specific metric called Normalized Invariability to Choice of Examples (NICE) that quantifies the learnability of tasks from a given instruction, and provides a heuristic to help decide whether to optimize instructions or ICE for a new task. 
Given a task, the proposed metric can reliably predict the utility of optimizing ICE compared to using random ICE. Our code is available at [https://github.com/microsoft/nice-icl](https://github.com/microsoft/nice-icl).

NICE: To Optimize In-Context Examples or Not?

We present the first domain-adapted and fully-trained large language model, RecGPT-7B, and its instruction-following variant, RecGPT-7B-Instruct, for text-based recommendation. Experimental results on rating prediction and sequential recommendation tasks show that our model, RecGPT-7B-Instruct, outperforms previous strong baselines. We are releasing our RecGPT models as well as their pre-training and fine-tuning datasets to facilitate future research and downstream applications in text-based recommendation. Public "huggingface" links to our RecGPT models and datasets are available at: https://github.com/VinAIResearch/RecGPT

RecGPT: Generative Pre-training for Text-based Recommendation

Pairwise Ranking Prompting (PRP) demonstrates impressive effectiveness in zero-shot document re-ranking tasks with large language models (LLMs). However, in the existing methods, PRP only outputs the same label for the comparison results of different confidence intervals without considering the uncertainty of pairwise comparison, which implies an underutilization of the generation probability information of LLMs. To bridge this gap, we propose PRP-Graph, a novel pairwise re-ranking approach, based on a refined scoring PRP unit that exploits the output probabilities of target labels to capture the degree of certainty of the comparison results. Specifically, the PRP-Graph consists of two stages, namely ranking graph construction and ranking graph aggregation. Extensive experiments conducted on the BEIR benchmark demonstrate the superiority of our approach over existing PRP-based methods. Comprehensive analysis reveals that the PRP-Graph displays strong robustness towards the initial ranking order and delivers exceptional re-ranking results with acceptable efficiency. Our code and data are available at https://github.com/Memelank/PRP-Graph.

PRP-Graph: Pairwise Ranking Prompting to LLMs with Graph Aggregation for Effective Text Re-ranking

The swift detection of multimedia fake news has emerged as a crucial task in combating malicious propaganda and safeguarding the security of the online environment. While existing methods have achieved commendable results in modeling entity-level inconsistency, addressing event-level inconsistency following the inherent subject-predicate logic of news and robustly learning news representations from poor-quality news samples remain two challenges. In this paper, we propose an Event-diven fake news detection framework (Event-Radar) based on multi-view learning, which integrates visual manipulation, textual emotion and multimodal inconsistency at event-level for fake news detection. Specifically, leveraging the capability of graph structures to capture interactions between events and parameters, Event-Radar captures event-level multimodal inconsistency by constructing an event graph that includes multimodal entity subject-predicate logic. Additionally, to mitigate the interference of poor-quality news, Event-Radar introduces a multi-view fusion mechanism, learning comprehensive and robust representations by computing the credibility of each view as a clue, thereby detecting fake news. Extensive experiments demonstrate that Event-Radar achieves outstanding performance on three large-scale fake news detection benchmarks. Our studies also confirm that Event-Radar exhibits strong robustness, providing a paradigm for detecting fake news from noisy news samples.

Downloads

Next from ACL 2024

Spectral Filters, Dark Signals, and Attention Sinks

Similar lecture

SafeKey: Amplifying Aha-Moment Insights for Safety Reasoning

Stay up to date with the latest Underline news!

PRESENTATIONS

CONFERENCES

COMPANY

RESOURCES