United States

For induction heads to copy forward information successfully, heads in earlier layers must first load previous token information into every hidden state, a process Olsson et al. (2022) call &quot;key shifting&quot;. While this information is hypothesized to exist, there have been few attempts to explicitly locate it in models. In this work, we use linear probes to identify the subspaces responsible for storing previous token information in Llama-2-7b and Llama-3-8b. We show that these subspaces are causally implicated in induction by using them to &quot;edit&quot; previous token information and trigger random token copying in new contexts.

EMNLP 2024

Inducing Induction in Llama via Linear Probe Interventions

induction heads

mechanistic interpretability

### 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!

ACL 2024

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.

For induction heads to copy forward information successfully, heads in earlier layers must first load previous token information into every hidden state, a process Olsson et al. (2022) call "key shifting". While this information is hypothesized to exist, there have been few attempts to explicitly locate it in models. In this work, we use linear probes to identify the subspaces responsible for storing previous token information in Llama-2-7b and Llama-3-8b. We show that these subspaces are causally implicated in induction by using them to "edit" previous token information and trigger random token copying in new contexts.

## Welcome to EMNLP 2024! 
We are excited to welcome you to one of the most prominent conferences in the field of Natural Language Processing. This year, EMNLP 2024 is being held in a hybrid format,
offering both virtual and in-person participation in beautiful Miami. Due to a record-breaking number of submissions, we've expanded the total number of accepted papers to accommodate more cutting-edge research from around the globe.
### [Conference Handbook](https://drive.google.com/file/d/1WPROgxjLAC96AJL7Ugy0tEnYm7dkrbHt/view?usp=sharing)

You are required to register for this event. **Please register [here](https://2024.emnlp.org/registration/).** The EMNLP 2024 event page on Underline will be open to public one week prior to the event.

EMNLP 2024 will take place in Miami, Florida from Nov 12th to Nov 16th, 2024, at the Hyatt Regency Miami Hote and on Underline for remote participants.

Whether LLMs can estimate their hallucination risk based on internal states before response generation.

LLM Internal States Reveal Hallucination Risk Faced With a Query

The surge of state-of-the-art transformer-based models has undoubtedly pushed the limits of NLP model performance, excelling in a variety of tasks. We cast the spotlight on the underexplored task of Natural Language Inference (NLI), since models trained on popular well-suited datasets are susceptible to adversarial attacks, allowing subtle input interventions to mislead the model. In this work, we validate the usage of natural language explanation as a model-agnostic defence strategy through extensive experimentation: only by fine-tuning a classifier on the explanation rather than premise-hypothesis inputs, robustness under various adversarial attacks is achieved in comparison to explanation-free baselines. Moreover, since there is no standard strategy for testing the semantic validity of the generated explanations, we research the correlation of widely used language generation metrics with human perception, in order for them to serve as a proxy towards robust NLI models. Our approach is resource-efficient and reproducible without significant computational limitations.

Enhancing adversarial robustness in Natural Language Inference using explanations

Faithfulness is a critical metric to assess the reliability of explainable AI. In NLP, current methods for faithfulness evaluation are fraught with discrepancies and biases, often failing to capture the true reasoning of models. We in- troduce Adversarial Sensitivity as a novel ap- proach to faithfulness evaluation, focusing on the explainer’s response when the model is un- der adversarial attack. Our method accounts for the faithfulness of explainers by capturing sensitivity to adversarial input changes. This work addresses significant limitations in ex- isting evaluation techniques, and furthermore, quantifies faithfulness from a crucial yet under- explored paradigm.

Faithfulness and the Notion of Adversarial Sensitivity in NLP Explanations

Language models have displayed a wide array of capabilities, but the reason for their performance remains a topic of heated debate and investigation. Do these models simply recite the observed training data, or are they able to abstract away surface statistics and learn the underlying processes from which the data was generated? To investigate this question, we explore the capabilities of a GPT model in the context of Markov Decision Processes (MDPs), where the underlying transition dynamics and policies are not directly observed. The model is trained to predict the next state or action without any initial knowledge of the MDPs or the players' policies. Despite this, we present evidence that the model develops emergent representations of the underlying parameters governing the MDPs.

Transformers Learn Hidden Dynamics when Trained to Predict Markov Decision Processes

Previous research on the predictive power (PP) of surprisal and entropy has focused on determining which language models (LMs) generate estimates with the highest PP on reading times, and examining for which populations the PP is strongest. In this study, we leverage eye movement data on texts that were generated using a range of decoding strategies with different LMs. We then extract the transition scores that reflect the models' production rather than comprehension effort. This allows us to investigate the alignment of LM language production and human language comprehension. Our findings reveal that there are differences in the strength of the alignment between reading behavior and certain LM decoding strategies and that this alignment further reflects different stages of language understanding (early, late, or global processes). Although we find lower PP of transition-based measures compared to surprisal and entropy for most decoding strategies, our results provide valuable insights into which decoding strategies impose less processing effort for readers. Our code is available via https://github.com/DiLi-Lab/LM-human-alignment.

On the alignment of LM language generation and human language comprehension

As intuitive units of speech, syllables have been widely studied in linguistics. A syllable can be defined as a three-constituent unit with a vocalic centre surrounded by two (in some languages optional) consonant clusters. Syllables are also used to design automatic speech recognition (ASR) models. The significance of knowledge-driven syllable-based tokenisation in ASR over data-driven byte-pair encoding has often been debated. However, the emergence of transformer-based ASR models employing self-attention (SA) overshadowed this debate. These models learn the nuances of speech from large corpora without prior knowledge of the domain; yet, they are not interpretable by design. Consequently, it is not clear if the recent performance improvements are related to the extraction of human-interpretable knowledge. We probe such models for syllable constituents and use an SA head pruning method to assess the relevance of the SA weights. We also investigate the role of vowel identification in syllable constituent probing. Our findings show that the general features of syllable constituents are extracted in the earlier layers of the model and the syllable-related features mostly depend on the temporal knowledge incorporated in specific SA heads rather than on vowel identification.

Uncovering Syllable Constituents in the Self-Attention-Based Speech Representations of Whisper

Large Language Models (LLMs) have shown remarkable progress, but their real-world application necessitates reliable calibration. This study conducts a comprehensive analysis of calibration degradation of LLMs across four dimensions: models, calibration metrics, tasks, and confidence extraction methods. Initial analysis showed that the relationship between alignment and calibration is not always a trade-off, but under stricter analysis conditions, we found the alignment process consistently harms calibration. This highlights the need for (1) a careful approach when measuring model confidences and calibration errors and (2) future research into algorithms that can help LLMs to achieve both instruction-following and calibration without sacrificing either.

Does Alignment Tuning Really Break LLMs’ Internal Confidence?

We present a comprehensive explainability dashboard designed for in-game chat toxicity. This dashboard integrates various existing explainable AI (XAI) techniques, including token importance analysis, model output visualization, and attribution to the training dataset. It also provides insights through the closest positive and negative examples, facilitating a deeper understanding and potential correction of the training data. Additionally, the dashboard includes word sense analysis—particularly useful for new moderators—and offers free-text explanations for both positive and negative predictions. This multi-faceted approach enhances the interpretability and transparency of toxicity detection models

ToxiSight: Insights Towards Detected Chat Toxicity

Semantic plausibility (e.g. knowing that “the actor won the award” is more likely than “the actor won the battle”) serves as an effective proxy for general world knowledge. Language models (LMs) capture vast amounts of world knowledge by learning distributional patterns in text, accessible via log probabilities they assign to plausible vs. implausible outputs. The new generation of instruction-tuned LMs can now also provide explicit estimates of plausibility via prompting. Here, we evaluate the effectiveness of log probabilities and basic prompting to measure semantic plausibility, both in single-sentence minimal pairs (Experiment 1) and short context- dependent scenarios (Experiment 2). We find that (i) in both base and instruction-tuned LMs, log probabilities offer a more reliable measure of semantic plausibility than direct zero-shot prompting, which yields inconsistent and often poor results; (ii) instruction-tuning generally does not alter the sensitivity of log probabilities to semantic plausibility (although sometimes decreases it); (iii) across models, context mostly modulates log probabilities in expected ways, as measured by three novel metrics of context-sensitive plausibility and their match to explicit human plausibility judgments. We conclude that, even in the era of prompt-based evaluations, log probabilities constitute a useful metric of semantic plausibility, both in base and instruction-tuned LMs.

Log Probabilities Are a Reliable Estimate of Semantic Plausibility in Base and Instruction-Tuned Language Models

Sparse autoencoders (SAEs) are an unsupervised method for learning a sparse decomposition of a neural network’s latent representations into seemingly interpretable features. Despite recent excitement about their potential, research applications outside of industry are limited by the high cost of training a comprehensive suite of SAEs. In this work, we introduce Gemma Scope, an open suite of JumpReLU SAEs trained on all layers and sub-layers of Gemma 2 2B and 9B and select layers of Gemma 2 27B base models. We primarily train SAEs on the Gemma 2 pre-trained models, but additionally release SAEs trained on instruction-tuned Gemma 2 9B for comparison. We evaluate the quality of each SAE on standard metrics and release these results. We hope that by releasing these SAE weights, we can help make more ambitious safety and interpretability research easier for the community. Weights and a tutorial can be found at https://huggingface.co/google/gemma-scope and an interactive demo can be found at https://neuronpedia.org/gemma-scope.

Next from EMNLP 2024

LLM Internal States Reveal Hallucination Risk Faced With a Query

Similar lecture

Identifying Semantic Induction Heads to Understand In-Context Learning